GhostReel v0.4.0

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage or .deb · CLI only: ghostreel-cli-*.

GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). On first start open Models and download the models.

---

Added

• Refining, as something that repeats and keeps the best. ghostreel script refine -p P --session N --rounds 3 asks the chat brain to improve a cut several times and keeps the best
  round the judge saw, never the last — a round can make a cut worse, and one did. On the Greet
  Mag brief: Jev's build at 61, then 66 / 67 / 63 across three rounds, with flow going 0.57 → 0.93
  and the ending 0.49 → 0.84. With Jev off there is no fitness to climb, so it degrades to plain
  repeated turns.
• "Jev drafts first" in the app. A toggle beside the chat box turns Send into the whole chain:
  Jev chooses a cut out of the index, the judge reads it, and whichever brain Settings names
  refines it from there. This is what makes the chain worth having for a local model — writing
  from scratch the local Qwen scored 38 mechanically and 56 editorially; handed a built cut it
  scored 98 and 64, the same structural lift agy gives, because it is no longer doing the research
  it is bad at.
• The cut closes on a picture rather than a talking head. hold_the_last_picture used to
  extend whatever clip was last, which on a speaker gave two seconds of somebody moving their
  hands with the sound already finished. It now cuts away to a described shot with nobody talking
  in it, chosen by what it shares with the closing line. Only under a bed: without one the closing
  voice is that picture's own audio.

Fixed

• The interviewer pass was asking Jev the wrong way, twice. Accuracy falls off with the length
  of the state rather than the number of questions — an unmistakable interviewer line reads 0.56
  against 63 lines and 0.96 against 20 — and serde_json::Map is a BTreeMap, so the framing key
  what_this_is sorted after lines and twenty lines of dialogue arrived before the sentence
  explaining what they were. That cost four of fourteen subject lines to false positives on a
  labelled set: the speaker's own words scoring 0.89 as the interviewer, 0.46 with the framing
  first. Windows are 20 lines now and the key is named to sort first, with a test that says so.
  GHOSTREEL_DEBUG_JEV dumps each request.
• A beat ran on into the reply over the top. One ended three segments past the speaker's last
  line, through "Yeah." "Yeah." into the interviewer starting a story of his own — ten words of
  ordinary English that no word count reaches and Jev scores 0.20. end_on_turns reads a run of
  two acknowledgements as a handover and keeps the longer side, so a pause inside somebody's own
  answer is left alone.
• Refining barely changed anything, because it was barely told anything. judge::notes only
  spoke below 0.4, so a cut scoring ending 0.49, opening 0.56 and brief 0.68 heard about none of
  them and changed only the one beat it was told about. The judge now always names the two
  dimensions losing the most points, ranked by points lost rather than by score.
• Every refine turn was judged against the wrong text. The turn passed its own message as the
  brief, so "says what was asked" scored the cut against "Improve this cut…" rather than against
  what was asked for. The same script read 67 in the turn and 76 against the real brief, four
  times running. It reads the session's first message now.
• "Build with Jev" in the app was a dead end. It saved the cut with no session and never
  judged it, so save_as_session had one caller in the repo — the CLI — and the built script was
  filed under the panel's orphans rather than under the conversation about to refine it.
• script import had drifted from the repair sequence it is meant to mirror, missing
  clamp_beds_to_beats entirely.