A translation pipeline for feature-film screenplays

What I designed: 11 agent roles, an orchestrator skill that calls them stage by stage, 9 ground rules and 26 tools (17 for the translation harness, 9 for the external benchmark). The roles sit in three layers: source anatomy (AI director, structure writer, detail writer, forms-of-address keeper), transfer (translator, cultural adviser, performance reviewer) and judging (auditor, format keeper, AI readers, AI investors). Nine run on the Opus family and two on Sonnet. The orchestrator is a skill: the main session runs the stages and calls each agent by naming its definition file.

Call records survive only for sessions from Sep 26 on. The Work C session called agents 33 times and the Work D session 51 times. Of Work D's 51 calls, 47 named one of 10 role definitions (8 each for the AI director, the translator and the auditor, among others) and 4 used no definition. The format keeper has no calls in the surviving records; the format-agent step was cut when the steps were trimmed on Sep 17.

The three decisions below that changed direction or standards are on record as the "user's", and I was the one running those sessions. On Sep 16 I pointed out that the sections translated in parallel contradicted each other and had them redone in sequence. The same day I set the principle that a proven fix is not finished in one script but goes into the whole pipeline. On Sep 17, noting "this is taking far too long", I had the steps that found nothing new removed. Market, purpose and naming for each script are put to a person; anything unanswered uses the previous script's decision as the default and goes to the report's user (production company) confirmation section as pending. The reports list 4 pending items for Work B, 3 for Work C and 8 for Work D.

Human decisions rank above the AI director. The order is "user decision > director's locks > lexicon", and the director never re-judges a user decision.

Whether meaning gets lost across cultures is checked separately. The cultural adviser writes neutral questions about what a reader of the source takes away, and AI readers who saw only the translation answer them. Answers are sorted into arrived, partial and lost; a lost meaning goes back to the AI director even if the item is locked.

The checking tools are checked too. The section gate plants quoted dialogue, a deleted scene and an altered locked string in a copy of the script, and if it misses any of them its results are not trusted. Packet leaks are measured by giving a fresh agent that does not know the answer either the source alone or the source plus the packet, and having it guess the twist. The hit rate has to stay within the source-only baseline plus 10 points.

The four AI readers and five AI investor profiles read only a copy of the translation kept outside the project. As the reports say, this is not audience research, so this write-up does not count screening or investment verdicts as results. Zero hallucinations and matching dialogue-block counts are also the verdicts of the pipeline's own gates. The record itself notes the limit that one model family is grading its own family's translations.

There was no CI, no PRs and no hooks. What enforces the gates is the wording of the skill and the role definitions, plus the tools' exit codes (the section gate blocking, the tone-mark check failing). Re-run on Nov 10 for this write-up, 11 of the 12 tool self-tests passed; one could not run because the DOCX library was not installed. These are not a test suite with counted assertions. The regression results from the Oct 2 batch of fixes (self-tests 8/8, gate self-checks 8/8) come from the work log and were not re-run.

Twice, a "pass" was not a pass of any check. The first leak check confirmed "zero remaining" with the same function that had filtered the packet. The record says "if the checking tool and the generating tool are the same, it is not a check", and the check became a blind guessing test by an agent that does not know the answer. On Oct 31, in the first Korean → Vietnamese script, a tool that assumed Vietnamese source text skipped every lock comparison and showed "pass" without a warning. The lesson written down afterwards: "on the first script in a new language pair, first check whether zero findings means zero or not run".

A tool deliberately left out at first became the core of the gates. The design document said writing a script parser and page counter in Python was "left out on purpose… if it becomes a bottleneck, a 150-line stdlib script then". In practice the first batch of commits already held three intake tools, including the parser (85 lines), and those tools became the core of the gates and, with the external-benchmark tools, grew to 26 (4,816 lines). Counting that needs judgment, though, was not handed to a machine (decision 3).

Records of finished runs were overwritten twice. On Oct 1, while taking a regression baseline for tool fixes, the originals of some of Work B's packets were lost; an output-location option and a rule that "finished run folders are records" were added. On Oct 31 the orchestrator again overwrote Work D's first-section packet and gate results. This time they were rebuilt from the revision backups, checked to be byte-identical, and restored.