The source behind each decision.
Specification excerpts, TypeScript, SQL and run logs for all twelve decisions. Excerpts are quoted verbatim; internal paths and identifiers are redacted.
Figures come from test environments, not customer traffic, unless noted.
Share the machinery, let the outside assistant plan
The rule that calls from outside assistants go only through the tool registry, the bridge code that follows it, and the usage accounting the bridge has to restore.
Read the sourceLet one required feature decide build versus buy
The pass-or-fail rule written before the sign-in provider test, the results table, and the response that ruled the provider out: unsupported scope intel:read.
Read the sourceOne owner per fix, reviewed with fresh eyes
The dated decision that merged the fixer and reviewer roles, and the review contract every production push goes through.
Read the sourceTest through the customer’s door
The closed issue: the assistant’s false “no profile linked” claim, the per-turn fix, the deployment, and a fresh-conversation retest by the customer-side tester.
Read the sourceSave the applied work before the optional review
The ordering fix and the restart log: resumed at segment 2, segments 0 to 4 applied once each, 360 credits on the receipt, the job record and the ledger.
Read the sourceSet the credits aside before admitting paid work
The database lock shared by reserving and spending credits, the capture path that rolls back on a failed charge, and the three-way concurrency test.
Read the sourceGive the AI four minutes at a time, from one shared document
How the video is split into four-minute segments, from the architecture spec and the planner code: a four-minute target, a clock fallback when no sentence or scene boundary fits, and a deliberately short final segment.
Read the sourcePlan each segment against the video as it is now
The locked decision to edit in order, and the planner’s input: the document is re-read before every segment, and the film-wide choices travel with each call.
Read the sourceTrusting the quality checks
Prove the checks can fail before paying to scale
The spend gate that blocks paid runs on a failed or stale calibration, the black-and-white false positive, and the test run that proves the fix.
Read the sourceAccept the exact film against a written rubric
The rubric’s gating rules, the recorded verdict for the reviewed export, and the caption gap a later NASA-footage run exposed.
Read the sourceBudget for the whole container, not the program
The fallback order, the assembler as built, and the receipt from the test run: 215 MB of process memory against a 7.23 GB container peak.
Read the sourceReprice from measured work
Measured rendering time (4.2 s per output second against 2.78 modelled), the cost comparison, and the 1.5 to 2.25 repricing decision.
Read the source