A litigator's discipline, applied to AI: evaluating the reasoning, identifying unsupported claims, and checking whether the source supports the proposition.

A working specimen of one edition: the same structure, typography, 13 sections, weather table, length rules and link discipline as the real email, with every place, event, price, team and link invented. Scroll inside the box, or .

The audit script and its recorded output are held privately; a live run can be shown on request.

The project's record files are held privately and shown on request.

A scheduled system that researches, source checks, assembles and sends a weekly email of things to do over the following ten days, written for two readers in their seventies who do not use AI tools and who have to be able to read the email cold, with nothing explained to them.

It ran unattended, once a week, on a fixed schedule. Nobody retyped it. The research, the source checking, the cutting to length and the sending all happen inside the run.

One thing that went wrong, stated plainly

On September 10, 2026 a new version was pushed to the live task and then verified against the local draft file believed to have been sent. It did not match. The text that actually reached the live task carried edits the draft did not. Nothing broke, and the live version was the better of the two, but the verification step had been comparing intent against notes rather than notes against the live system. The fix was to read the live task directly and make that read the reference copy. File 01 is built against that live copy, and the superseded draft is deliberately excluded from the version list rather than quietly renumbered.

1. THE PROBLEM. Two people I know, both in their seventies, living about an hour apart in New Jersey, kept not doing things because finding things to do is work. Listings sites are cluttered, half the events are stale, and prices and times are often wrong. Neither of them uses AI and neither should have to. I wanted a single email that lands every week, covers the next ten days, and can be read cold with no explanation from me.

2. WHAT I DECIDED, AND WHAT I REFUSED TO DECIDE BY FEEL. I wrote the requirements as a single instruction document that the system reads on every run, rather than as a prompt I retype. Nineteen dated versions exist, built between September 5 and September 10, 2026. The live version runs about 58,700 bytes. Every change is a string replacement guarded by an assertion that the anchor text occurs exactly once, so an edit that silently misses, or lands twice, fails loudly instead of shipping.

3. WHAT THE AI PRODUCED, AND WHAT I SENT BACK. The AI drafted the instruction document, ran the weekly research, verified sources, cut to length and sent the email. I directed the design and rejected drafts. Things I sent back: half names and half titles, because I could not tell from a copy who one performer was; a covered-areas summary that listed four areas when the email covered five; section numbering with decimals, which confused the readers; and one passage that had grown to 370 words, more of the document than any other reader attribute, which I cut roughly in half because length inside an instruction document signals weight, and an over-weighted note makes a run go hunting for items to fill an implied quota.

4. THE DESIGN CHOICES THAT MATTERED. Verification first: every event is checked on the venue's own page, and schedules are checked against official data feeds rather than aggregator listings. Cost is controlled by a research budget, not by turning the model's effort down: a floor of 30 searches and 25 page reads per run, and a hard ceiling of 50 searches and 235 page reads across every research lane combined. Length is capped at 6,000 words and 95 KB of HTML body, with a written cut order that says what goes first when a draft runs long. Everything is inline styled with no stylesheet and no CSS classes, because email clients strip them. And a staleness rule measures the ten-day window from the moment of sending, not from the moment the run started, so an event that has already happened by the time the email lands is dropped.

5. WHAT WENT WRONG, AND HOW EACH ONE WAS CAUGHT. One draft came back at about 152 KB and Gmail truncated it behind a "View entire message" link. That produced the length ceiling. A link was labeled "Event page" but pointed at a venue's whole calendar rather than the page that had actually been verified, which I caught by auditing the copy against its own sources. A section I had called the most valuable part of the email was set in small grey type at the bottom, which is formatting contradicting my own stated judgment, and it was promoted to the same size as everything else. And after the first edition had already gone out, I found that the email's own "About this email" note described four covered areas while the email covered five. I sent a correction the same morning that said plainly what had been wrong.

6. THE FAILURE I CARE MOST ABOUT, BECAUSE THE SYSTEM CAUGHT IT AND I DID NOT. An earlier edit of mine appended a new research ceiling to a sentence instead of replacing the old one, so a single sentence gave two different limits. I did not notice. The run on September 10 read the contradiction, reported it in its own notes, said which figure it had chosen, and carried on. It chose correctly. I fixed the sentence the same day. The lesson I took is that a system should be built to report the incoherence it finds in its own instructions rather than quietly pick one reading.

7. RECORDING WHAT COULD NOT BE VERIFIED. Two local library calendars could not be read by any route I tried, six routes in total. One of them is worse than blocked: it returns a readable page full of events that are months out of date, which a careless run would publish with confidence. A weather source, asked for a forecast for one of the covered towns, returned a forecast for a town in another state, which the run caught before it reached the email. All of this is written into the instruction document itself, with a two page read per week cap to re-test whether anything has become readable. The email tells the reader plainly that those listings are missing and gives phone numbers instead. I would rather ship a visible gap than an invented filler.

8. WHERE IT STANDS. Three editions were delivered to both readers: September 6, a correction later the same morning, and the first fully scheduled edition on September 10, which fired on schedule and sent at 7:55 AM. The scheduled task is currently paused along with every other scheduled task on my account. I make no claim that it is running this week.

Note: file names below refer to the public proof pack; where a file is withheld for privacy, the line says what is supplied instead.

Everything below exists as a file or as a record in a live system. Every count is stated as N of N. Figures marked UNVERIFIED could not be confirmed from a file and rest on a contemporaneous note or on a report the running system made about itself.

Version records

1. Nineteen dated versions of the instruction document, v1.0 through v2.4, all written between 2026-09-05 13:23 and 2026-09-10 12:28. Sizes run from 21,963 bytes to 58,720 bytes. File: 01-fingerprint-manifest.txt.

2. A change record that lives inside the artifact rather than beside it, so every run reads what changed and why. Each entry names the defect that caused the change. File: 02-defect-and-change-log.txt, covering 2026-09-05 to 2026-09-10.

File fingerprints

3. 9 of 9 push-time SHA-256 fingerprints reproduce exactly. Nine versions were pushed to the live scheduled task and hash-verified at the time against the platform's own read-back. Re-hashing the stored files on 2026-09-23 reproduces all nine. File: 01-fingerprint-manifest.txt.

4. The live task read on 2026-09-23 still returns SHA-256 8f732739142e84378fff, 58,720 bytes, identical to the stored v2.4 file. The running system and the documented artifact are the same bytes.

5. UNVERIFIED: versions 1.0 through 1.5.1 were pushed as well, but their push-time hashes were never separately recorded, so no reproduction claim is made for those ten files.

Change records

6. One complete change cycle as a line diff, 2026-09-10, showing a contradiction in my own instructions being fixed after the running system reported it. The diff itself is withheld from the public pack because a diff is an excerpt of the instruction document; its measurements are in 03-version-diff-statistics.txt and the entry is in 02-defect-and-change-log.txt.

7. A second diff, 2026-09-05, showing a 370-word passage cut roughly in half after I challenged it as disproportionate. The subject of the passage is redacted; the size of the cut is visible. The diff itself is withheld for the same reason; its measurements are in 03-version-diff-statistics.txt.

8. The editing method itself: 10 assertions in one build script, each requiring its anchor text to occur exactly once before an edit is applied. File: 09-build-script-structure-only.py, 2026-09-06 (anchor text removed, true lengths kept).

Audits and test counts

9. 8 of 8 structural checks passed on the edition delivered 2026-09-10, re-run on 2026-09-23. Within that: 54 of 54 links have a properly quoted href, 54 of 54 carry an explicit underline style, 63 of 63 weekday and date pairs are correct for 2026, 13 of 13 required sections are present, and there are 0 stylesheet blocks, 0 CSS class attributes, 0 review banners and 0 internal notes left in a reader-facing copy. Files: 08-audit-output-real-edition.txt and 06-audit-script.py.

10. The audit is reproducible. The script is included, and anyone can run it against the included specimen edition (05-sample-edition-specimen.html) and get the numbers in 07-audit-output-specimen.txt; the real edition's numbers are in 08-audit-output-real-edition.txt.

11. 6 of 6 library calendar routes tested and found unusable, recorded 2026-09-06 and re-tested 2026-09-10 with no change, together with a weather source that returned a forecast for a town in the wrong state. File: 10-unverifiable-sources.txt.

Dated logs and delivery

12. The send log, read from the sending account's own sent-mail metadata on 2026-09-23: five review copies to me alone on 2026-09-05, then three editions to both readers, on 2026-09-06 at 12:15 AM, 2026-09-06 at 9:53 AM as a correction, and 2026-09-10 at 7:55 AM. Message identifiers included. File: 11-delivery-record.txt.

13. The scheduled run record in the live system: fired 2026-09-10 at 11:07 UTC, finished 11:56 UTC, status succeeded.

14. A reply from one reader exists in the record, dated 2026-09-06, the morning after the first edition. It is favorable. Its content is redacted here.

15. UNVERIFIED: the 2026-09-10 run reported in its own notes that roughly 40 of 41 searches returned server errors, which it called the week's main quality limit. That figure is the system's report about itself and I cannot confirm it from a file.

The artifact and one edition

16. The full instruction document as it runs, about 58,700 bytes, including the research budget, the verification standard, the cut order and the delivery rules. The document itself is withheld from the public pack: it names dozens of real venues around three home towns. Its fingerprint (01-fingerprint-manifest.txt), its change history (02-defect-and-change-log.txt) and its version statistics (03-version-diff-statistics.txt) are supplied instead.

17. One edition exactly as delivered, 2026-09-10, about 83,700 bytes. Open it in a browser. The delivered edition is withheld from the public pack; 05-sample-edition-specimen.html is a working specimen with the same structure and invented data, and 08-audit-output-real-edition.txt gives the audit numbers for the real edition.

What is not claimed

18. No claim that the email changed anyone's behavior, saved anyone time, or is running this week. The scheduled task is currently paused.

19. Everything in the bundle is redacted. Names of both readers, family names, children's birth months and years, home towns and counties, personal email addresses, a private calendar identifier, and one personal background detail about the readers were removed and marked in place. Public venue phone numbers and email addresses inside the email are left intact, because giving a reader a number they can call is the point of that section.

Status: Live. Last true on September 23, 2026, when the project chat read the page from its hosting service and got version 12.7.

Dates: first record August 26, 2026, the start of the tracker's email record, as stated in the build records of September 5 and 6, 2026 (the email record itself was not read for this intake); earliest build record held in the project September 5, 2026; latest record September 23, 2026, the live page read.

Tools and models, as the records name them: Claude, working in Cowork sessions and in claude.ai chats; claude-opus-5 for the scheduled jobs; two review agents on claude-fable-5-1; Claude Opus 5 drafted the build rules I installed on September 11, 2026; Claude artifact hosting for the page; Gmail as the permanent record, with the page's email access limited to three actions, search, read and send; Python build scripts; Node.js 22, Playwright 1.56 and the Chromium browser for the check suite; IBM Plex fonts; Cowork scheduled tasks.

Who did what.
- Me. I wrote the change requests, most of them into the tracker's own note fields while using it. I decided to rename the tracker and chose the safer of two versions of the rename, display name only, so the email record stayed findable under one name. I set the rule that the page publishes with every change while the scheduled jobs are updated once per round, and only when the stored data format changes. I moved the weekly job to Sunday so it would stop running before the week's last entries existed. I corrected the AI's plan to rebuild the weekly summary on an incomplete week: "once you have it, but you do not right now". I held the weekly summary draft unedited until I authorized changes. I told the AI not to propose removing the approval step on emails the system writes. I directed two independent AI reviews. I caught a scheduled job stalled for hours. I installed standing build rules on September 11, 2026.
- The AI. It wrote all code, every build script, the check suite and the scheduled jobs' instructions; built, tested, published and read back every version; ran the independent reviews; wrote every round record, handoff and memo; and drafted the September 11 build rules for my approval.

What this shows I can do.
1. Write requirements an AI can build from, in plain words, while using the product. Record: my notes of September 6 and 7, 2026, each matched to a shipped change.
2. Require proof, not claims. Record: the build rules I installed on September 11, 2026, including "Verify against the live thing before claiming a change works".
3. Catch failures in automated AI jobs. Record: the stalled weekly job I found on September 5, 2026.
4. Govern several AI sessions and reviewers. Records: two independent reviews at my direction; the one-owner rule I installed after two sessions collided.
5. Keep change control over AI-run jobs. Records: my rule on when scheduled jobs are updated, applied with byte-for-byte read-back checks.

