Aaron Zeanah · Product Owner & AI Workflow Architect · Atlanta
Red Eagle: a governed AI product system, documented end to end.
In 2012 I published a peer-reviewed study of William Weatherford and the end of the Creek frontier. In 2026 I rebuilt it as a product: an AI production pipeline where no generated line, image, or scene exists without a traceable path back to a primary source. This portfolio documents the system — its architecture, its evidence chains, its failures, and the controls those failures produced.
✓ Every figure on this page was re-verified against the project repository — file counts, register ids, and decision numbers — on 26 Aug 2026, before publication.
In brief · 60 seconds
A historian's method, run as a delivery pipeline.
Stories are how an idea becomes convincing. The hard part was never the telling — it is finding and defining the data points a story can honestly rest on. That is a historian's problem before it is an AI problem, and it is the same problem I spent ten years solving in product organizations: turning scattered, unreliable evidence into something a leader can act on.
The problem
A story is the most effective way to demonstrate an idea — and the easiest to get wrong, because the critical data points underneath it are hard to locate, harder to define, and trivial for a generative model to invent. Fluency is not evidence.
What I built
A historiographically-trained, SDLC-backed storytelling engine with programmatic fact-checking wrapped around the generative layer — eight agent roles, evidence tags, a citation register, and two human gates — built to breathe life into old evidence without letting it drift from the record.
What it produced
A structured, feature-length screenplay — 24 story units and a seven-beat epilogue — plus an engineered prompt library targeting animation APIs. Every line traces to a source or carries an explicit invention tag. Zero fabricated claims in audited output; 66/75 on an independent rubric.
Why it transfers
Progressing from coordinator to strategist taught me that digital leaders need confidence in their decisions, and the fastest route to it is a user story told quickly with clear data behind it. Same instruments I ran at FanDuel and NBA Digital — pointed at a generative system.
01 — The Strategist
Five roles, one method
The trajectory runs Project Coordinator → Business Analyst → Product Operations Manager → Voice of the Customer Product Partner → Lead CX Strategist. Read as a job history it is a climb through product operations. Read as a method, every step is the discipline a historian brings to an archive: find the records, source them, weigh them, and claim only what they support.
Coordinated technology delivery across global reservation & revenue systems with 20+ development resources — and learned that shipped software is an operational event, not an endpoint. Usage and SLA records were the first archive: the truth about a rollout lived in what hotels actually did with it, not in the launch plan.
Built a Salesforce case-management system for the ethics & compliance team — requirements, test scripts, and user stories in an environment where every claim had to survive an audit. The user story became the unit of record: actor, motive, consequence, written so a stranger could verify it.
Owned live services & support across 16 apps and 370k+ subscribers, and defined the contact-reason taxonomy governing how every fan interaction was classified, routed, and reported. This is source criticism at scale: a support contact is testimony, and the taxonomy decides what the record can later prove. Elevated the BBB rating from “D” to “A” on the strength of that record.
Mapped end-to-end fan support journeys across automation, live, self-service, and product touchpoints, and owned the customer satisfaction roadmap for FanDuel Racing & TVG. UX research done as archival work: surveys, app reviews, and session records read together as corroborating — or contradicting — witnesses, reducing reliance on support 3–5% annually.
Led enterprise Voice of Customer strategy in a regulated, compliance-heavy environment — a unified insight framework across live support, surveys, reviews, and product analytics, quantifying up to $42M in annualized revenue risk and turning it into prioritized roadmap. The historian's endpoint: an argument executives can act on because every claim in it traces to a record.
User feedback is a primary source — and the structure of the story is the user story.
Feedback arrives the way primary sources do: fragmentary, biased, self-interested. It rewards the same criticism — who is speaking, what they want, what corroborates them. And a user story is an actor, a motive, a consequence. Ten years of product operations and a historian's training are the same craft on different archives.
AI raises the stakes on exactly these skills. Building in digital environments now depends on knowing how to research — interrogate sources, weigh testimony, keep provenance — and how to tell the story the evidence supports. Stitching together the historical record of users — what they did, said, and asked for over time — is what separates refined detail from casual prompting. Anyone can generate; the discipline is in what the generation is allowed to rest on. The system documented on this page is that discipline, built and run end to end.
