MorganTitcher's picture
v1.1.0 Sync from inkwell 9bdb21e57
63d3244 verified
|
Raw History Blame Contribute Delete
3.04 kB

Evaluate a constrained revision

Give a writer or model a record's input and instruction, then preserve its exact response before inspecting it. cases.jsonl contains the tasks; calibration.json contains example responses with expected mechanical outcomes and a separate explanation for a human reviewer. These examples are not unique correct rewrites; reference_output remains null.

In the Inkwell workspace:

import { evaluateRevision } from "@inkwell/screenwriting-research/revision-evaluation"

const task = {
  id: "example",
  anchor: "as far as the gate",
  max_words: 50,
  review_questions: ["Does help remain an offer rather than pressure?"],
}
const result = evaluateRevision(task, "Ren carries the case as far as the gate.")
console.log(result.wordCount, result.nonempty, result.checks)
console.log(result.reviewQuestions)

Run the evaluator and calibration tests from the repository root:

pnpm exec vitest run packages/screenwriting-research/src/datasets/revision-evaluation.test.ts packages/screenwriting-research/src/datasets/hub-kits.test.ts

Interpret the result

  • nonempty rejects an absent response independently of its word count.
  • max_words counts all whitespace-delimited runs, including headings, speaker cues, punctuation-only runs, and markup. The ceiling is inclusive. This is an English-oriented counting convention, not multilingual word segmentation.
  • contains_exact is a case-sensitive match of the anchor as whole words: it is not found inside a longer word, so key does not match monkey (a hyphen or punctuation next to it is a boundary). It does not normalize Unicode, punctuation, or spacing. It establishes presence, not placement: a preserved final sentence still needs review to confirm it is last.
  • no_preface is flagged or not_flagged, not a proof of compliance. It flags leading code fences, “Sure”, “Certainly”, “Of course”, “Here's”, “Here is”, “Here are”, and “Revised scene/version:” or “Rewritten scene/version:”. Case is ignored. A legitimate opening can match; an unusual preface can evade it.
  • requiresHumanReview is always true. Answer each task's review questions with evidence from the response. Do not combine these into a literary-quality score.

The mechanical-pass-dramatic-failure example retains the anchor and word ceiling but reverses Sol's career decision. It is deliberately unacceptable for its task. Use it to calibrate reviewers who might otherwise mistake a mechanical pass for instruction-following or dramatic success. The over-budget and prefaced examples isolate different failures; the empty response passes the ceiling but is no scene.

Keep each task and its calibration examples with the same split_group as its source scene. These are development materials with public counterparts. Use independently authored unseen scenes for comparative evaluation claims.

Original AI-assisted Inkwell calibration material, November 7, 2026; CC0-1.0.