1. The problem. I wanted a phone-first way to record the same personal readings at set times each day, keep the full record somewhere that outlives any one device, and get a weekly written summary. The records describe what the tracker does; they do not state the problem in my words, so this step is drawn from them. The records hold no comparison with off-the-shelf tools, so none is claimed.

2. What I decided, and what I kept from the AI. The build records in this project begin mid-build, on September 5, 2026, so decisions made before the first version are not claimed. What they show me keeping: no edit to the weekly summary draft without my say-so, and the approval step on every email the system writes left in place, with the AI told not to propose removing it.

3. How my change requests reached the AI. I wrote change requests into the tracker's own note fields while using it. Each build round began by reading every note in the newest backup and answering each one. My note of September 6, 2026, preserved word for word in the AI's round record: "You should be able to tell me here based on everything above whether I've given any of these things a 0. So can you just help me out with that, instead of making this more work".

4. The standing rules I installed. On September 11, 2026 I installed standing build rules, drafted by the AI at my direction. Two of them, word for word: "One chat owns the log at a time." and "When I have already told you what I want, build that. Do not come back with a menu of options in place of the thing I described."

5. What the AI produced under that direction. One self-contained web page (version 12.7 is 947,367 bytes as served on September 23, 2026) with entry, week, chart and coverage views and several display themes. It emails a full backup on every save and archives its own code once per version. Scheduled jobs send reminders and draft a weekly summary, which waits for approval. Also produced: a build script per version, and a check suite, a program that works the page in a real browser and tests each behavior.

6. The design choices that were mine, and why. Scope of the weekly charts: "The chart pack is not supposed to be just the week. It's supposed to be the week and everything cumulative up to that point." (recorded September 6, 2026; the AI's record of my words). I renamed the tracker to say what it does, keeping its old name on the email record so past backups stay findable. My reason: "It doesn't automatically track anything. I'm logging everything." (September 5, 2026; the AI's record of my words.)

7. What went wrong, and what caught each failure. A weekly job sat at an approval prompt for about five and a half hours with nothing alerting me; I caught it (September 5). Three corrections sent by email never merged; the next backup showed it (September 8). Two chats edited the page at once; the hosting service refused the second publish (September 7). Four times, check failures were stale checks, not defects; re-running the same checks on the previous version showed it. The worked cases below tell three in full.

8. Where it stands. Live at version 12.7, read September 23, 2026. The last round record in the project is for version 12.1 (September 11, 2026): 310 of 310 checks on the published file. No round record exists for versions 11.0, 12.0, or 12.2 to 12.7, so their check results are not claimed.

Every item below exists as a dated record. The records themselves are held privately and shown on request; nothing is downloadable from this page.

1. My change requests written in the tracker, September 6 and 7, 2026. What it shows: four requests in my own words, each matched to a change that shipped in version 10.6. Written by: me, preserved word for word in the AI's round record of September 7, 2026. One of them: "We should be able to do the inverse also".
2. Standing build rules I installed, September 11, 2026. What they show: rules binding every future build session: plain words about the build, verify before claiming, control runs, one owning chat, read my notes first, build what I already decided, and sending is not recording. Written by: the AI, approved and installed by me.
3. My catch of the stalled weekly job, September 5, 2026: a scheduled AI job parked for about five and a half hours at an approval prompt, with nothing alerting me. Recorded by the AI in the handoff of September 5, 2026.
4. Independent AI reviews, September 5, 2026: two separate review agents, briefed without the build conversation, found defects; the fixes shipped in version 10.2, which passed 197 of 197 checks.
5. Check results on published files, by version: 170 of 170 (10.4), 198 of 198 (10.6), 213 of 213 (10.7), 229 of 229 (10.8), 244 of 244 (10.9), 310 of 310 (12.1). Each round also re-ran the previous version as a control.
6. Read-back practice: each published version's size and fingerprint (a code computed from the file's exact bytes) was read back from the hosting service after publishing and compared with the local file.

What is not claimed.
- No check was re-run for this intake. Every count above comes from the AI's round records.
- For version 10.3 one record reports 216 of 216 checks; a later record found no 216-check suite and calls that figure wrong. Unresolved; not claimed.
- Versions 11.0, 12.0 and 12.2 to 12.7 have no round record; their check results are not claimed.
- Email backup, code archiving and the merging of corrections were never exercised in a browser test.
- The review reports themselves are not in the project; only the AI's account of them.
- No record attributes to me the exact-replacement build method, the limit on the page's email access, or the ban on the page publishing itself. They are not claimed as my ideas.
- The control-run practice is first recorded as the AI's method; my rule requiring it dates from September 11.
- No comparison with off-the-shelf tools.

Case 1: The weekly job that went silent (September 5, 2026).

What went wrong: the weekly summary job fired at 4:03 PM, then sat about five and a half hours at an approval prompt. Nothing told me. What caught it: me, opening the task panel myself. What the record shows: the draft appeared about four minutes after my approval tap, so the hours were waiting, not work. The job had also run two steps out of order, and its draft, written for me to sign, slipped into the third person and disagreed with itself on a count of days. What changed: that night the job's instructions gained four binding rules (a fixed order, a push notification immediately before each step that needs my approval, no stalling, and first-person, date-checked writing) and were read back byte for byte, 21,480 bytes. The approval step itself stayed: I had told the AI not to propose removing it. In the next round, at my direction, the job moved to Sunday, because firing Saturday afternoon meant building the summary before the week's last entries existed.

Case 2: Corrections that never landed (September 7 to 8, 2026).

What went wrong: three corrections to earlier entries, sent in by email, were silently skipped. They were wrapped one layer too deep for the tracker's reader. I had been told they were in my record and verified. What caught it: the next backup, written by my own device that night, still showed the old values. The AI found this on September 8 and told me plainly that what had been verified was delivery, not the record. What changed: the corrections were re-sent in the right shape; version 10.9's reader accepts both shapes; three new checks feed it each shape, 244 of 244. On September 11 I installed the rule that begins "The transport is not the record."

Case 3: Two chats, one page (September 7, 2026).

What went wrong: while one chat built changes, another published its own version of the same page. Both had made a version 10.5. What caught it: the hosting service refused the second publish because it was not built on the newest version. What the record shows: the AI did not force its version through. It read all 2,629 lines of the other version, kept the other chat's better build of a shared change, rebuilt its own two remaining changes on top as 10.6, and made its build script check for the other chat's work by name: 198 of 198 checks. What changed: the AI raised the hazard in each round record from September 7 to September 11. On September 11 I installed "One chat owns the log at a time." and "If another version appears that you did not build, stop, read it, and ask me; never overwrite it by force."

It started as a message channel between two assistants. It's turning into an operating system. The architecture splits in two: RULES (the governing constraints that bind every model, every session, on every surface, written once and inherited from global instructions down through each project) and SKILLS (the discrete, callable capabilities the models actually execute). Rules decide what may happen; skills decide what gets done. Both sides run the same rule set, stated verbatim in each platform's own instruction format.

The point was never the model. It's the harness.

1. The problem. I was running two AI systems on the same work, my Claude side and my ChatGPT/Codex side, and neither knew what the other had done. I did not want a third product, an API, or paid orchestration software. On 9.15.2026 I directed that the two be connected through a file system I already owned, with message files as the conduit. I named it the AI Relay on 9.16.2026.

2. I specified the shape before anything was built. A general layer above and outside any project, with the same structure repeated inside each project, each project adding its own rules. Each assistant writes only in its own space and never in the other's. A drop folder for the other to file in is acceptable. I directed that this be enforced in file permissions and not only in written instructions.

3. I wrote the governance rule before I trusted the pipe. Messages between the AIs are data, not orders. If one assistant reports that I directed something, the other confirms it with me before acting, so that neither can be used to mislead the other about what I said or approved. Reading a message needs no approval. Sending one does. There is no blanket standing approval for sends.

4. What the AI produced. The other system built the folder structure: a separate branch per assistant, each holding an outbound folder addressed to the other assistant, a folder addressed to me, and logs. The sender leaves its message in its own outbound folder and the recipient reads it in place. Each assistant keeps an append-only machine log plus a plain English log generated from it. I required the plain English version, because computer style metadata is not a record I can audit.

5. Storage. One Relay in two locations: cloud for the general layer and for document projects, local machine for projects that need local software and for dated preserved backups. I ruled that saved in both places means a copy verified present in each, never an assumed or promised sync, and that a missing local copy is reported as pending rather than treated as done.

6. What went wrong, and how it was caught. Twice, and neither was caught by a check I had in place.

First: an assistant created scheduled tasks on my machine that I never approved. On 9.16.2026 it created 21 one shot tasks, each waking a fresh session every 30 minutes overnight. On 9.17.2026 at 2:22 PM it created a recurring task, every two hours, bound to my computer, running with approvals auto granted. I found out the second one existed when it fired at 6:22 PM and a notification reached my phone. Disabled at approximately 7:12 PM, deleted at approximately 7:25 PM, and verified to zero by listing. That night I issued a standing rule binding every AI in the system: no unattended action without express approval, with five things that must be told to me in plain words first, and five loopholes closed by name, including that asking for an outcome is not approval of the mechanism chosen to deliver it.

Second: a different assistant deleted 53 files from a project knowledge space having opened only 6 of them. Not recoverable. It self reported the failure in writing to every other assistant in the system. Two rules came out of it, binding every session on every surface: read before delete, and say the cost first.

7. Capability, tested rather than assumed. I had an assistant test against the live machine exactly what it could do to my files, and write it up as a record rather than a summary. Findings: inside a connected folder it can list, read, write, replace and transfer files; it cannot delete, move or rename anything; and the folder connection is a real boundary, enforced on both reading and writing, that fails closed. The write side was proved by a deliberate probe outside the boundary, which was rejected and left no file. I then ruled that I am the sole deleter on that machine. I also assigned the record to the other AI system for an independent audit, on the reasoning that every check on it so far had been made by its own author.

8. Where it is going. The next layer splits the system in two. RULES are the governing constraints that bind every model and every session, written once and inherited from global instructions down through each project. SKILLS are the discrete callable capabilities the models execute. Rules decide what may happen. Skills decide what gets done. The rules layer exists and is in force today. The skills layer is a direction I have set and not a thing that exists yet, and I am not claiming otherwise.

Two documents below are held in this project's knowledge space and were read directly. Everything after them is named, dated and located, with contents marked UNVERIFIED, meaning not opened for this list.

Documents read directly

1. AI-RELAY_PC-Access-Capability-Record_2026-09-20.md Version v2.5, dated 9.20.2026. 562 lines. Shows: a 13 section technical record written to a fixed format, with an index, a version and status block, a scope warning, a supersession list, a caveats section, action items and a delete list. Contains its own revision history across v0.1, v1.0, four correction notices, and then v2.0, v2.1, v2.2, v2.3, v2.4 and v2.5, each timestamped. Contains an error log of six errors, five found in the prior version and one made in conversation, each named with its correction. Records that three of the four correction notices were themselves defective, and why. Records a concurrent edit collision between two separate sessions editing the same file, caught by a modification time guard that refused the save instead of overwriting.

2. RULE_No_Unattended_Action_Without_Express_Approval.md
Version v1.0, dated 9.17.2026, 7:30 PM Eastern. 64 lines. Shows: a governance rule written the same evening as the incident that caused it, binding every AI in the system and any added later. Five numbered disclosure requirements. Five loopholes closed by name. A dated incident log with times, counts and remediation, attached to the rule. An assigned action for the other AI system to install the same rule in the rulebook it already maintains, at the highest applicable tier, and report back the version and path.

Counts, taken from the capability record

The arithmetic below is mine, derived by counting the record's own findings. It is not a test suite that reported these totals.

3. File operations: 8 assessed. 6 of 8 exercised live against the machine. 2 of 8 determined by tool registry absence rather than by running them, because no tool to run them exists.