02 — The Product Challenge
Generative speed is a solvent. It dissolves evidence, authorship, and ethics in that order.
A model will write you a Creek War screenplay in an afternoon. It will also invent the quotations, attribute them to real people, and cite page numbers that do not exist — fluently enough that nobody catches it until the work is finished and wrong.
That is not a hypothetical. An audit run ahead of drafting found that roughly two-thirds of the quotations the draft scripts presented as historical did not survive contact with the sources. Four of the errors reversed the meaning of the scene they sat in. Both of the drafts' page citations to my own published paper were fabricated — and one of them was serving as the project's thesis statement.
The problem is structural, and it has three faces:
Evidence
A fluent model produces confident detail about sources it cannot see. Plausibility is not provenance, and the failure is invisible in the output — it is only visible against the source.
Authorship
Adapting my own scholarship means the argument has to stay mine. A system that generates freely and asks forgiveness later ends up authoring the thesis by drift.
Ethics
Fort Mims, Holy Ground, and Horseshoe Bend are real mass killings of real people whose descendants are living communities. Speed is exactly the wrong instinct here.
The reframe
Treat invention as a first-class artifact rather than a defect. The film is allowed to invent freely — it simply has to know, in writing, every time it does.
03 — System Architecture
Seven stages, two human gates
Unstructured archival material enters on the left and leaves as a locked production asset on the right. Nothing crosses a stage boundary without a gate, and two of those gates are human. Material moves upward only — which is what makes provenance a property of the system rather than a habit of the operator.
Agent roster
Eight bounded roles. Each carries a purpose, an input contract, a required output shape, and a forbidden-behavior list — the last of which is the field that actually does the work. Select a role to open its dossier; the stages it operates in highlight in the diagram above.
04 — SDLC Mapping
The lifecycle, stage by stage
Every stage produces a durable artifact that the next stage is required to read. The artifacts are the interface between stages, and between me and the agents.
| Stage | What I did | Artifact produced |
|---|
05 — Research-to-Product Traceability
Most AI-generated historical content cannot answer a simple question: where did this come from? Not the topic — the specific line, the specific staging choice, the specific object in the frame. The output is fluent, internally consistent, and unfalsifiable, because there is no chain to walk back.
That absence is not a quality problem. It is a structural one. A model asked for a Creek War scene will produce quotations, attribute them to real people, and cite page numbers — and every part of that can be invented without a single signal in the text that anything is wrong. On this project's inherited drafts, roughly two-thirds of the quotations presented as historical did not survive contact with the sources, and both citations to my own published paper were fabricated. One of them was serving as the thesis statement.
So the system was built the other way around. Every element in the finished product carries a path back through four recorded steps: what the record actually says, what we decided it means, what we made from that, and what checked it. The chain is the product. The screenplay is just the part you can read.
| 00Product element | 01Historical basis | 02Interpretive decision | 03Creative output | 04Validation step |
|---|---|---|---|---|
| UNIT 17 The surrender at Fort Jackson April 1814 · Weatherford's central decision — riding alone into the enemy camp to ask for food |
Two independent accounts describe a plain, brief exchange: Woodward (1859) and William Gates, orderly sergeant, in an 1893 letter printed in the Transactions of the Alabama Historical Society (1898). Registered Q-11 VERIFIED. The famous oration — “I am in your power… I cannot animate the dead” — first prints in Reid & Eaton, 1817. Reid was Jackson's aide and was present. But Pickett called it “camp gossip,” Brewer “fictitious,” and Halbert & Ball reject it as lacking “sufficient evidence of genuineness.” Registered Q-12 COMPOSITE. Jackson's own hand, to David Holmes, 18 April 1814: the surrendered chiefs “have surrendered up Eighteen of the negroes taken at Mimms.” Q-51 VERIFIED. QUOTE_REGISTER.md · Q-11, Q-12, Q-51, Q-94 · Passes 003, 006, 020 |
