Appendix B — Capability matrix and tool inventory

Status: DRAFT 1 (2026-09-21), DESIGN v6 §7, §12. Column "verified on" is the date of a ledger row or file that shows the capability in use on HALFWAY; "reproduced on another tool" is empty by decision: the §7 test was not run (Director, 2026-09-22).

B.1 Capabilities the process depends on

Capability Required? We used (brand) Verified on (evidence) Fallback when absent Reproduced on another tool
Identity from ordered reference images required Nano Banana Pro, Seedance 2.0 image references 2026-09-11 (ledger row 1, tags @char_HW_Crystal_hairup_v1 …); 2026-09-16 (row 588, refs uploaded into project, price read 2) one hero still per shot as start frame + text identity block + face-check sentence; more stress takes —
Start frame from an in-project still required for dialogue shots Seedance Start Frame 2026-09-13 (row 272: "tile → Turn to video → Use as → Start Frame"); LESSONS.md L7 take 10, 2.7 px drift image-to-video on any tool; else shorter shots, cut on action —
Motion reference video optional Seedance video input 2026-09-16 (rows 571–572, job 26d72daa, 5C stubble retake) describe the motion beat; tier budget covers the retake —
Audio reference conditioning optional Seedance audio input 2026-09-14 (rows 393–395, 1B proof: conditions only, timing not honoured; no controlled comparison) generate silent, place the line in the edit, accept per DESIGN §6 sync rule —
Dialogue sync acceptance required (some path) placement by window + Director ear/eye; Sync Lipsync 3 prepared, declined 2026-09-16 (rows 499, 504, 511) dedicated lip-sync tool, coverage rewrite, or shorter line —
Price visible before spend required Cinema Studio composer button, zoom-read every job row ("button read 30 struck → 23", row 589) compute from the rate card, ledger as computed not read —
Batch TTS, one file per line, tone tags required ElevenLabs v3 in Higgsfield (Naomi, Holden) 2026-09-12 (row 65, 0.15 per line); AUDIO-PLAN.md:61–72 any per-line TTS; tone by rewording —
Voice change keeping timing optional Higgsfield Voice Change 2026-09-16 (rows 540, 557; 0.1/s rounded up) re-record with the target voice and re-cut —
AI score required (some AI music source) MusicGen large on a private GPU 2026-09-16 (rows 576–578, 593; cues uploaded, assets 266 → 271) hosted AI music tool; no GPU on the required path —
Read a take: strips, crops, transcript, loudness required ffmpeg 9, whisper.cpp 1.9.1, Pillow, numpy every read row; tools/media.sh same, free everywhere —
Assembly from a manifest, never overwrite required ffmpeg build script per cut every cut row; templates/build_cut.py tested 2026-09-21 (240/240) any NLE with a written shot table and per-cut folders —
Ledger required csv + git, Python csv.writer production/phase3/ledger.csv (606 rows at 2026-09-21) spreadsheet with the same columns —
Jobs and balance readable from the account strongly recommended platform jobs API and user API 2026-09-16 (rows 550, 603: balance and 273-job export) the account page, read by hand, after every job —

B.2 What was not on the required path

B.3 Tool inventory as run (moved from Chapter 1 draft 1)

Versions read 2026-09-21 on the production machine.

Tool Version What it did
ffmpeg / ffprobe 9.0.1 (Homebrew); no drawtext filter in this build every trim, grade, sheet, strip, loudness read, assembly, full decode
whisper.cpp whisper-cli 1.9.1, model ggml-large-v3-turbo.bin (1.62 GB) transcripts; 10 s clip in 1.8 s wall on this Mac
Python 3.9.6, Pillow 11.3.0, numpy 2.0.2 build scripts, caption PNGs, RMS windows, ledger rows, Appendix A
git 2.54.0 history of every text file
Agent Claude Code with a Chrome extension driving the Director's signed-in browser; Codex for adversarial passes operator

tools/media.sh commands used: probe, frames, sheet, first, firstsec, audio, loudness [--target --tol --peak], transcribe <file> <lang>, img [--luma-min …]. Two properties: a bare call only reports and never claims a pass; every failure exits non-zero with a sentence. One known defect, kept as a teaching case (Chapter 6): firstsec printed PASS on an empty mask (LESSONS.md:128).

B.4 Reproducibility test (DESIGN §7): NOT RUN, by decision

One representative scene of a fresh story through every mandatory stage on a clean laptop with no GPU and none of our private skills: bible → planning card → final prompt → fenced generation → read → line placed and sync-accepted → assembled → hosted AI score under it → exported and decoded. The Director decided on 2026-09-22 not to run it. The "reproduced on another tool" column therefore stays empty, Chapters 9–15 carry the status "verified on one tool", and the first learner on another platform is the test; their Chapter 16 forms are the result.