Against the plan

1. v0 · built the log and one worker planned: an events table, projections as views, one worker that wakes departments by subscription in the prototype: an append-only array in the browser, projections rebuilt on every event, the five-role harness, subscription and staleness
2. v1 · not built durable execution planned: workflows that sleep for days waiting on a human, activities with retries and heartbeats in the prototype: a wake is a chain of fetches in a tab; close the tab mid-round and the round is lost, though the log is not
3. v2 · built the board planned: board, margins, call sheet, ledger over server-sent events; inspector last in the prototype: all six surfaces, including the inspector, the scrubber and export / import of the whole diary
4. v3 · not built telemetry and provenance planned: OpenTelemetry spans; a manifest on every rendered asset; signed credentials at export in the prototype: frames are rendered by hand outside the council, so there is nothing yet to attach a manifest to
5. v4 · built offers and dailies planned: the extra question on every wake, a surprise budget, the recorder’s digest in the prototype: one offer per department per wake; dailies written from the log at every gate close, with the roads not taken

The design document laid out a build order: the log first, because everything else is a view of it; durable execution second, because the first render that dies at minute nineteen teaches you why; then the board, then telemetry and provenance, then offers and dailies. The prototype took the first, third and fifth steps in miniature and skipped the middle two.

What held

The one idea. The event log is the production diary, and every screen is a projection of it. This was the first thing built and it has not changed. Time travel, the why chain, the ledger, export and import all fell out of it as the design said they would.

Departments as data, one harness. Seven rows, one five-role wake, mandates as prompts. Adding the production and post gates was adding rows and prompts; the harness did not change.

Single writer, flags to the Director, guess or ask, convergence as a table. All four rules are in the code and enforced by code, not by prompts.

Offers and dailies. The last thing the design said to build, and the thing people notice first. Built.

What changed

the design · nine

• Producer
• Director
• Story
• Characters
• Art
• Camera
• Sound
• Unit
• Editorial

→

the build · seven

• Producer
• Director
• Story
• Look & Cast Characters + Art — on a short film, whose clothes these are is one question about the person and the world
• Camera
• Sound & Music
• Unit + Edit Unit + Editorial — nothing renders inside the council, so the editor’s job at production is to write the sheet the human renders from

Nine became seven. On a ten-second film, Characters and Art are one conversation, and with nothing rendering inside the council the render unit and the editor have one job at production. Both merges are one row each.

First drafts are ordered. The design said order falls out of subscriptions and the spine, with no orchestrator. That holds for everything after the first draft: revisions, reviews and rulings are driven by staleness, and nothing schedules them. But an empty artifact emits nothing to subscribe to, and the first shot list cited character ids nobody had written. So the opening move of each gate is scripted: Story, then Look & Cast, then Camera, Sound and Unit in parallel. The orchestrator the design refused to have exists for exactly one step.

Artifacts are written in parts. The design assumed one call per artifact per wake. The models could not hold a shot list or a render sheet in one answer, so every artifact is written in parts, pinned to the ids asked for, with retry passes that shrink the batch until it is one id.

Rounds are capped at four, not three. One more round of argument was worth the cost on the runs so far.

What was skipped

Durable execution. A wake is a chain of fetches in a browser tab. Close the tab mid-round and the round is lost, though the log is not: every completed step was appended before the next began, and the board’s recovery buttons are computed from what is missing. What is missing is the ability to wait a week for an answer without a tab open, and to tell a slow render from a dead one.

The tool gateway as a service. It is a function. It enforces the same rules (only the owner writes, flags target artifacts, a denial is an event) but a function can be bypassed by code that does not call it, and a service cannot.

Telemetry and provenance. No spans. No manifests, because frames are rendered by hand outside the council and there is nothing yet to attach a manifest to. The design’s argument for provenance (signed credentials on every frame as a legal and festival surface) waits on renders happening inside the system.

What is untested

A critic from a different model family. The design’s bet is that two families rarely share a blind spot, and a critic that agrees with everything is worse than none. The prototype runs one model in every role. The actor field on every event records which, so the bet can be measured when it is tried.

Autonomy that departments earn. The design’s slider from “ask me for everything” to “take the guess and tell me later” is a mandate in the prototype, fixed per gate.

Forks and skill replay. Branch the project at any event and answer differently; replay last month’s project against a new version of Camera’s rubric and diff the flags. Both are a copy of the log up to an event plus a re-run, and the prototype has the copy but not the re-run.

More than one user. The question the first design left open (turning a single-user design into a multi-user system) is still open. The event log is the answer’s first half.