D-15 — play the plain version. The oration exists in the film only as the version other people tell afterward. The most famous speech in the story enters through the rumor register, not through his mouth. D-83 — the split ledger. The surrender is forgivable for Weatherford; Jackson is forgivable only for a grief that is not in this scene. So in the one room these two men share, the defeated man may be forgiven and the victor may not. Do not resolve the asymmetry by staging — play the transaction flat and let both ledgers sit in the same shot. Rebuilt on Q-94. Woodward disavows the standard story in a parenthesis and gives a different mechanism: an intermediary sent up first, safety confirmed through his brother-in-law, and only then the ride in. The asking is the act, not the nerve. OPEN_DECISIONS.md · D-15, D-79, D-80, D-83 · SCENE_LEDGER unit 17 |
17-the-surrender.fountain — seven movements, plus a design document recording why each is staged as it is. The scene is built on a register entry nobody had ever staged. The food is not free: it is bought with eighteen people, in Jackson's own sentence, and nobody in the scene says so. There is a number in a ledger and the audience can read. Final image: two pages on one table — Jackson's ten true words, and Reid composing the 1817 oration three feet away while the plain exchange happens. History keeps the wrong one. 3-CREATIVE-WRITING/screenplay/17-the-surrender.fountain · _DESIGN.md |
C4 — transfer check. Forgiveness for the surrender is correct and is not a finding. The audit looks only for that forgiveness leaking onto what he does to Abram. C5 — the wall. One test: does a character who cannot see the system state it? Nobody may narrate the price. C6 — figure audit. Every number in the scene listed without judgment and checked on our side. The count of eighteen is the fifth instance of the counting motif and the one time the arithmetic is exactly right. Chain repaired: Pass 016 sampled Q-94 by search and recorded a caveat. Pass 020 read the transcription entire and returned a fuller, corrected wording. The earlier verdict stays visible with the correction attached. |
| UNITS 2 & 22 The Tensaw ground, twice c. 1799 and c. 1820 · location depiction — the same place under two legal orders |
Tensaw as a functioning bicultural settlement: who greets whom, who may speak in council, what a Wind Clan woman's word is worth, how a Scots trader and a Creek matriarch conduct a marriage as a treaty. Held in the tier-1 world bible. The apparatus that replaced it is equally documented. Jackson to Blount, 9 August 1814: “about twenty two millions of acres are contained in this cession.” Q-15 VERIFIED. Then section lines, the St. Stephens land office, and Monroe County incorporated 1820 — the survey grid arriving on ground that had been held by town and clan. 2-INTERPRETIVE/world-building/tensaw-creek-community.md · QUOTE_REGISTER Q-15, Q-51, Q-52, Q-63 |
The measured world — a structural principle for the whole picture, locked 2026-08-21. Build Scene 2 specifically enough to be remembered, then replay the identical categories on the same ground in Scene 22 under the new system. Authority descends through the maternal line → authority is a deed with a name on it. Land is where a town is → land is a plat with a number. A man's standing is his clan's → a man's standing is his colour. Discipline: none of this is stated. No character observes that anything has changed. The audience recognises it because they were shown the original — which is why Scene 2 is the film's control group. CANON.md § The measured world · § The wall between story and argument |
02-tensaw.fountain — the plantation set for a wedding. 22-twilight.fountain — the same ground, sorted into fixed and legible castes. Location plates are registered as separate reusable assets: Tensaw plantation, set for a wedding; Little River house and section line; the St. Stephens land office; a clerk's office, Monroe County. A scene shot at a location never re-invents that location. Each plate carries a registry row: id, model, mode, seed, prompt file, reference image, locked. Without the seed and the exact prompt, the plate is not reusable and the work is lost. screenplay/02-tensaw.fountain · 22-twilight.fountain · 4-PRODUCTION/locations/REGISTRY.md |
C5 — the wall. This is the case the rule exists to protect. The contrast carries the film's entire argument, and a single line of a character noticing it would collapse the device into commentary. Stage 3 production gate. No shot is generated without a numbered ledger entry, a written scene, locked character sheets for everyone in frame, and a location plate. The plate is a hard prerequisite, not an asset to improvise at generation time. Registry completeness check. A plate without a recorded seed and prompt fails the gate even if the image is good. Known debt: ~20 images already exist from an earlier pass with no prompts, models, or seeds stored. They are strong and stylistically coherent — and they are logged as casting references and mood, not locked plates, because they cannot be reproduced or extended. |