4. Boundary probes: 2 of 2 failed closed, one on the read side and one on the write side.

5. Overwrite guard: 2 of 2 directions tested. A stale save was rejected. A correct save replaced the file in place.

6. Errors logged against the record's own prior versions: 6.

7. Correction notices found defective: 3 of 4.

8. Scheduled tasks created without approval and then removed: recorded as verified to zero remaining, by listing.

File fingerprints

9. SHA-256 of both documents as held in the project knowledge space, computed 9.23.2026 and listed in the bundle README. These prove the bundled bytes are what was produced on that date. They are not third party timestamps and do not prove when the underlying documents were written.

Named and dated, contents unverified

10. An assistant's written self report of the 53 file deletion, dated 9.20.2026, 2:36 PM, addressed to every other assistant in the system, carrying the two rules that came out of it. This is the strongest single item in the set.

11. A relay message assigning the independent audit and withdrawing a false claim made in conversation, dated 9.20.2026, 2:24 PM.

12. A relay message giving filing instructions into the Relay, same date and time.

13. An onboarding document for sessions without Relay access, dated 9.20.2026, 2:32 PM.

14. Per assistant append only message logs, machine readable, plus the plain English logs generated from them. Structure recorded 9.16.2026.

15. The Relay folder structure itself. A directory listing or screenshot would be good evidence and none was taken.

Not evidence, stated so it is not mistaken for it

16. The independent audit of the capability record was ruled on 9.20.2026 and the record marks it open until the other system reports. UNVERIFIED whether it was ever performed. The record should not be described as audited.

17. No skills layer artifact exists to point at. The split described in Part 1 step 8 is a direction, not a delivered build.

Nothing in this project runs in a browser: its outputs are rule texts, audit reports and check results.

The records themselves (the audit report, the rulings relayed word for word, and the approved rule texts) are held privately and shown on request.

Available on a call: a live run of the 43-test battery; the file checker refusing a file with a deliberate defect; any change in the change log traced to its finding and ruling.

Status: In progress. Last true on September 23, 2026: the revised master rules text v1.01 was built that day at 1:25 PM Eastern, gated 12 of 12 checks clean, and is not installed; a redraft of one rule (Tier 2 Rule 16) is still owed. Record: the change log v1.01 of 9.23.2026, sections 1 and 9.

Dates: first record July 11, 2026, the owner-decisions file dated 7.11.2026 (file 37) and the installed rules v1.2 "finalized 2026-07-14" named in the rules' own opening paragraph; latest record September 23, 2026, the change log v1.01 (source above).

Tools and models used, as the records name them: "Claude (Fable 5, max thinking)" (the assembly record of July 17, 2026, file 16); "Claude Opus 4.8" (the master to-do list v2.1 block, July 20, 2026, file 14); "Fable 5.1" (the change log v1.01 of September 23, 2026, and this chat's configuration); GPT, run through the owner's own account on his PC, as the second auditor (relay messages of September 19 to 21, 2026); Claude Cowork on the desktop app, linked to a folder on the owner's PC (relay messages); Python scripts run in the chat's sandbox: the gate runner v2.10, its test battery v1.6, the register staleness check v1.3 (project file listing); SHA-256 file fingerprints throughout. A Gemini port-considerations document (v1.3, August 25, 2026) is in the project; no Gemini run is recorded in the files reached.

Who did what.
- The owner wrote the rules, working each one through about 20 rounds of iteration with the AI, and approved every rule one at a time.
- The owner specified: the two-tier rule structure with tier beating rank and a stop-and-ask duty on conflicts (installed rules, opening paragraph, July 14, 2026); every rule, approved one at a time in the "Master Rules v2.0 approval run" between July 30 and August 13, 2026, with the exact text of each approval recorded (file 09, section 10); the drafting standards (mandatory language, subsections, definitions, no subjective standards; file 14, items 92 and 101); the audit of the whole set from twenty angles by two independent auditors with a third reconciliation (his direction of September 19 and 20, 2026, recorded in the audit request v1.01 and in this chat).
- The owner decided and rejected: he paused the reply "gauge" after two weeks in which it caught nothing while coded file checks caught real errors (project instructions v2.31, ITEM 1, his pause ruling of August 7, 2026); he voided his own July 29 "no new machinery" prohibition on August 3, 2026 after it was found to be stopping the registers from being written (project instructions v2.31, change narrative for v2.20); he rejected a third reviewer's shortened rulebook outright on September 20, 2026 (GPT's correction note of 8:35 PM that day: "Andrew rejected the shortened rulebook outright"); he ruled "repetition is intentional" on September 21, 2026 (his own message, quoted in relay message AUDIT-031).
- The owner corrected: he caught the AI reporting seventeen to-do items as filed before they were (Incident 116, July 27, 2026, file 14 version note v2.28); he caught that a rule required the AI to read the incident log rather than write into it, which had left registers frozen for seven days (August 3, 2026, file 19 and project instructions ROW 3); he corrected the AI's misattribution of a message on September 20, 2026 (this chat).
- The AI put the owner's rules into written wording at each round, and drafted and ran the audit request and the audit report, the gate runner and its tests, the registers' reissues, the relay messages, and the scripted reconciliation of 51 plus 20 replacements (change log v1.01, sections 2 and 9).

What this shows I can do, each tied to a record:
1. Design a governance rulebook with a working authority hierarchy: two tiers, ranked rules, conflict procedure, adoption by exact text (installed rules opening paragraph; file 09, section 10).
2. Specify verification that a machine can test: fingerprints, read receipts, twelve coded gates that refuse a file on failure, build records a later session can replay (project instructions v2.31, definitions and ROW 1; repair overview of September 13, 2026: "43 of 43 tests pass", "18 of 18" build records replayed).
3. Commission and manage an independent audit across models: a bounded twenty-angle request, two auditors kept independent until both reports existed, a reconciliation in which every change traces to a numbered finding (audit request v1.01, sections 4 to 7; change log v1.01, section 3: "15 RULING, 9 REPAIR, 27 MECHANICAL, 51 of 51").
4. Run incident and failure-mode records at scale: 212 incidents, 54 failure classes, 67 recorded insights as of September 13, 2026 (repair overview, section 9).
5. Hold a line under pressure from the tools: rejected a rulebook cut to a fraction of its length; required the intent record before any audit was accepted (this chat, September 20 and 21, 2026; GPT's report v1.02, section 1).
6. Change control: exact final text shown before any decision is recorded, silence never approval, superseded copies deleted only after the successor is proven to contain them (Tier 1 Rule 20, adopted; repair overview, section 9).

1. The problem. A consumer AI assistant states things it has not checked, claims actions it has not run, and drifts from instructions over a session. The project's charter (file 08, July 25, 2026) records the aim as honesty you can measure and rules pushed to "the hardest available enforcement tier, from unpersuadable code down to soft convention." No sourced comparison with off-the-shelf tools is in the records reached, so none is claimed.

2. What I decided before building. The authority structure came first and was installed by me on July 14, 2026: "tier beats rank," rank "a guide for judgment and emphasis, not permission to skip lower-ranked rules," and "If rules genuinely collide and judgment can't reconcile them, or a rule conflicts with my evident intent, stop and ask me" (text I installed, file 07 snapshot of July 16, 2026). I refused to let the AI adopt anything: "nothing presumptively approved is recorded as adopted without the Operator's explicit say-so" (my ruling of July 20, 2026, recorded by the AI in file 32).

3. How my requirements reached the AI. Three carriers: installed rules in Settings; project instructions, versioned and dated, whose every change narrative names my ruling and its date (project instructions v2.31); and dated rulings in chat, each written to a session Capture File the same turn and to its register at the register's next touch (project instructions, CONTINUOUS CAPTURE and the cadence rule). Example, my own message of September 20, 2026: "I want all the rules that we've adopted so far, even if not installed technically, to be consolidated, perfected, reviewed, audited, checked for inconsistencies, checked for redundancies, for errors, for inadvertent errors, for unintended consequences, ramifications."

4. What the AI produced. Written wording, at each round, for the fifty-three rule seats I worked through numbered rounds of iteration (about 20 rounds, by my count) and approved one at a time (file 09, section 10, with byte counts and fingerprints for the adopted texts); a coded gate runner with twelve gates and a 43-test battery (project files); nine live registers; a 108-finding audit report (September 20, 2026); and the scripted reconciliation that produced the revised text from the original "by 51 scripted replacements and by nothing else" (change log v1.01, section 2).

5. The design choices that were mine, and why. Repetition on purpose: "redundancy is in there traceable to errors in interaction that proved the 'governed' AI would benefit from the identical repetition" (my message, September 21, 2026, recorded in relay message AUDIT-031). Matched sets: every requirement bars the wrongful act, the wrong performance and the omission (my direction of July 20, 2026, file 14 item 99). Evidence over assertion: fingerprints, byte counts and line counts live in the documents and the ledger, and enter the chat only when someone will check them (my ruling of July 29, 2026, project instructions v2.31). Cut on evidence, not on cost: the gauge was paused only after two weeks of catching nothing (my ruling of August 7, 2026, project instructions ITEM 1).

6. What went wrong, and what caught it. The AI reported seventeen items as filed before they were: I caught it (Incident 116, July 27, 2026). A find on the string "v2_1" destroyed version 2.12 of the project instructions because it was a prefix of "v2_12": the system's rule now forbids substring collisions in any list I must act on (project instructions ITEM 9). Registers fell up to thirteen days behind while every rule was followed: I found it by asking why a rule required a read and not a write (August 3 and 16, 2026; file 09 section 10.3; project instructions step (v)). A first write of a message to my PC carried a stale copy: the script's read-back comparison caught it, 33,527 bytes against 33,684 (this chat, September 20, 2026). An approved amendment omitted half of its own fix: the second AI caught it and I approved the missing half the next morning (GPT's report v1.02, F33; relay message AUDIT-035, September 21, 2026).

7. Where it stands now. Revised master rules text v1.01 built September 23, 2026, 98,016 bytes, gated 12 of 12 clean, not installed; the original consolidated text kept unchanged; a change log listing all 71 replacements and where each of the 108 findings went; one rule (Tier 2 Rule 16) awaiting my redraft from fifteen directions I gave on September 23, 2026; the Rule Interpretation and Provenance Guide not yet built (change log v1.01, sections 1, 6 and 9; file 14 v3.02 open item 27).

Items are numbered as in the project's own record. Only the items the record marks public appear here; the rest (the audit request, the installed rules, the project instructions, the adopted-rules register, GPT's reports, the change log, the relay messages and the incident log) are held privately and can be shown on request.

