Repository documentation
Production plan — Eight Hours to Make Coffee
State: completed 1.3.1 media goal. The 12-chapter/86-turn, 12:04.896 film uses newly recorded approved-voice takes. The standard verifier accepts its 86 captions and continuous media. Fresh scoped execution evidence is retained, and the user approved listening, factual, accessibility and normal-speed playback against exact media hashes. The review-bound rebuild is byte-identical and is published alongside its captions, receipt, poster, manifest, QA and review provenance.
The exact candidate revision's fresh evidence records a ready host and read-only adoption plan over an unchanged eight-file native fixture. It does not claim apply, build, retained-harness or native qualification.
Version and factual contract
Use the research ledger and snapshot as source authority for project case studies. Freeze the actual framework version used for filming and refresh the inspected project commits before production. The prior film demonstrates v1.2.0; v1.3.0 is the currently published wheel and this replacement is a blocking v1.3.1 goal. Identify development-only functionality on screen; do not imply that downloading an earlier wheel enables every source-tree feature.
The narrative's project test counts are records at named commits, not tests rerun for this film, comparisons of whole-pipeline speed, current-HEAD green guarantees or equivalent units: BookScan 126 retained tests; Aviation 273 retained tests; ElixirSSI four verification stages; LeRTX 160 tests in a recorded ARM64 native admission milestone, distinct from its standard project receipt. Daemonarchy is an in-place contract adoption with native lifecycle work still open. Keep those limits beside the facts. Public source evidence cannot establish present runtime behavior on the filmmaker's untested machine.
Sam's eight-hour pipeline is a fictional anecdote that motivates the argument. Never render a comparative stopwatch, invent elapsed times for commands, or imply an eight-hour native workload universally finishes in minutes. The mechanism being sold is earlier useful feedback, deliberate target selection, eligible parallel execution, reproducible inputs and retained evidence—not less testing.
Chapter construction and measured timing
Compose the episode from separately prepared and reviewed chapters. The current pilot authoring path caps a scene package at 180 seconds and 40 dialogue turns. Check each chapter against measured recordings and split it if necessary; do not raise limits just to fit a draft. The numbered speech lines are the exact dialogue; stage directions, citations and proof labels must never be sent to TTS as narration.
Retain the approved stock requests:
- Sam:
Magpie-Multilingual.EN-US.Jason.Neutral. - LitAI:
Magpie-Multilingual.EN-US.Aria.
Cast approval does not approve new utterances. Audition the dry punchlines, technical names and turn-taking with the actual provider's available voices. Do not guess style IDs or silently fall back to system speech. Keep new recordings and generation provenance in new custody, preserve the old WAVs, and map phrase cues from the new recordings. Do not reuse the pilot's word times for changed dialogue. Model-effective voice/checkpoint fields remain unknown unless the provider actually reports them.
Allow quiet reading time for source excerpts and the five project case studies. Short punchline pauses are authored intentions; only measured audio and a review cut establish the final timeline. No canned laughter or laugh animation is needed. Keep Sam's objections intelligent until his waiting excuses become self-defeating. LitAI punctures the ritual while preserving the engineering rigor he values.
Animation realization
Retain the accepted stable workbench, two-joint foreground arms, meaningful gaze, expressions, purposeful holds and static alternative. The existing pilot only implements its own finite target/pose vocabulary. The full script now compiles into finite per-turn intentions and twelve chapter-specific visual panels, with a bounded reach-and-settle performance instead of an idle gesture loop. Synthetic structural QA does not approve the performance: inspect every actual voiced chapter and reject any unsupported beat rather than mapping new dialogue back to a generic gesture.
| Capability | Direction and realization requirement |
|---|---|
| Excuse stack and mug | New simple owned vector props; COFFEE, LAUNDRY, GARAGE, SPICES and THINKING cards. Contact points must follow the hand while moved, then release and settle. No idle wobble loop. |
| Recipe and dependency graph | Explicit compatible targets, worker placement, blocked cells, dependent order and eligible parallel branches. Use labelled diagrams, not fabricated execution progress. |
| Independent evidence | Distinct plan, generated tests, native result, independent acceptance and required CI/release cards. An old pass cannot migrate to changed inputs. |
| Actual project views | Legible source excerpts, real captured UI where available, and scoped committed results. Keep source provenance off the spoken soundtrack; retain short on-screen source captions. |
| Native/emulator distinctions | Explicit emulator, Metal, NVIDIA, development-only FreeBSD and open native-gate labels. No badge may imply untested hardware or a completed adoption. |
| Ending | Sam arranges the next engineering step himself. Coffee is beside the graph, not a dependency. Both settle; hold useful commands rather than wave. |
At each phrase cue, prioritize the speaking actor, the listener's delayed response or a diagram reveal—one at a time. Do not move both sets of hands throughout a technical explanation. Let Sam's growing competence alter who directs attention. The joke props support the argument rather than replace the real project visuals.
Static/reduced-motion export must preserve the same facts, exact distinctions, commands and source scope through settled poses, readable labels and transcript. No crucial meaning may require noticing an eyebrow or hand movement.
Fresh capture plan before rendering as a demonstration
| Chapters | Required material | Honest alternative until obtained |
|---|---|---|
| 01–02 | Original diagrams and recipe/authority illustration | Always labelled illustration; eight-hour anecdote remains fictional. |
| 03 | Exact wheel acquisition/install, activation and litai doctor; read-only plan on a disposable nontrivial repository |
Commands labelled walkthrough; retain the already proven pilot only at its exact two-file scope. No implied apply. |
| 04–05 | Selected target/worker readiness and one actual scoped failure/recovery before CI; real typed/native or API checks with retained receipts | Source contract and planned-check diagrams; no fake green terminal output, broad matrix or measured speed claim. |
| 06 | BookScan UI capture if authorized/available; named test source and recorded retained result | Source-backed workflow diagram and explicitly recorded repository evidence. Do not imply a fresh camera test. |
| 07 | Aviation retained Make entrypoints, analyzer contract, test/source excerpts and scoped receipt | Source walkthrough; do not imply every optional suite ran or certify real flight operations. |
| 08 | ElixirSSI verifier, four-stage image report and clearly emulated three-node topology | Existing committed source/report view with scope labels, not new physical-board execution. |
| 09 | LeRTX native milestone record and actual native frames only when retained provenance is available | Contract/record view; no mock frame masquerading as current RTX success. New SDK refresh stays pending. |
| 10 | Daemonarchy retained contracts and explicit native-admission limits | Source/architecture view; no disk writes, boot or release-success demonstration. |
| 11–12 | Actual project gate policy and useful next-step card | Workflow illustration; required CI and release gates stay in force. |
Fresh execution remains subject to the repository's normal host/worker permissions, readiness and disposable-project workflow. Research did not execute the example applications or qualify their current HEADs. Capture source revision, commands, exit status, checked target and retained artifacts before calling any new result an observed demonstration. Source-backed diagrams need no invented live result.
Review and delivery gates
The full script, compiler and chapter choreography are authored. Every line was recorded and aligned after live catalog discovery. Every chapter and the assembled film were checked at playback size and normal speed; the final approval is recorded against the exact reviewed bytes rather than inferred from the accepted pilot.
Run score admission, audio/source drift refusals, full-timeline pose continuity, caption bounds, offline/keyboard and reduced-motion/static checks at the actual chapter scope. Assemble with the existing video CLI and independently verify the final manifest and media. Retain captions, source records and a local review copy. Voice, visual, factual and accessibility judgments are recorded in the exact approval evidence. The user authorized GitHub publication as part of v1.3.1. Retain the prior film and make all new front-door links resolve to the approved media identity.