| UNIT 9 Fort Mims 30 August 1813 · mass-violence event — roughly 250 people killed |
Three documented points of view carry the scene. The enslaved man tied to the post with the lash raised as the attack begins DOC. The Creek woman ordered to go home DOC. Weatherford driving the warriors on, then imploring and being driven off with clubs raised over his head V-01, V-04. The gate line — “Dixon Bailey. To-day one or both of us must die.” — is V-17 TRADITIONAL: Charles Weatherford Jr.'s letter reporting an aunt who saw him, printed in Halbert & Ball (1895), ch. IX. Family testimony recorded eighty years later. QUOTE_REGISTER.md · V-01–V-04, V-07, V-17 |
D-99 — the thesis correction. The attack is the Muscogee Creek warriors' and would have happened whether or not Weatherford led. The scene now has three leaders in three places along the ravine — three middles and no top. The rising starts at Francis's end and travels; nobody looks to Weatherford, and he does not give the signal. The stated reason is a control, not a preference: a reader who thinks the assault is his doing is one step from thinking his restraint could have stopped it — which is the redemption reading the canon forbids. If Fort Mims happens without him, his presence is not the hinge and his mercy is not a counterfactual. Muscogee-language names for the two invented observers remain placeholders. Do not bake a name into a registry row. OPEN_DECISIONS.md · D-46, D-79, D-80, D-99 · CANON.md § Handling the hard material |
09-fort-mims.fountain — roughly sixteen pages, plus a design document. Three points of view, all three documented. The gate line reaches the page carrying its own provenance and its own guardrail, both inline: a TRAD annotation naming the 1895 source, and a direction that it must not read as a personal quarrel driving the massacre, because the killing that follows has nothing to do with either of them. 3-CREATIVE-WRITING/screenplay/09-fort-mims.fountain · _DESIGN.md |
C7 — atrocity check. Flag anything that makes violence or suffering beautiful, thrilling, or satisfying: lyrical language about the dead, violence choreographed for excitement, a death staged to be moving rather than to be seen, the camera lingering where it has no reason to. Plain, unsparing description is correct and is not a hit. Result: no FABRICATED claim anywhere in the drafted scenes, and all four anachronism hard bans clean. D-08 — the gate that has not opened. Depiction of Fort Mims, Holy Ground, and Horseshoe Bend is blocked until consultation with the Poarch Band of Creek Indians and the Muscogee (Creek) Nation. The scene is written, sourced, and staged. Nothing may be drawn. Defect this chain caught: an adversarial pass found that a character's age did not close across three scenes — and that a fix logged complete was complete in one place of three. Neither had been caught internally. |
| id | Quotation | Status | Note |
|---|---|---|---|
| Q-30 | “The Great Spirit gave this great island to his red children…” | FABRICATED | Modern pastiche. No verbatim record of the address exists. Sugden calls the standard southern-tour text “fraudulent,” “unhistorical,” “bogus.” |
| Q-31 | “…walk from the headwaters of the Alabama to its mouth on the bodies of her soldiers.” | FABRICATED | There is no documented Weatherford–Tecumseh exchange at all. |
| Q-32 | “The blood of our women and children, recently spilt by the savage hand…” | COMPOSITE | → Use the real letter, Q-33. |
| Q-33 | ⚠ Misquoted by us — corrected, re-registered as Q-42. | SUPERSEDED | The letter is 4 June 1812, not 5 June, and “wanton massacre of our women and children” is not in it (0 hits in Bassett). Real text: “my heart bleeds within me… But they must be punished… lay their Towns in ashes.” The real text is colder and better. |
| Q-34 | Burr–Jackson dialogue, all of it | INVENTED | Refuted 2026-08-25. The earlier finding of “no usable Burr–Jackson quotations” was a search failure, not an archival fact — the volume had been in the corpus the whole time and was never opened for this question. It carries six letters verbatim. The narrow claim survives as a rule: no record exists of anything either man said aloud. |
| V-17 | “Dixon Bailey, to-day one or both of us must die.” | TRADITIONAL | The only battlefield line attributed to Weatherford with a named source — a family letter reporting an aunt who saw him, in Halbert & Ball 1895. Family testimony, not contemporary. Used at the gate. |
What a broken chain looks like
The strongest argument for this table is the row that could not be written. Before the register existed, a draft carried a thesis statement citing page 7 of my own article. There is no page 7 in that journal's pagination. The citation was fabricated, the sentence it supported was fluent, and nothing in the text indicated a problem. It took a deliberate audit to find the break.