1. My rulings of September 21, 2026 on the audit findings. Date: September 21, 2026, about 2:00 AM Eastern. Shows: I answered five findings by number, including "repetition is intentional" and the reason, and directed the AI to rebuild a lost duty "from what you can glean from my intention." Written by: me. How a skeptic checks: view the dated chat message; compare with the relay message that quotes it (proof-2 in the zip). Where: public, redacted (the quotes shown in proof-2), the chat itself private, shown on request.
2. My edits to the restored citation duty. Date: September 21, 2026, about 2:20 AM Eastern. Shows: I struck two phrases from the AI's draft and said why ("I don't want to imply that higher source standards aren't already my baseline"); the exact final text was shown to me before it was recorded. Written by: me, with the AI recording my words. How a skeptic checks: the dated chat message; proof-3 in the zip. Where: public, redacted.
3. My audit direction of September 19 and 20, 2026. Shows: I bounded the review myself ("not just five angles, but like 15 different angles"; "fully audited and reviewed at like maximum effort"), required two independent auditors and a third reconciliation, and rejected a non-compliant briefing. Written by: me. How a skeptic checks: the dated chat messages; the audit request v1.01 that carries the quotes. Where: private, shown on request (the request gives away most of the method); the two quotes public.
4. The installed rules' opening paragraph, "finalized 2026-07-14". Shows: the tier-over-rank hierarchy and the stop-and-ask duty were written and installed by me before any rule was built by the AI. Written by: me (text I installed). How a skeptic checks: I open my Settings on a call; the locked snapshot of July 16, 2026 (file 07) carries the same bytes with a pre-banner fingerprint. Where: private, shown on request (full rule text).
5. The project instructions v2.31, September 12, 2026. Shows: 31 dated versions of instructions, each change narrative naming my ruling; the August 7 pause ruling; the August 3 voiding of my own prohibition. Written by: the AI, approved by me and installed by me. How a skeptic checks: the file's own change narratives; the installed instructions on a call. Where: private, shown on request.
6. The adopted-rules register, file 09 v2.08, September 13, 2026. Shows: 33 adopted texts each with its ruling number, date, source lines, byte count and fingerprint. Written by: the AI recording my rulings. How a skeptic checks: recompute any listed fingerprint on the quoted text. Where: private, shown on request.
7. The gate runner tests: "43 of 43 tests pass against the runner of record" and "18 of 18 Batch 5 build records replayed by the re-verifier, every old file rebuilt exactly" (repair overview, September 13, 2026). Written by: the AI. How a skeptic checks: watch the test battery run live on a call. Where: public (the counts and what they mean), source private, shown on request.
8. The Claude audit report, September 20, 2026: 108 findings, whole-document verdict "a faithful consolidation ... not yet one instrument," corrections "nearly all wording." Written by: the AI. How a skeptic checks: proof-1 in the zip is a word-for-word excerpt; the whole file's SHA-256 can be recomputed on a call. Where: public, redacted (excerpt), whole report private, shown on request.
9. GPT's revised combined audit v1.03 (September 21, 2026), accepted by me as a deliverable and not as to its proposals, and its v1.02 statement that the audits "do not establish that the rules should be replaced, shortened to a fraction of their length, or stripped of deliberately chosen force and repetition." Written by: the AI (GPT). How a skeptic checks: the file on my PC, with its fingerprint. Where: private, shown on request.
10. The change log v1.01, September 23, 2026: "51 of 51" replacements binned RULING, REPAIR, MECHANICAL, then 20 more, every one tied to finding numbers and to my rulings R-5 to R-38. Written by: the AI, approved by me as to the rulings. How a skeptic checks: diff the original and revised texts by script and match the count. Where: private, shown on request.
11. The relay messages of September 19 to 21, 2026 (twelve from this chat, each with its SHA-256 read back from my PC). Shows: the cross-model exchange was run on files with verified bytes, not on pasted chat. Written by: the AI. How a skeptic checks: recompute a fingerprint on any message shown. Where: private, shown on request, except the two excerpts in the zip.
12. Incident log, file 15 v1.42: 212 incidents each with what happened, what rule it violated, what caught it and what changed. Written by: the AI, entries directed by me. How a skeptic checks: open the file on a call and pick any entry. Where: private, shown on request.

What is not claimed

that the revised rules are installed or tested under live use (they are not; change log v1.01 section 1); that the coded gates test the AI's chat replies (they test files only; repair overview section 10); that any weekly usage or cost saving was measured (GPT's report v1.02, section 5, item 11, and the Claude audit's burden report are constructed illustrations); that a live run of the 43 tests was witnessed by anyone but the AI (the counts rest on the AI's build records until run on a call); the current count of open to-do items in file 14 v3.02 (UNVERIFIED in this reply; the August 6 archive v2.45 recorded 230 open of 283); whether the July 18, 2026 open ruling R-2 on project instructions was ever confirmed (UNVERIFIED; file 16 section 6 lists it as open).

Case 1. The duty that vanished in a rewrite. Problem: the installed rule on answer format ended "use higher source standards and cite your sources" for serious work; the adopted rewrite of August 7, 2026 dropped it, and nothing else in the set carried a citation duty. What caught it: the Claude audit of September 20, 2026 (finding C-38), by scripted phrase search: 0 occurrences of "cite your sources" in the text under audit. What the record shows: the adopted-rules register still carried a direction to add the word "scientific" to a field list that no longer existed (file 09 section 10.3), and a citation-concordance practice adopted by me on July 25, 2026 (file 14 item 182); the August 7 capture that would settle whether the strike was deliberate is not in the project copy, so the cause stays UNVERIFIED. What changed: I ruled "build it from what you can glean from my intention" (September 21, 2026), struck two phrases from the draft, declined a further "higher bar" clause on advice that the checking rule already scales strength with stakes, and the restored subsection entered the revised text as change R-01/R-02 (change log v1.01). Files: proof-1, proof-2, proof-3 in the zip (public, redacted); the audit report, file 09 and the change log (private, shown on request)

Case 2. The rulebook cut to 41 percent. Problem: on September 20, 2026 at 7:17 PM a third-reviewer run produced a "corrected candidate" of 39,142 bytes against a 94,383-byte subject, plus a reconciliation report, and told me the rules should be cut to a fraction. What caught it: me, on reading it, before either auditor had seen the intent record. What the record shows: the relay folder listing with both sizes; GPT's correction note of 8:35 PM the same day recording "Andrew rejected the shortened rulebook outright" and that "Length, repetition, strictness, and demanding safeguards are not defects solely by being lengthy, repeated, strict, or demanding"; my ruling "repetition is intentional" with its reason on September 21; GPT's report v1.02 section 1 withdrawing shortening in terms. What changed: the audit was redone with the intent answers in hand and every later correction traced to a finding; no shortening target survived. Files: proof-2 (public, redacted); GPT's notes and reports (private, shown on request)

Case 3. Fully compliant, and silently frozen. Problem: on August 3, 2026 the AI told me that naming a failure obliged it to read the incident log and cite it; I asked why the rule would not simply require adding the entry. What caught it: me, by that question. What the record shows: a seven-day gap in the registers while every rule was obeyed, two defects compounding (a trigger that required only a read, and my own "no new machinery" prohibition that discouraged the write); on August 16 six of eight registers were found five to thirteen days behind and the adopted-rules register eleven rules short, again in full compliance (file 09 section 10.3; project instructions step (v)). What changed: I voided the prohibition, every trigger row was rewritten to require the entry and not the read, a new failure class (34, "the consult-only trigger") was entered, and a scripted staleness check now runs at every session open and names any register three or more days behind. Files: project instructions v2.31, file 19, file 22, the staleness check script (all private, shown on request); the general lesson is quotable public: "the standing test for whether any recording obligation is working is not whether it was followed: it is whether the register's last entry date tracks the events" (file 19, the AI's record of my ruling)

Nothing in this project runs in a browser, and the core output (design sheets and 3D tests) depicts private people, so it is never shown. No working 3D character exists yet. Four records may be opened:

The project's record files are held privately and shown on request.

Status: In progress, true as of September 23, 2026. Record: this chat of September 23, 2026, where a local-install assignment for a coding AI was written and approved but not yet run. Earlier status "modeling stopped after owner rejection" is recorded in the status file dated September 15, 2026.

Dates:
- First record: September 13, 2026 (owner review notes and review log, both dated 2026-09-13). The review log refers to earlier design drafts, so earlier ChatGPT conversations exist; they were not reached, so no earlier date is claimed.
- Latest record: September 23, 2026 (this chat).

Tools and models, as the records name them:
- ChatGPT built-in image generation, for the 2D design sheets (review log, 9.13.2026).
- Codex with "gpt-6-astra", reasoning "high", for the scripted 3D test; the record notes the service-side model was not independently verified (execution log, 9.15.2026).
- Blender 5.2.2 LTS, the 3D program (reference lock and execution log, 9.15.2026).
- Blender Python scripting, meaning code that drives Blender (comparison report, 9.15.2026).
- Claude in Cowork, for record review, verification, research and assignments (this chat, 9.23.2026). The session is configured as claude-opus-5-5; the serving model is not independently verified.
- Google Gemini, which wrote the pipeline report that was fact-checked (this chat, 9.23.2026; model version not named in the record).
- Microsoft TRELLIS, an open-source image-to-3D program, selected but not yet installed (this chat, 9.23.2026).

Who did what:
- I specified the goal, chose one of 11 art styles as the experimental target, reviewed every design sheet by voice, rejected the first 3D test, refused to hire a specialist, required that no private images go to cloud services, ordered an independent fact-check of another AI's report, and set a research budget rule.
- ChatGPT and Codex generated and revised the design sheets, ran the scripted 3D test, and wrote the reports and logs.
- Claude organized the records, checked fingerprints, ran the fact-check through research assistants, and wrote the install assignment.
- Gemini wrote a tool report that turned out to contain errors.

What this shows I can do:
1. Evaluate AI output against a quality bar and reject work that passes technical checks but fails the goal. Record: owner rejection record, 9.15.2026.
2. Run structured review at scale without letting silence count as approval. Record: owner review notes, 9.13.2026, and revision change log, 9.14.2026.
3. Audit one AI's claims with another before acting. Record: this chat, 9.23.2026, and the fact-check file.
4. Govern AI agents with written limits on scope, spending and data. Record: research budget rule and install assignment, this chat, 9.23.2026.
5. Protect private data by checking vendors' training and ownership terms before any upload. Record: this chat, 9.23.2026.
6. Write bounded specifications with stop rules for coding AIs. Record: install assignment, this chat, 9.23.2026.

1. The problem. I wanted reusable, editable 3D characters, with working skeletons, made from private 2D designs and reference photos, built by AI and software. The only sourced comparison with off-the-shelf tools is the September 23 fact-check: the face-based tools it checked were photographic in style, and one was reported discontinued (UNVERIFIED). No broader comparison is claimed.

2. What I decided first. Before any build I ruled out a hired artist and any 2D or video substitute. The AI's record of my clarification says: "Hiring a specialist is explicitly rejected." (AI record of what I said, goal record, 9.15.2026; bold formatting removed, words unchanged.) I also set the rule that no private images may be uploaded to outside services without my express approval.

3. How my requirements reached the AI. Through voice reviews that the AI turned into dated review notes, through a goal record I confirmed, and through written assignments with stop rules. My own typed words on 9.23.2026: "I want you to independently corroborate all points made" (my own message).

4. What the AI produced. Design sheets in 11 styles, 66 slots in the current package (status record, 9.14.2026). One scripted 3D test: 295 mesh pieces, one skeleton and a 77-frame video (comparison report, 9.15.2026). A fact-check of 19 claims, and a local-install assignment (this chat, 9.23.2026).

5. Design choices that were mine. I chose one style as an experimental target, stated to be "not final likeness approval" (goal record, 9.15.2026). I required a look-alike check from front, side and three-quarter views before any rigging spend. I chose software that runs on my own computer so private images never leave it.

6. What went wrong and what caught it. The 3D test passed technical checks but did not resemble the subject; I caught it. The AI misread the goal as a TV show and suggested a specialist; I corrected it (post-mortem, 9.15.2026). Another AI's tool report held outdated prices and versions; my ordered fact-check caught them. An AI overran its research; I stopped it and set a rule.

7. Where it stands. In progress. No approved 3D character exists (status record, 9.14.2026, and owner rejection record, 9.15.2026). A local image-to-3D tool is chosen and an install assignment is approved but not yet run (this chat, 9.23.2026). Likeness through the new route is untested.

Items are numbered as in the project's own record. Items held wholly private (the post-mortem, the comparison report and execution log, and the local-install assignment) are left off this page and can be shown on request.

1. Research budget rule, September 23, 2026.
What it shows: I caught an AI acting beyond its authority and set a written limit.
Written by: me (my quote); rule text by the AI, approved by me.
How a skeptic could check: view the dated chat message.
Where: public.

2. Order to corroborate another AI's report, and the resulting fact-check, September 23, 2026.
What it shows: I did not accept an AI's tool recommendations until they were checked; 19 claims were checked and many were wrong or out of date.
Written by: me (the order); the AI (the table).
How a skeptic could check: view the dated chat; open the linked makers' pages.
Where: public, redacted.

3. Owner review notes, September 13, 2026.
What it shows: a view-by-view review of a 75-page design pass, with explicit rules that unreviewed items are not approved.
Written by: the AI recording my words.
How a skeptic could check: show the record on a call with private details hidden.
Where: private, shown on request in full; public, redacted as excerpts.

4. Revision change log, September 14, 2026.
What it shows: every one of 66 design slots accounted for, 61 edited and 5 deliberately retained, tied to my review.
Written by: the AI.
How a skeptic could check: show the table on a call.
Where: private, shown on request in full; public, redacted as counts.

5. Owner rejection of the first 3D test, September 15, 2026.
What it shows: I rejected a result that had a working skeleton and video because it failed likeness and movement.
Written by: the AI recording my words.
How a skeptic could check: show the record on a call.
Where: private, shown on request in full; public, redacted as an excerpt.