This is what fluent generation does by default, and better prompts do not fix it. The four columns above are the defense: a recorded path from every element back to a page someone can open, and an honest label wherever no such page exists.
Source IP: Aaron Zeanah, “The End of the Alabama Frontier: Weatherford's Perspective,” International Journal of Interdisciplinary Social Sciences, Vol. 6, No. 4 (2012). Register ids, decision numbers, and audit-pass numbers on this page correspond to live entries in the project repository. Depictions of the Creek War involve real mass killing of real people whose descendants are living communities; production is gated on consultation.
06 — Prompt Engineering
The audit that inverted
The most instructive prompt failure on this project did not hallucinate anything. It read the project's own safety rule — the sentence forbidding a redemptive reading of a slaveholder — and reported it as evidence of the reading it forbids.
The reviewer was auditing scenes against a blanket rule: the audience should pity Weatherford and not forgive him. Handed the working .md files, it received the screenplay and its [[ ]] annotation apparatus in one document and reviewed all of it as if it were the scene. Task C5 returned 57 hits; two survived. C4 inverted outright, quoting the canon's own prohibition back as a risk.
Before
Three defects. The rule is undifferentiated, so every sympathetic beat is a candidate. Nothing tells the reviewer that a prohibition is not an offence. And the input is the annotated working file, so the directions written to block a reading arrive as candidate evidence for it.
After
Four fixes. The rule becomes a scoped contract (two things forgiven, three not) with the principle stated. The job is renamed — find the transfer, not the sympathy. An explicit negative-case instruction retires the inversion. And the input changes to a script-generated _clean export: apparatus-free, title-block-free, opening on the first slug line.
07 — Governance
The same instruments I ran in regulated product operations
None of this is novel. It is source-of-truth hierarchy, data classification, change control, and release gating — the standard kit from a compliance-heavy environment, pointed at a generative system instead of a support queue.
Canon hierarchy
0-PROJECT/CANON.md · § Authority order
Five ranked levels resolve every source conflict by rule rather than by argument. My published research sits at level one; superseded material sits at level five and is never citable as canon, though nothing is ever deleted.
Evidence tags
CANON.md § The four tags · QUOTE_REGISTER.md
Every claim entering the interpretive tier carries [DOC], [TRAD], [INTERP], or [INVENTED]. The quote register runs six statuses across 20 cumulative audit passes — a later pass may downgrade an earlier verdict, but the earlier verdict stays visible with its correction attached.
Open-decisions register
0-PROJECT/OPEN_DECISIONS.md
105 numbered decisions carrying open, parked, or closed status. Closed decisions are not re-litigated. Every canon amendment cites the decision that authorized it, so a rule can always be traced to the moment and the reason it changed.
Production gates
0-PROJECT/PIPELINE.md · Stage 3
A shot is generated only when it has a numbered ledger entry, a written scene, locked character sheets for everyone in frame, and a location plate. Every generated asset records its model, mode, seed, and exact prompt — without those, the work is not reusable and is therefore lost.
08 — Independent Quality Assurance
A 75-point rubric, run by a reviewer briefed to find failure
The internal apparatus — canon hierarchy, evidence tags, the citation register — is self-graded by design; it has to be, since it runs inside the production. To check it from outside, the project scores itself in dated rounds against a separate Historical Drama Evaluation Rubric and Critical Analysis Policy: 15 categories worth five points each, 13 named critical-failure conditions checked first and capable of overriding any score, and a five-stage review process. Round 1 was run against the full screenplay and research corpus as they stood on 25–26 Aug 2026.
Total score, Round 1
STRONG BAND · APPROVE WITH TARGETED REVISIONS
Critical-failure conditions clear
NONE TRIGGERED · ONE OPEN DEPENDENCY FLAGGED
Scored categories
0–5 EACH, EVIDENCE CITED PER SCORE
Open questions
CARRIED FORWARD TO ROUND 2 BY NAME
Round completed to date
RUN AS MATERIAL WARRANTS, NOT ON A CALENDAR
Round 1 landed at 66/75 — inside Strong, five points under Exemplary. The category holding it there hardest is scored honestly below, not smoothed over.
Source criticism & contextualization — 5/5
Rubric §4
Cited as the project's strongest category: the analysis of Weatherford's two "Creek names" traces both back to a single 1859 source with a documented rehabilitation motive, and refuses to let either replace the poet-invented "Red Eagle" on that basis — full authorship/purpose/audience/bias/transmission analysis applied to one sourcing question.
Verification of names, dates, quotations, events — 5/5
Rubric §7
119 tracked quotations, each carrying a status, with corrections layered on top of earlier verdicts rather than overwriting them. Cross-checks against published scholarship corrected a casualty figure, a command structure, and a burial-record silence.
Legal, ethical & privacy compliance — 3/5
Rubric §13
Scored down deliberately, not softened: no dedicated permissions/copyright risk-assessment document exists yet, and the Indigenous-nation consultation the project's own decision log requires before production has not happened — while 24 units plus the Epilogue are already drafted.
Scholarly & audience-facing disclosure — 2/5
Rubric §15
The lowest score in the round, and mostly a function of production stage: no historical note, bibliography, or composite-character disclosure exists yet as something a reader outside the production would ever see — that apparatus is deliberately walled off from the page until the work is closer to finished.
- Fabricated sources or quotations — not present
- AI text presented as authentic record — not present
- Uncorrected OCR/scrape errors — not present
- One archive as proxy for a society — guarded against
- Archival silence read as nonexistence — guarded & self-corrected
- Cultural material without consultation — open, not yet triggered
Five open questions, carried by name
- 01Indigenous-nation consultation.
Has outreach to the Poarch Band or the Muscogee (Creek) Nation begun, even informally? The closest thing to a live critical-failure risk in the project.
- 02Copyright/permissions assessment.
No such document exists yet. Draft one from the acquisitions log, or commission separately?
- 03Ten reference videos, unknown license.
Correctly walled off as "modern interpretations, never sources" — but origin and license are still unrecorded.
- 04Scoring philosophy for stage-appropriate gaps.
Scored low because the deliverables don't exist yet, not because anything about them is wrong. Treated as N/A instead, the total lands near 72–73/75.
- 05Independent bias/red-team readers.
The rubric calls for reviewers outside the original research/writing loop. Planned, and on what schedule?
This is the same discipline the architecture diagram names as Citation Audit and Final QA — made concrete, dated, and scored, with its own gaps published alongside its strengths.
Reusable assets
Four classes of artifact, built once for Season One and written to be storyline-agnostic. The badge on each group says how much survives a change of subject — the honest answer is that configs and checklists port intact, while anything carrying period detail is rebuilt.
Templates
6 · document formats
- Delegation packetStanding format, naming convention
DATE__SERIES-X__slug__vNN, and instructions carried in every packet. - Disposition recordRaw return preserved verbatim beside the judged version, so a reviewer's wrong call stays inspectable.
- Scene design docWhy each movement is staged as it is, with the decisions it answers to.
- Registry row
id | subject | model | mode | seed | prompt file | reference | locked - Condition reportWhat in the corpus is readable, what is not, and what that means for the work.
- Session-start checklistSix steps, in order, before any task begins.
Prompt library
9 · 2 retired clean
- G0 — access gateRuns before any task. Line counts and first/last lines, or URL plus repository plus printed page.
- C4 — transfer checkScoped forgiveness contract. Carries the negative-case clause.
- C5 — the wallOne test: does a character who cannot see the system state it?
- C6 — numbers and datesList without judging. Highest-yield task in the packet.
- C7 — atrocityFlag beauty applied to violence; plain description is correct.
- C1 · C2 — voice, anachronismClosed: clean across all scenes, independently re-verified.
- Series M · V · WResearch, page verification, and provenance dating briefs.
Agent configs
4 · persistent instruction blocks
- EMPLOYEE_BRIEF.mdThe permanent context, deliberately separate from task Directions so a task can stay short and a fresh session still behaves.
- Scope contractA ✅/❌ table of what the role is and is not asked to do — written after work was sent that it could not do.
- Reply contract
TASK / STATUS: COMPLETE · PARTIAL · NOT FOUND · CANNOT DO, then the answer, nothing else. - PROJECT_INSTRUCTIONS.mdOrients a session; explicitly does not carry the decisions.
QA checklists & scripts
5 checks · 5 scripts
- Stage 3 gateLedger entry + written scene + locked sheets + location plate. All four, or no shot exists.