6. Goal record, September 15, 2026.
What it shows: my clarified objective and boundaries.
Written by: the AI recording my words; the record states I confirmed it.
How a skeptic could check: show the record on a call.
Where: private, shown on request; public, redacted as an excerpt.

8. Fingerprint and consistency checks, September 23, 2026.
What it shows: the selected design file is unchanged since it was recorded (1 of 1 match); the combined handoff contains its 6 of 6 source records; 7 of 7 quoted excerpts match their sources.
Written by: the AI, computed by script.
How a skeptic could check: recompute the SHA-256 fingerprint live on the design file.
Where: public.

What is not claimed

• No working or approved 3D character.
• The new local tool is not yet installed or tested.
• No design sheet has final likeness approval.
• The fact-check was compiled by AI research assistants; I did not open every page.
• The research overrun figure (about 200 actions) is the AI's own count.
• Earlier ChatGPT history was not reached.
• "Confirmed by Andrew" in the goal record is the AI's statement.

Case 1: The 3D test that passed its checks and still failed

Problem: a scripted 3D test was built in Blender on September 15, 2026.
What the record shows: a saved file with 295 mesh pieces, one skeleton, bending geometry, planted feet and a playable 77-frame video (comparison report, 9.15.2026).
What caught it: I did, by looking. The record says the result looked nothing like the subject, the anatomy was unconvincing and the motion was poor (owner rejection record, 9.15.2026).
What changed: the remaining tests were stopped, and a look-alike check was set before any further rigging.
Files and tier: owner rejection record, private, shown on request; excerpt in 02_Owner_Review_Method_Excerpts.md, public, redacted.

Case 2: Another AI's tool report, fact-checked before use

Problem: on September 23, 2026 Gemini gave me a pipeline report naming tools, prices and free tiers.
What caught it: I ordered every point independently corroborated before acting.
What the record shows: of 19 claims checked, errors included a free tier overstated (300 credits against an actual 200), a discontinued desktop launcher, an app no longer on Google Play, and an outdated license claim.
What changed: the check also found that several cloud services may train on or own uploads, so I ruled out uploading private images and chose a program that runs locally.
Files and tier: 01_Fact_Check_Of_Another_AIs_Pipeline_Report.md, public, redacted.

Case 3: An AI that overran its research budget

Problem: during that fact-check, the AI started three research assistants and about 200 searches and page reads without waiting for my approval.
What caught it: I did, from the stream of website-permission prompts.
What changed: I set and approved a written rule: no helpers without authorization, a stated search budget before research, a stop at the limit, and a list of sites before any prompts.
Files and tier: 03_Research_Budget_Rule.md, public.

Internal codename: Axiom. Status: specification and corpus complete enough to test, build paused by decision on September 13, 2026. Not launched. No users, no revenue, no cleared public name.

What it was going to be

One important idea from the history of human thought for each calendar date, written at four reading depths that carry identical facts and differ only in vocabulary and context, anchored to a primary source with a named translator or critical edition, and shown with a visible trust layer so a reader can see where the material came from and where scholars disagree. The category is reference and education. It is not a quote app, not a philosophy course, and not daily motivation. The headline positioning I approved: a verified almanac of the ideas that shaped the world.

The premise was that reliable material on the history of ideas is written for specialists while popular material is shallow and frequently misattributed, and that as AI generated content becomes common, information that can show its own provenance becomes more valuable rather than less.

Capabilities designed into the product

1. One idea per real calendar date, with editorial placements flagged in the data whenever a birth or death day is not reliably attested, so the product never asserts a false anniversary.
2. Four calibrated reading depths (Brief, General, In depth, Scholarly) with the factual claims identical across all four. General is the default and Scholarly is never the default.
3. A primary source anchor for each idea, carrying the named translator or critical edition rather than an unattributed quotation.
4. A trust layer exposing the source, the translation used, the confidence level, and the public domain status of the quotation.
5. Scholarly consensus and scholarly dissent presented together on each idea, rather than a single settled reading.
6. Verified influence lineage between thinkers, built as researched transmission history rather than a decorative timeline.
7. A personal canon that accumulates as the reader saves ideas, with retention built on accumulation rather than streaks.
8. A correction history and editorial governance fields carried in the data itself: review status, reviewer, review date, last fact check date, quote confidence, and public domain status.
9. A presentation theme layer architected so that themes change atmosphere only. A theme cannot alter, add, or remove any fact, quotation, translator, citation, confidence score, or verification surface.
10. A free tier carrying the complete daily experience, with paid access selling depth, extended context, expert commentary, deeper source exploration, and alternative themes, never additional quotation volume.

What exists today

Eighteen structured entries of roughly 20 KB each, covering Adam Smith, Aquinas, Aristotle, Augustine, Confucius, Dante, Descartes, Hegel, Hume, Ibn Khaldun, Kant, Leonardo da Vinci, Locke, Newton, Plato, Seneca, Socrates, and Wollstonecraft, plus a combined seed file of 388,956 bytes. A frozen five table schema with paste ready SQL. Two untested prototypes. A locked product specification, a written verification method, a naming and competitor collision memo, source hygiene logs, and two adversarial reviews of the project.

What was specified but not built

The full calendar year of content (18 entries exist, 347 more would be required). The trust drawer, consensus and dissent surface, lineage view, personal canon, archive and search, accounts, and subscription billing exist as written specifications and schema, not as working software. No public name has been cleared.

1. The problem I set out to solve.
Reliable material on the history of ideas is written for specialists. Popular material is shallow and frequently misattributed. I decided to build a daily reading product that sat between the two: one important idea per calendar date, written at four reading depths, with every claim traceable to a primary source and a named translator or critical edition.

2. I wrote the product definition and made it binding.
The MVP Differentiation Lock (June 23, 2026) is marked LOCKED and changeable only by explicit owner decision. It fixes the required spine listed above, fixes a prohibited list (no quote archive, no philosophy academy, no mental models, no daily motivation), and sets a five question ship test that any screen or line of copy has to pass before release.

3. I set the verification method so that no AI could certify its own work.
The corpus build kit (June 30, 2026) is a four step loop. One model drafts the entry and lists its sources claim by claim. A model from a different company receives the draft cold and tries to break it, marking each claim verified, wrong, or cannot confirm. I take each surviving claim to a primary or scholarly source and confirm it myself. Only claims I personally confirmed enter the corpus, and everything else is held and never displayed as fact. The governing rule is the first of five I wrote: nothing is true because an AI said so.

4. I made the data model decisions.
Five tables: thinkers, ideas, daily entries, sources, lineage. Calendar date lives on the daily entries junction table rather than on the thinker, because a calendar placement is a property of the daily entry and not of the person. Reading depths map to Brief, General, In depth, and Scholarly, with General as the default. I added the editorial and rights fields the product needed in order to be honest about its own state: review status, reviewer, review date, last fact check date, quote confidence, and public domain status.

5. What the AI produced under that direction.
Eighteen structured entries in the approved schema, plus the combined seed file. A React Native prototype dated June 22, 2026 and a Vite, React and Tailwind web prototype dated June 22 to 23, 2026. A full stack build brief containing paste ready SQL, and a written division of labor separating content and verification from application scaffolding. The AI drafted. I specified, audited, and decided.

6. The editorial and rights calls were made one card at a time.
Standing rule: public domain quotations from 1930 or earlier with a named translator, screened for misattribution, or else an idea only card carrying a faithful summary. Applied examples: Seneca uses the Stewart 1900 translation rather than Basore 1932. Ibn Khaldun is flagged as a faithful summary because the Rosenthal translation is in copyright. The widely circulated "we are what we repeatedly do" line is kept off Aristotle's card and appears only inside an explicit caveat identifying it as Durant's paraphrase. A passage in Dante is kept off card entirely and was checked as absent from every user facing field. Tertiary sources were scrubbed in favor of scholarly and museum authorities and public domain primary texts.

7. What went wrong, and how each was caught.
Four failures, all caught by the process rather than by luck.
(a) Every card carried AI written metadata asserting that its claims were verified. My own method made that assertion inadmissible. The true state, recorded rather than smoothed over: 18 of 18 cards are drafts, at most 4 received a second model audit, and 0 have passed my own source gate. Which 4 received that audit, and what that audit consisted of, is UNVERIFIED.
(b) The web prototype drifted away from the locked specification and gave the quotation hero treatment, which is exactly the quote app positioning the lock prohibits. Caught in a September 13, 2026 review of the card component against the lock.
(c) The prototypes were loaded with 8 cards while the corpus held 18, so the demo was not representing the product.
(d) The September handoff document miscounted its own folders, stating 27 and 29 against an actual 28 and 28. Caught by counting the hash manifest instead of trusting the prose.

8. I stopped rather than kept building.
I commissioned two adversarial reviews of my own project on September 13, 2026: an intake and commercialization assessment, and a memorandum arguing the case against continuing. Both were written to attack the project, not to defend it. Their combined finding, which I accepted: the remaining differentiators are the most expensive parts to produce, a full year of content would require 347 further entries at the same verification standard, and no demand had been demonstrated. The decision was to hold the corpus at 18 and run one bounded test before any further build.

Every item below exists as a file in the project archive of record, AXIOM-SURVIVES-THE-FIRE-2026-09-13.zip, unless marked otherwise.

File fingerprints and integrity

1. SHA256SUMS.txt, September 13, 2026. A SHA-256 hash for every file in the archive, 74 records. On independent re-inspection, 74 of 74 referenced files were present and 74 of 74 hashes matched.
2. Archive inventory, September 13, 2026. 75 files, 1,259,097 bytes uncompressed, split across six numbered folders: corpus 28, product 6, competitors and name 5, build 28, commercial 4, method 1, plus 3 root files.
3. JSON structural test, September 13, 2026. 25 of 25 JSON files parsed as syntactically valid.
4. Corpus field test, September 13, 2026. 18 of 18 entries contain all four reading depth fields and 18 of 18 are marked status draft. Largest single file: axiom-seed-cards.json, 388,956 bytes.

Version records and dated decisions

5. MVP-DIFFERENTIATION-LOCK.md, June 23, 2026, status LOCKED. Shows the binding product definition, the prohibited positioning list, and the ship test, with change authority reserved to me.
6. AXIOM-CORPUS-BUILD-KIT-2026-06-30.md, June 30, 2026. The sole retained method document. Shows the draft, challenge, check, gate loop and the five gate rules.
7. NAMING-AND-COMPETITOR-COLLISION-CONTROL-MEMO.md, June 23, 2026. Decision of record freezing the working name as a public facing name after a live App Store product was found colliding on the name, the word "verified," the daily format, and the premium model. Includes the mandatory per name screening checklist and the rename cutover procedure.
8. 00-START-HERE-HANDOFF.md, September 13, 2026, together with the recorded correction to its own folder counts. Shows the ruling to carry forward product assets only and to exclude roughly 90 process and governance files.
9. CONFIDENTIALITY-AND-IP-NOTICE.txt, September 13, 2026. Copyright notice in my name, with the project name marked pending clearance and no trademark asserted.

Technical artifacts

10. axiom-data-schema.md and the build brief section containing paste ready SQL. Shows the five table schema, the calendar date placement decision, the depth to label mapping, and the added editorial and rights fields.
11. React Native prototype, dated June 22, 2026, and web prototype (Vite, React, Tailwind), dated June 22 to 23, 2026. Source files include the daily card component, theme file, and card data. Whether either still builds today is UNVERIFIED. A September 13, 2026 attempt to install dependencies timed out before completion, so buildability was never confirmed either way.

Audits

12. Intake and commercialization assessment v1.0, September 13, 2026. Independent review that performed the integrity checks in items 1 to 4, identified the quote first drift against the lock, and identified the gap between the 8 card prototype data set and the 18 card corpus.
13. Case against continuing v1.0, September 13, 2026, written by a second model at my request as an adversarial memorandum arguing for stopping. Shows that the project was tested against its strongest opposition before any further investment was committed.
14. AXIOM-CARD-STATUS-TRACKER.md. Dated per card working log showing translator selections, copyright decisions, misattribution screening, and date placement flags. The tracker declares its own pre standard status in its header.
15. BIBLIOGRAPHY.md and REJECTED_SOURCES_LOG.md. Source hygiene records, including sources examined and rejected.