- Continuity checklistOrder of operations for applying a dispositioned return.
- Cost disciplineModel cost before any batch; one test image before a set.
- check_age_spine.pyToken list carries word and digit forms.
- check_scene_refs.pyEvery register id cited on a page resolves.
- build_exchange.pyRegenerates clean exports immediately before send.
Five lessons, logged as changes
Each of these is a dated failure in the delegation log with a root cause and a control that followed. None were anticipated in the plan. The rightmost column is what makes them reusable rather than merely learned — a lesson without an artifact is a lesson that has to be learned again.
A research return that was fabricated in full
2026-08-22Problem
An agent answered a complete research task about a 673-page volume it had never received — inventing page numbers, transcriptions, an index, and an editorial note. All internally consistent. All false. A later report then repeated one of its claims by citing "the Volume III research."
Root cause
The access check tested content. It asked the agent to transcribe two lines as proof, and the agent transcribed two invented lines. A model without a source produces exactly the artifact the check was looking for.
Change
Gate on a fact already held, never on content. Blind-first ordering: the task naming the string being chased runs last, after the page is already transcribed. The void report was quarantined by name so nothing reasons from it.
Reusable artifact
G0 gate block, plus the access section of EMPLOYEE_BRIEF.md. Both are subject-independent and shipped in every packet since.
An audit that reported the rule as the violation
2026-08-24Problem
The constraint reviewer read the canon's own sentences forbidding a redemptive reading and returned them as evidence of that reading. One task returned 57 hits of which 2 survived; another inverted outright, quoting the project's rule against itself.
Root cause
Two defects in what the packet supplied. The rule was undifferentiated, so every sympathetic beat qualified. And the input was the annotated working file, so directions written to block a reading arrived as candidate evidence for it.
Change
Rule rewritten as a scoped contract — two things forgivable, three not. An explicit negative-case clause: a prohibition is not the offence it prohibits. Input switched to a script-generated apparatus-free export.
Reusable artifact
build_exchange.py and the _clean/ export convention with its manifest, plus the negative-case clause now standing in every constraint prompt.
Two verdicts condemned by one source's silence
Pass 006Problem
Two quotations were marked FABRICATED and purged. Both were real. The register — the instrument built to stop fabrication — had produced two false positives of its own.
Root cause
Absence of evidence read as evidence of absence. A single source was searched, came back empty, and the empty result was recorded as a finding rather than as a search boundary.
Change
No quotation is condemned on one source's silence. A FABRICATED verdict now means not found where we looked, and the looking is named. The register is cumulative and never rewritten backward — a downgraded verdict keeps its history with the correction attached.
Reusable artifact
The six-status register schema, and the append-only convention that makes a reversal visible instead of silently overwriting the record.
A correction stamp that outlived its correctness
2026-08-25Problem
A line carrying four character ages was stamped "corrected to the written pages" — two days before the decision that moved the underlying spine. All four ages were wrong, in the file that tells an illustrator what to draw.
Root cause
Two, compounding. A stamp was trusted on its own confidence rather than its date. And the validation script that did surface the lines was never read — and could not have caught them anyway, because its token list held word-forms while the file used digits.
Change
Check a stamp's date against the decision it claims to satisfy. Token lists carry both notations. Validation output gets a named owner, because a check nobody reads is not a check.
Reusable artifact
check_age_spine.py with its dual-notation token list, and the standing rule: a correction stamp is not a guarantee.
Thirteen delegated tasks superseded by one command
2026-08-23Problem
A whole task series was written to have an agent read volumes it could not receive, to answer questions a single search would resolve. Two more packets had a first step that was an impossible upload. The defect was invisible in the shape of the document.
Root cause
The boundary between mechanical work and judgment was never written down, so tasks were scoped by what seemed helpful rather than by what the agent was actually good at.
Change
The coordinator searches; the delegate reads and judges. A ✅/❌ scope table now sits in the standing config, and the first step of every packet is checked against it before writing the rest. Five packets have been withdrawn pre-send since.
Reusable artifact
The scope contract in EMPLOYEE_BRIEF.md and the scope rule at the head of TASK_QUEUE.md — plus the archive convention that keeps withdrawn packets with their stated reason.
Season Two roadmap
The expansion is already reserved in canon. Texas and the Confederacy are named as the far end of the argument and are deliberately never depicted in Season One — carried by geography alone, with Montgomery planted once by a surveyor as pure place-name.
The constraint is the asset
Season One forbids anticipation: no character foresees what is coming and no line points forward. That rule exists to protect historical integrity — the horizon is a place on a map, not a prophecy. It also happens to be what keeps the sequel viable. A first season that foreshadowed its own continuation would have spent the second season's material to decorate the first. The discipline that protects the history protects the franchise. That is not luck; it is what happens when constraints are written down and held rather than negotiated per scene.
PHASE 0 · NOW
Close the first cycle
Segment 07 is the only open segment: style bake-off, then character sheets, location plates, shot specs, animation. Nothing about the framework is proven until one cycle ships end to end.
GATE — style locked on four images before four hundred
PHASE 1 · BEFORE ANY ASSET
Consultation
Depiction of the killings is gated on consultation with the Poarch Band of Creek Indians and the Muscogee (Creek) Nation. It also settles the Muscogee-language name placeholders — which is why it must happen before registry rows exist, not after.
GATE — D-08 · no name baked into a registry row
PHASE 2 · SEASON TWO
The Texas storyline
Offscreen in Season One, heavily alluded to, reserved. It inherits every control unchanged and needs its own tier-1 corpus, its own condition report, and its own audit passes before a page is written.
GATE — audit before draft, by act
PHASE 3 · UNIVERSE
Connected storylines
Recurring figures carry their locked sheets and their forgiveness ledgers across seasons. Shared ground reuses its plates. The registry becomes the continuity spine of a universe rather than a film.
GATE — seed and prompt recorded, or the asset is lost
ZERO REWORK
Ports unchanged
Agent configs, the reply and scope contracts, the six-status register schema, the four evidence tags, the five-level authority order, gate definitions, packet format, the QA prompt library, and every validation script. These describe how work is governed, not what it is about.
INHERITED, EXTENDED
Carries forward
Locked character sheets with their seeds, for anyone who recurs — several already built to age across four decades. Location plates for shared ground. The style skeleton, if the visual world holds. Structural devices whose payoff spans seasons.
REBUILT PER STORYLINE
Starts from zero
The source corpus and its condition report. Story-specific canon — POV assignments, the wall's particular prohibitions, the forgiveness ledger for new figures. Character bibles, the scene ledger, and every quotation in the register. Nothing about the evidence is reusable, and it should not be.
12 — The System in One Paragraph · Contact
What it is
A four-tier artifact pipeline — academic sources, interpretive bibles, screenplay, production — where material moves upward only and nothing enters a tier without a locked entry beneath it. Four evidence tags ([DOC] [TRAD] [INTERP] [INVENTED]) classify every claim; a six-status citation register clears every quotation; 105 numbered decisions govern change; and two of the gates are human. Invention is permitted everywhere — it just has to declare itself, in writing, every time.
Why it exists
An audit ahead of drafting found roughly two-thirds of the quotations in the inherited drafts did not survive contact with their sources — including two fabricated citations to my own published paper, one serving as the thesis statement. The system was built against that failure mode, and its controls have since caught a fully fabricated agent research return, two false verdicts of its own, and two continuity defects no human had noticed. The film is the output. The framework is the product.
Aaron Zeanah
Product & Customer Strategist — Project Coordinator to Business Analyst to Product Operations Manager to VoC Product Partner to Lead CX Strategist, across Intercontinental Hotels Group, Coca-Cola, WarnerMedia / NBA Digital, and FanDuel. Trained historian (B.A. History, Auburn; published 2012). I design product systems for environments where provenance is not optional, and I'm bringing that discipline to AI-first product work.
Scan to open this portfolio
Source IP: Aaron Zeanah, “The End of the Alabama Frontier: Weatherford's Perspective,” International Journal of Interdisciplinary Social Sciences, Vol. 6, No. 4 (2012). All figures on this page were audited against the project repository — file counts, register ids, decision numbers — on 26 Aug 2026, before publication; the zero-fabrication result is scoped to the scenes covered by the adversarial review series. Depictions of the Creek War involve real mass killing of real people whose descendants are living communities — the Poarch Band of Creek Indians and the Muscogee (Creek) Nation; production is gated on consultation.