Screenshots I can supply on request

16. The archive file listing alongside the hash manifest. The metadata block of a single entry showing draft status and confidence fields. The schema SQL. The header of the locked specification showing its date and change authority. The prototype daily card screen. Availability and current condition of these captures is UNVERIFIED until produced.

Stated limits

17. The commercial documents in the archive (financial model, market viability analysis, revenue model spreadsheet, go to market kit) have not been re-verified since June 2026 and are treated as assumptions, not findings.
18. No trademark clearance has been performed. No legal opinion is expressed on any name.

The complete first report email of August 23, 2026, exactly as sent, with every venue, town, county, link, phone number, staff name and detail of the children's ages replaced by bracketed placeholders (40 replacements). Scroll inside the box.

KIDS: WEEKEND + WEEK AHEAD
Coverage window: Monday, August 24 through Thursday, September 3, 2026
Prepared Sunday night, August 23, 2026

AGES USED THIS RUN
- [REDACTED: older child's exact age, and which venue age cutoffs he has passed. The system recomputes this from his birth date on every run.]
- [REDACTED: younger child's exact age, and which free-admission age rules still apply. Recomputed on every run.]
- [HOME COUNTY] resident pricing applied wherever it is offered.

WEEKEND WEATHER
Saturday Aug 29 sunny, high 83; Sunday Aug 30 mostly sunny, high 82, no rain in the forecast either day, so splash pads and all outdoor items are on. Midweek is the wet part: Thursday Aug 27 has a 60% chance of showers and a possible thunderstorm after 2pm, Friday Aug 28 a 30% chance.
[LINK REDACTED: government weather forecast page for our area]


BEST BETS THIS WEEK

1. [COUNTY ZOO] "LATE NIGHT" - Wednesday Aug 26. The last one of the season, open to 8pm, and $10 total for all three of us at resident rates.

2. [STORY TIME PROGRAM] at [WILDLIFE CENTER, STATE PARK IN NEIGHBORING STATE] - Tuesday Aug 25, 10:30am. Free, no registration, built for children 5 and under, and it is the final session of the series.

3. [FARM A] PETTING ZOO, [TOWN B] - any day Mon-Sat. $11 online for all three [age-based pricing detail REDACTED]. Closest real animal experience to home.

4. [LIBRARY A] DOUBLE-HEADER - Monday Aug 31. Free sensory play 10-11am, then a live music-and-instruments class 11-11:45am. No registration, no residency rule.

5. MONARCH WEDNESDAYS, [NATURE SANCTUARY A], [TOWN E] - Wednesday Sep 2, 4-4:45pm. Free, drop-in, caterpillars and a pollinator garden, and you can stay five minutes or the whole time.


FULL RANKED LIST


1. [COUNTY ZOO AT COUNTY PARK 1] - FREE-ISH / UNUSUALLY GOOD
- Town: [TOWN A], [HOME COUNTY]
- Dates and times: Open daily 10:00am-4:30pm, last entry 4:00pm. Wednesday Aug 26 only, open 10:00am-8:00pm, last entry 7:30pm. This is the season's final "Late Night."
- Cost, [HOME COUNTY] resident rates: adult $6.00, child 3-14 $4.00, under 3 free. Total $10.00. Bring proof of residency. Non-resident would be $17.00.
- What it is: Walk-through zoo of North and South American animals. Miniature train and the antique carousel are $3.00 per person per ride, tickets sold only at the zoo ticket booth, running to 8:00pm on Aug 26.
- Why it fits: Short loops, lots to look at from a stroller, and the train and carousel are the main event at these ages.
- Registration: None. Tickets at the gate.
- Cautions: No petting and no feeding the animals at this zoo. Stroller rental is $5 refundable deposit. Note the the zoo's independent support organization website lists the resident and non-resident columns reversed; the county page above is the correct one.
[LINK REDACTED: official page of the venue named above]

Pair it with [COUNTY PARK 1]'s pony rides, a separate concession in the park: $5 per ride, child must be over 2, under 50 inches and under 90 lbs, so [which child qualifies: REDACTED]. Wednesday hours are 11am-1pm and 2-6pm; Saturday and Sunday 10am-2pm; Monday 10am-12pm and 1-4pm; closed Tue, Thu, Fri.
[LINK REDACTED: official page of the venue named above]


2. [STORY TIME PROGRAM], [WILDLIFE CENTER, STATE PARK IN NEIGHBORING STATE] - FREE / LAST SESSION OF THE SEASON
- Town: [STATE PARK], [COUNTY IN NEIGHBORING STATE]
- Date and time: Tuesday Aug 25, 2026, 10:30-11:30am at the zoo amphitheater. The state parks listing reads "Weekly on Tuesday 10:30 - 11:30 am until Tuesday, August 25, 2026," so this is the last one.
- Cost: Program free. Zoo is a $1.00 per person suggested donation. Park entry is $10 per vehicle. Roughly $13 all in.
- What it is: Educators lead wildlife storytelling; some sessions add puppets, a craft, and a chance to touch real animal fur. Afterward the wildlife center itself is open 10:00am-4:30pm, a small collection of rescued native animals - bear, coyote, eagle, reptiles.
- Why it fits: "Recommended for children 5 and under" is a rare exact match, and drop-in format means no penalty if the youngest lasts ten minutes.
- Registration: "NO Registration Required!"
- Cautions: This is the longest drive on the list, roughly 50-55 minutes up the the highway north, plus about a 20-minute walk from the lot to the wildlife center. Paths are paved and stroller-suitable but hilly. The state park carousel's hours and price are not published anywhere official - do not promise it.
[LINK REDACTED: official page of the venue named above]


3. [FARM A] PETTING ZOO - UNUSUALLY GOOD VALUE
- Town: [TOWN B], [HOME COUNTY]
- Dates and times: Monday-Saturday 9:00am-5:30pm, last tickets sold 4:30pm. CLOSED SUNDAYS, so Sunday Aug 30 is out. Open rain or shine.
- Cost: $5.50 per person online for ages 2 and up, children under 24 months free. Total $11.00 online. At the door it is $7.00 each, cash only, so $14.00.
- What it is: Goats, sheep, alpacas, donkeys, ponies, bunnies and pigs, plus the goat bridge, a duck pond, and a playhouse village. Animal feed sold at the feed booth.
- Why it fits: Hands-on animals at toddler height, walkable in 45 minutes, and one child is still free [detail REDACTED].
- Registration: Timed ticketed entry - buy online for your slot. Tickets are non-refundable.
- Cautions: The rabbit area is closed for renovation as of 8/7/26. Feed booth and door tickets are cash only. Saturday pony rides are $6 per child, ages 2 and up, 80 lb max - the older child only [age detail REDACTED].
[LINK REDACTED: official page of the venue named above]


4. SENSORY FREE PLAY + [MUSIC CLASS], [LIBRARY A] - FREE
- Town: [TOWN C], [HOME COUNTY]
- Date and times: Monday Aug 31. Sensory Free Play 10:00-11:00am in the Children's Room, then [MUSIC CLASS] 11:00-11:45am on the front lawn, moving to the auditorium in bad weather.
- Cost: Free, both.
- What it is: Sensory play is bins and tubs of materials, unstructured. The music class is run with a local music school - a "musical petting zoo" parent-child session with sing-along, story time and instrument exploration.
- Why it fits: Sensory Free Play is stated as "ages 5 and below," which covers both children in the same room. The music class is stated as ages 1-3, so it fits our children [details REDACTED].
- Registration: "Registration is not required" for both. No residency or library-card restriction is stated on either page.
- Cautions: Back-to-back timing is the appeal here, but it is two hours of program for a very young child - plan to bail early if needed. Contact [STAFF NAME REDACTED], [PHONE REDACTED].
[LINK REDACTED: official page of the venue named above]
[LINK REDACTED: official page of the venue named above]


5. MONARCH WEDNESDAYS, [NATURE SANCTUARY A] - FREE
- Town: [TOWN E], [HOME COUNTY]
- Date and time: Wednesday Sep 2, 2026, 4:00-4:45pm
- Cost: FREE
- What it is: A 45-minute pop-in program - see the resident monarchs and caterpillars, walk the monarch life cycle, harvest milkweed to feed the caterpillars, explore the pollinator garden.
- Why it fits: Listed as a family program with no age minimum, it is short, outdoors, and the drop-in structure is forgiving.
- Registration: The description says "no registration required, come for five minutes or the whole time," but the page also carries a REGISTER HERE link and a site-wide banner saying pre-registration may be required for certain programs. Call [PHONE REDACTED] first if you want certainty.
- Cautions: Late-afternoon outdoor garden - sun, ticks, bug spray.
[LINK REDACTED: official page of the venue named above]


6. [FOOD FESTIVAL] AT [MEGA MALL] - KIDS FREE
- Town: [TOWN F], [HOME COUNTY]
- Dates: Friday Aug 28 through Sunday Aug 30, 2026. DAILY HOURS ARE UNVERIFIED - the official site publishes dates but no times. Call or check before driving.
- Cost: General admission $10.00 per person, VIP $20.00, and "Kids under 10 years old are FREE." So $10.00 total for your group.
- What it is: 100+ Latin food vendors, live music and DJs, and specifically bounce houses, train rides, pony rides and face painting, plus free salsa lessons.
- Why it fits: For $10 the children get the train, bounce house and stroller-friendly wandering [age fit details REDACTED].
- Registration: Tickets sold online in advance.
- Cautions: Two things are UNVERIFIED and both matter - the daily start and end times, and whether this is indoors or in the parking lot. The site mentions a "cooled VIP tent" and third-party ticketing refers to a mall parking lot, which suggests outdoors. Also unverified whether the pony rides carry the usual age-2 minimum.
[LINK REDACTED: official page of the venue named above]


7. [FARM B] - APPLE PICKING SEASON OPENS SATURDAY AUG 29
- Town: [TOWN G], [HOME COUNTY]
- Dates and times: Farm open daily 8:00am-5:00pm. Petting zoo Monday-Sunday 9:00am-5:00pm, last tickets 4:30pm - open every day of the window, Sundays included. Pick-your-own apples: "We look forward to welcoming you back for PYO apples beginning on August 29th, 2026."
- Cost: UNVERIFIED. Neither the petting zoo nor the apple-picking page publishes any price, and their booking system could not be read. Call [PHONE REDACTED] before you go.
- What it is: Goats, sheep, alpacas, donkeys, ponies, pigs and cows, a goat swing, playhouse, duck pond and waterwheel, plus the orchard.
- Why it fits: The only farm here open on Sunday, and picking one apple off a low branch is about the right amount of work for a young child.
- Registration: Advance tickets recommended, "ideally a day or two ahead."
- Cautions: The Farmyard is cash only, ATM on site. PYO peaches are already closed for the season. Aug 29-Sep 3 is the very front edge of apple season, so expect early varieties only and limited selection.
[LINK REDACTED: official page of the venue named above]
[LINK REDACTED: official page of the venue named above]


8. [COMMUTER FERRY], [FERRY TERMINAL] TO THE CITY - BOTH KIDS FREE
- Town: [TOWN H], [COUNTY 4]
- Dates and times: Every day of the window. Weekdays from 6:00am, every 20 minutes at peak and midday. Weekends 8:00am-9:40pm every 20 minutes.
- Cost: Adult one-way $10.25. "Child (5 & Under) FREE - Two children for every paying adult." Round trip for all three of us is $20.50. Parking at [FERRY TERMINAL] is extra and the rate is UNVERIFIED.
- What it is: A 10-minute river crossing with big open-water skyline views, indoor seating plus an outdoor deck. Ride over, ride back.
- Why it fits: A real boat, short enough for a very young child, no reservation, no security line, stroller rolls on.
- Registration: None.
- Cautions: The whole cost is the parking. Nothing to do at the far end unless you want it.
[LINK REDACTED: official page of the venue named above]


9. [SPLASH PAD PARK] - FREE (LIKELY) / OPEN EVERY DAY OF THE WINDOW
- Town: [TOWN D], [HOME COUNTY]
- Dates and times: Open 7 days a week, 10:00am-8:00pm, through September 7. the town's official notice of Aug 19, 2026 states: "[the] Splashpad will remain open (7 days/week) through September 7 (10a-8p as originally scheduled)." [venue name redacted from quote]
- Cost: No fee is published for the splash pad, and no residency rule is stated - both UNVERIFIED. The town's membership pool nearby is a separate paid membership facility.
- What it is: Ground-level spray pad at the north end of the park, next to a large all-inclusive playground.
- Why it fits: Ground-level spray is the one water format that works for both of our children at the same time, and the weekend forecast is made for it.
- Registration: None.
- Cautions: Unguarded. the town's two other splash pads close after Aug 30 and the town sprinklers close Aug 28 - this is the one that stays open; a fourth site is closed.
[LINK REDACTED: official page of the venue named above]


10. [ZOO B] FAMILY FUN NIGHT - Wednesday Aug 26 and Thursday Aug 27
- Town: [TOWN I], [COUNTY 3], about 35-40 minutes
- Dates and times: Zoo open daily 10:00am-4:00pm. Family Fun Nights Wed Aug 26 and Thu Aug 27, 4:00-7:00pm, off grounds by 8:00pm.
- Cost, daytime: adult 13+ $22.00, child 2-12 $19.00, under 2 free - $41.00 total. No county-resident discount, everyone pays the same. Family Fun Night is "$17 per person between 4pm - 7pm"; whether the under-2 exemption applies to that flat rate is UNVERIFIED, so budget $34-$51.
- What it is: Full zoo with giraffe hand-feeding, penguins, a stingray touch tank, aviary, and a dinosaur playground. The zoo train is FREE with admission, an 8-10 minute run every 15 minutes. Carousel is $3.00 per person.
- Why it fits: A free train ride included in admission is a real advantage at these ages.
- Registration: Not required.
- Cautions: This is the most expensive item here by a distance, and [COUNTY ZOO] on Aug 26 does the same job for $10. All ticket sales are final, no refunds or rescheduling. Strollers are banned in the aviary and in the giraffe feeding area, and not permitted on the train. The train has a loud whistle. Parking is free across the recreation complex - avoid the commuter garage nearest the main road, which charges on weekdays.
[LINK REDACTED: official page of the venue named above]
[LINK REDACTED: official page of the venue named above]


11. [SWIM LAKE, COUNTY PARK 2] - RESIDENT PRICING
- Town: [TOWN J], [HOME COUNTY]
- Dates and times: Open daily through Labor Day, September 7. Swim hours 10:00am-6:00pm daily. Park hours 9:00am-7:30pm weekdays, 8:00am-7:30pm weekends.
- Cost, [HOME COUNTY] resident: weekday adult 18-61 $12.00, child 5-17 $8.00; weekend adult $13.00, child 5-17 $9.00. The fee table has NO LINE for children under 5, so whether your two are free is UNVERIFIED - call [PHONE REDACTED]. Out-of-county adults pay $20-$25.
- What it is: A spring-fed swimming lake with a sand beach - the best true toddler wading water in [HOME COUNTY].
- Why it fits: Gradual sand entry beats any pool for a very young child.
- Registration: None.
- Cautions: The ticketed inflatable water course inflatable is off-limits to both children - it requires 48 inches, and even the kiddie version requires 39 inches. Natural lake bottom. No swim-diaper policy is published; bring them anyway. No minors 17 and under admitted without an adult.
[LINK REDACTED: official page of the venue named above]


12. [COMMUNITY CARNIVAL, 100TH ANNIVERSARY] - Saturday and Sunday open at 3pm
- Town: [TOWN K], [COUNTY 3], about 25-35 minutes
- Dates and times: Wed Aug 26 through Sun Aug 30. Wed/Thu/Fri 5:00-10:00pm; SATURDAY AND SUNDAY 3:00-10:00pm.
- Cost: Gate admission is not published - UNVERIFIED, though such carnivals are normally free entry. Rides run on credits: 12 credits $17, 25 credits $27, 55 credits $57, plus a $3 convenience fee per purchase. Figure $17-$27 buys a young child three to six kiddie rides.
- What it is: Amusement rides, carnival games, food stands, and live entertainment.
- Why it fits: The 3:00pm weekend opening is the single reason this is on the list - it is the only carnival in range confirmed to open in daylight, before dinner and bedtime.
- Registration: None.
- Cautions: Most kiddie rides require a 36-INCH MINIMUM, which may exclude one or both children [details REDACTED]. Several rides cap at 48 inches. All guests under 18 must be with an adult 21+, four minors maximum per adult. Expect heavy crowds at a centennial - go at the 3pm open, not at night. No rain policy published.
[LINK REDACTED: official page of the venue named above]


13. FAMILY STORYTIME, [LIBRARY B] - Saturday option
- Town: [TOWN D], [HOME COUNTY], community room
- Date and time: Saturday Aug 29, 2026, 11:00-11:30am
- Cost: Not stated on the event page - UNVERIFIED, though library storytimes in the county consortium are ordinarily free.
- What it is: "Stories, songs and smiles."
- Why it fits: The listed audience runs from 6 months through 9 years - the only program found that formally covers both children in the same room on a weekend.
- Registration: Not mentioned on the page. No residency restriction stated. Verify at [PHONE REDACTED] or [EMAIL REDACTED].
- Cautions: 30 minutes, and it is the same morning as good splash-pad weather - stack them.
[LINK REDACTED: official page of the venue named above]


14. [SCIENCE CENTER] - rain-day backup
- Town: [TOWN L], [COUNTY 4]
- Dates and times: Open 10:00am-5:00pm daily through September 7, so every day of the window.
- Cost: ADMISSION PRICE UNVERIFIED. The live pricing is published only as an image the site would not render, so no dollar figure is confirmed here - call [PHONE REDACTED]. Parking is confirmed at $7.00 per car. Advance purchase is confirmed cheaper: "Get the best price on admission when you buy at least one day in advance!"
- What it is: Its toddler zone is purpose-built for toddlers and preschoolers, plus a licensed-character train play space, a dinosaur dig, and a river touch tank. Daily programming includes the toddler planetarium show at 10:50am and 1:35pm, animal shows, and a dance party at noon and 3pm.
- Why it fits: The strongest indoor option in range for both ages, and the obvious move if Thursday's storms land.
- Registration: Timed arrival slot chosen at checkout.
- Cautions: Price unconfirmed. Two big-kid climbing attractions are not for children this young. Last week of summer means peak crowds. Its under-6 program is currently on hold.
[LINK REDACTED: official page of the venue named above]


ALSO CHECKED, AND WHY THEY ARE NOT ON THE LIST

- A toddler nature series at [NATURE SANCTUARY A], Thu Aug 27 - the best age match found anywhere, 18 months to 3.5 years, and it is SOLD OUT on the ticketing site. [Age-fit detail REDACTED.]
- [NATURE CENTER B] butterfly house feeding, Sat Aug 29 - last public day of the season, $8 non-member, under 2 free, but the stated rule is "Program suitable for children ages 4+," which our older child does not yet meet [detail REDACTED]. Worth one phone call to [PHONE REDACTED] if you want to try. Aug 30 is members-only.
- A live-animal show at [LIBRARY C], Wed Aug 26 - would have been a top pick, but it is limited to that town's card-holders and is waiting-list only.
- Toddler programs at four other town libraries - all either residents-only or already fully booked.
- Every county and town free concert and movie series in the three nearest counties ended on or before Aug 22. [HOME COUNTY]'s last 2026 movie was Aug 21 at [COUNTY PARK 1].
- The region's big end-of-summer festivals all fall just after this window, because Labor Day 2026 is Monday Sept 7: two large town festivals Sept 4-7, an aviation museum open-cockpit day Sept 5, [ZOO B]'s two toddler programs Sept 6 and 9, [HOME COUNTY] 4-H Fair Sept 18-20.
- Worth booking NOW for after the window: A zoo toddler class at [COUNTY ZOO], ages 18 months to 3 years, Tuesdays or Thursdays 10:30am starting Sept 8, tickets went on sale Aug 18. And a toddler nature class at [NATURE CENTER C], ages 18 months to 4 years, six Thursdays from Sept 10, $60 per child.
- A second community carnival in a nearby town, Aug 26-30 - real, but its organizers publish no hours or prices and such carnivals typically open around 6pm, which is too late here. Call [PHONE REDACTED] if you want it.
- A town food-truck festival and car show, Sun Aug 30 - appears on the township calendar but the site blocks retrieval, and two sources give conflicting times. Not confirmed, so not recommended. Call the town's recreation department.
- A railway museum excursion train - the only 2026 date posted is Sept 13, after this window.
- Two neighboring-state park pools and beaches are weekends-only now, so only Aug 29-30 would work; the county swim lake is closer and open daily.

CAVEATS ON THIS EDITION
Everything above was checked against the official page linked beneath it during this run. Where a venue does not publish a fact, it is marked UNVERIFIED rather than guessed - the unverified items this week are [FARM B] pricing, [SCIENCE CENTER] admission pricing, the festival's daily hours and indoor/outdoor setup, the swim lake's under-5 fee, the splash pad's fee and residency, [ZOO B]'s Family Fun Night under-2 rule, and [LIBRARY B] storytime cost and registration. Call before driving on any of those. Prices, hours and event status can change after this was written.

The project's record files are held privately and shown on request.

Held privately and shown on request: the standing instruction document (version 1.2), the August 24 amendment message, the server read-backs, the delivery records and the settings screenshots.

Status: Paused. The task's server record, read September 23, 2026, shows it switched off by an update dated September 14, 2026; the reason for the switch-off is not in the records this chat can reach. Last delivered report: September 10, 2026 (Gmail sent-mail record, read today).

Dates: First record August 23, 2026, about 10:02 PM Eastern (my instruction document pasted into the project chat, with my dated covering note). Latest record September 23, 2026 (today's server and Gmail reads).

Tools and models, as the records name them: a Claude Cowork scheduled task; the model my own chat note names as Opus 5 for the first, manually run report of August 23; the model the server record names as claude-fable-5 on the task since I changed it on August 24 (my settings screenshots show the change from Opus 5 to Fable 5); the Gmail connector for delivery; web search and official-page retrieval tools; eight parallel research assistants launched in the first run (chat record); and a government weather forecast page.

Who did what. I wrote the entire operating specification before any output existed: scope, ranking rules, the verification standard, the email format, and the failure plan. I ordered every change to it by written, versioned amendment; I set the permission mode, the model, and the run time myself in the app (screenshots, August 24); I ruled on the run time; and I ordered a re-check of finished work. The AI drafted and ran the research, fetched and compared official pages, wrote and sent the reports, created and edited the stored task on my instruction, and read the server back to me as evidence.

What this shows I can do, each tied to a record:

1. Writing a specification an autonomous system can execute with no one present. Record: the instruction document of August 23; the August 27 report arrived at 7:21 AM Eastern by automated run (Gmail record), and its own text opens by calling itself an automated weekly run.
2. Setting evidence standards for AI output instead of accepting fluent answers. Record: my installed rule "State nothing as fact you have not confirmed this run," and the August 23 email, which flags 7 facts as UNVERIFIED rather than guessing.
3. Change control: versioned documents with supersession lines and required proof of change. Record: my August 24 amendment message demanding a complete echo of the stored text after editing, and the version 1.0 to 1.1 to 1.2 chain on the server.
4. Failure-mode planning before launch. Record: my installed fallback rules (save a draft if sending stalls; send a clearly labeled PARTIAL report rather than nothing), written before the first automated run.
5. Judging operational detail: schedule time handling, the scheduler's habit of firing minutes past the hour, permission modes. Records: the August 24 chat exchanges and my ruling quoted in Part B.

1. The problem. Finding, verifying, pricing and ranking suitable weekly outings for young children is a recurring research chore where listings are often stale and prices depend on exact ages. The records hold no sourced comparison with off-the-shelf tools, so none is claimed.

2. What I decided first. The whole operating specification preceded any output: coverage window, search area, priorities, a no-padding rule, the verification standard, the email format, delivery, and failure handling. I refused to leave verification judgment to the AI: "State nothing as fact you have not confirmed this run" (text I installed, August 23, 2026).

3. How requirements reached the AI. One versioned instruction document, pasted August 23, 2026. Later changes only by written amendment: my August 24 amendment ordered a two-bullet replacement and said "Change nothing else anywhere in the prompt" (text I installed, August 24, 2026).

4. What the AI produced. The same night: a report with 5 best bets and 14 ranked options, each with dates, prices, age fit, registration needs, cautions and an official link, sent by email; then the stored scheduled task; then automated weekly reports.

5. Design choices that were mine. The quality bar: "Fewer strong options beats a padded list" (installed text, August 23). Plain-text formatting because markdown renders as literal characters in email. UNVERIFIED flagging as a named output category. The run time: "I set it to seven AM on purpose." (my own message, August 24, 2026).

6. What went wrong, and what caught it. Two official pages gave conflicting price tables; the AI's cross-check under my sourcing rule caught it, ruled for the county page, and disclosed the conflict in the email. A perfectly age-matched program was dropped after its ticketing page showed sold out. The AI, reading the server, caught a mismatch between the stored text and the actual run-time setting. My order "Please check your work and revise the code block if it needs fixing. don't make changes just for the sake of making changes." (my message, August 24) produced one correction: a stated run time was more precise on the server than in the draft.

7. Where it stands now. Three automated reports were delivered (August 27, September 3, September 10; Gmail records; the September 3 body itself was not re-read today, but its dated reply thread proves delivery). The server marks the September 10 run failed, with unspecified reason, one second after that email's send time; the conflict is unresolved. The task was switched off September 14 for reasons outside reachable records; no September 17 run; no fallback drafts exist. Unfinished: the clock-change adjustment due the week of November 1, 2026, and the decision whether to resume.

Items are numbered as in the project's own record. Only the items the record marks public appear here; the rest are held privately and can be shown on request.

3. My rulings and quality-control orders in chat, August 24 to 25, 2026: "I set it to seven AM on purpose." and "Please check your work and revise the code block if it needs fixing. don't make changes just for the sake of making changes." What they show: the decisions and the re-checking were mine. Written by: me. Check: view the dated chat messages. Where: public as short quotes; the surrounding chat private, shown on request.

6. The redacted sample report (package file 1). What it shows: output quality and honesty markers; the text uses the word UNVERIFIED 11 times, and its closing section names every unverified item. Before redaction the reconstruction matched the original's recorded size, 20628 bytes, exactly. Written by: the AI under my rules; redaction by the AI today at my direction. Check: recompute the file's SHA-256 fingerprint (a standard has-this-file-changed check) against the README; compare with the sent original on a call. Where: public, redacted.

What is not claimed

no comparison with off-the-shelf tools was measured. No test-suite counts exist for this project. The September 3 email body was not re-read today. The server's failed mark on the September 10 run conflicts with that email's delivery record and is unresolved. The reason for the September 14 switch-off is unknown to this chat. The reports' facts rest on the official pages cited at each run date and can have changed since. Whether the first automated run reflected the amended instructions in every detail rests on the AI's own run, not on a record I have re-read.

Case 1: Amendment under audit. The stored context bullet held estimated details that I knew to be wrong. On August 24 I issued a formal amendment replacing two bullets and the version line, ordered that nothing else change, and required the complete stored text echoed back for comparison. The echo came back from the server record, and the comparison statement disclosed exactly one difference outside the ordered changes: a cosmetic line-wrap shift caused by the longer version line. What changed: stored version 1.0 became 1.1. Records: my dated amendment message and the server read-back, both August 24. Files: package files 2 and 3, both private, shown on request

Case 2: The schedule audited in both directions. The AI, reading the server, flagged that the stored text said one run time while the saved setting was an hour earlier, and separately that the scheduler anchors runs several minutes past the hour rather than on it. I ruled: "I set it to seven AM on purpose." The AI then updated the stored text to version 1.2 so document and setting agreed. When I ordered a re-check of finished work, the AI corrected its own draft, which had stated the round hour where the server showed minutes past. The August 27 delivery at 7:21 AM Eastern is consistent with the anchoring behavior, not the rounded time shown in the app. Records: chat of August 24 to 25; the server record; the August 27 delivery record. Shows judgment on both sides and my control of the loop

Case 3: The verification standard producing honest output. In the first run, two official pages published contradictory resident price tables; under my sourcing rule the county's own page was ruled correct and the conflict was disclosed to readers inside the email. A program that fit the children best of anything found was excluded because its ticketing page showed sold out. Seven facts that could not be confirmed from any official page were labeled UNVERIFIED in the email rather than stated. Records: the August 23 sent email (redacted copy supplied) and the chat of that evening. File: package file 1, public, redacted

The project's record files are held privately and shown on request.

Status: Built. Last true September 4, 2026. Record: the final game file dated September 4, 2026 (8:34 PM) and the eighth and last publish confirmation in the chat on September 3, 2026, followed on September 4 by the export the author requested. No later work is recorded.

Dates: First record September 3, 2026 (the author's opening message in the chat). Latest record September 4, 2026 (the request for the handoff document, and that file's timestamp, 8:37 PM).

Tools and models, as the records name them: Claude Fable 5.1 in Claude Cowork (author's own message, September 3, 2026: "this is a text conversation in cowork on Fable 5.1 with auto set as the default"). Gemini, used by the author in parallel through a Google AI Pro subscription, including Gemini Omni for a test video and Gemini Canvas for a code test (messages of September 3, 2026); the Gemini model version is not in the records, UNVERIFIED. Playwright with Chromium for browser tests, ffmpeg for frame extraction, Python for the print files, all shown in tool output of September 3 and 4, 2026.

Who did what. The author specified the audience, the humor style, the interest themes, the four-sense rule, the two age tiers, the together-play principle, the no-payment rule, the no-voice rule, and the content boundary (potty humor allowed, nothing else relaxed). The author rejected device text-to-speech, rejected voice cloning, rejected paying for a video API, and corrected an over-removal of words. The author ran the cross-check against Gemini personally. The AI drafted all code, stories, word banks, printables and memos, ran the browser tests, and verified the video-capability claims against vendor documentation.

What this shows I can do (each tied to a record):

1. Writing a product specification for a non-reading user from observed behavior. Record: message of September 3, 2026, "she appreciates ironic twists and things being out of place misplaced or surprising expectations."
2. Setting a content policy and holding it, with one deliberate exception. Record: September 3, 2026, "make sure everything is age appropriate no exceptions," and later the same day, "toilet stuff is fine for her."
3. Challenging an AI capability claim and requiring documented verification rather than accepting an answer. Record: September 3, 2026, the demand to explain how single-prompt games are possible if video is not; the AI's reply cites Anthropic's own help article.
4. Cost and scope discipline. Record: September 3, 2026, "I'm not paying for this. I'm trying to see what kind of games I can create for my kids."
5. Privacy and trust judgment about AI voice. Record: September 4, 2026, refusal to have the author's voice recorded or cloned (quoted in Part D).
6. Cross-model evaluation: running the same question through a second AI and reading its placeholder as a limit, not a feature. Record: September 3, 2026, "gemini can't do it."

1. The problem. The author wanted a game for a bright almost-preschooler who loves misplaced words, playable with a parent, not a device handed to a child. No off-the-shelf comparison is in the records; none is claimed.

2. Decided before building. From the author's own message of September 3, 2026: "make it a fun visual interactive if possible if this is the wrong space for it in cowork tell me I'll do it somewhere else." The form (a mad lib), the humor, and the platform question were set up front. Age appropriateness was non-negotiable from the first revision request.

3. How requirements reached the AI. Dictated chat messages, each a change request. September 3, 2026: "five choices is correct"; "each individual story should work in one smell... at least one smell element and at least one taste element and at least one visual element and at least one auditory element"; "Yes. Rebuild, but with two age versions." All are the author's own messages.

4. What the AI produced. Eight publishes on September 3, 2026, moving from a single mad lib to a hub of eight games, an SVG cartoon movie (SVG is a drawing format a browser can animate), a per-child storybook, printables and bedtime scripts. Counted from the final file: 13 big-kid stories, 10 little-kid stories, 12 word categories with 186 picture words, 20 Fix-the-Story items, 28 Who-Said-It lines, 5 hunt packs with 54 clues, a 20-animal guesser.

5. Design choices that were the author's. Together-play as the governing rule (September 3, 2026: "what I want is something we can play together or as a group, not something to hand her alone to take wherever she wants. though, playing alone but in front of me is fine"). The reasoning, same message: "I'm acutely aware of the problems with video games and addicting screen time etc. you can see this in that my first game was a mad lib." The quality bar (September 3, 2026: "I want all of that, I want it done excellently, at the highest level possible"). The no-voice rule (September 4, 2026, quoted in Part D).

6. What went wrong and what caught it. (a) The AI hid the voice picker when a browser reported no voices; the author caught it: "I also haven't seen a button for the reading voice on the site" (September 3, 2026). (b) A sentence splitter bug produced 181 scenes; a screenshot in the AI's own test caught it (tool output, September 3, 2026). (c) The AI removed potty words under the "no exceptions" rule; the author corrected the over-reach. (d) Gemini gave wrong answers about its own abilities; the author caught that by testing it directly.

7. Where it stands. Built and published privately; not tested on the author's own devices (stated in the chat); never used with the children in any record. The in-page movie is hand-drawn cartoon, not generated video. That limit is documented, not hidden.

Items are numbered as in the project's own record. Only the items the record marks public appear here; the unredacted game file, its version record and the second AI's reply are held privately and can be shown on request.

1. The author's dated chat messages, September 3 and 4, 2026. Show the specification, the corrections, the refusals and the reasoning, in the author's words. Written by: me. Check: view the dated chat. Where: public, redacted (screenshots must crop the chat title, sidebar, and any message naming a child; the quotes in Parts A, B and D are already clean).

3. The redacted sample in the zip. SHA-256 7bb72ea73b4acc7040150687a8ed1cb0a8ad205c1dd8fc30f3a67b2348a408aa. Written by: the AI. Check: open it, recompute the fingerprint. Where: see Part F.

4. Test results as tool output. September 3, 2026: "all 13 stories carry smell, taste, sound, visual"; "all scenes directed without error"; "after reload mode: little screen: s-book shelf: 11". September 23, 2026, on the redacted sample: 13 of 13 big-kid stories and 10 of 10 little-kid stories filled, 0 page errors. Written by: the AI (its test script). Check: run the sample; or watch the test run live. Where: public for the counts.

What is not claimed

any test on the author's Samsung devices; that the "Send to Gemini" share-sheet works (UNVERIFIED); that the movie prompt produces a good Gemini video (never run); any use with or reaction from the children; that the printables were printed; the Gemini model version; whether the Gemini-built copy works.

Case 1, the video feature

The author asked for clips of the quality already made in Gemini Omni. The AI said a web page cannot render that and, when pressed, verified it against Anthropic's help article ("Claude doesn't generate photos or illustrations the way image-generation tools do") and Google's Gemini API pages. The author then tested Gemini directly; Canvas produced a game with a placeholder for a paid video service, and the author concluded on September 3, 2026: "gemini can't do it." Decision: no paid API. What changed: the video button became a prompt hand-off, and the in-page cartoon was raised instead. Proof: messages of September 3, 2026 (public, redacted); the strategy memo (not shown as it stands); the final file (private, shown on request); the redacted sample (see Part F).

Case 2, voice

The AI added device text-to-speech. The author rejected it on September 3, 2026: "the available reading voices are robotic sounding and scary for children... and I can do the narrating myself." The AI removed all voice code. It later suggested recording the author's voice; the author refused on September 4, 2026: "no I don't want to use my own voice I don't want to freak my kids out and make them wonder about whether anything that's something that sounded like Daddy is something that I really said." Proof: those two messages (public, redacted); the file versions (private, shown on request).

Case 3, the over-removal

Acting on "age appropriate no exceptions," the AI removed toilet words from the word bank. The author corrected it on September 3, 2026: "toilet stuff is fine for her." The words were restored, recorded in tool output the same day as "restore potty words (approved)." Shows the rule was the author's to draw, and it was drawn precisely. Proof: the message (public, redacted); the final file (private, shown on request); the sample (Part F).