Fold HIGH alexwestco/llm-to-jev: conversion assistant on-ramp

[24601]

PRIMARY user-linked HIGH fold of alexwestco/llm-to-jev (2026-09-20 ~09:40 Boise). Rebased onto latest main after merged #37 (501ba28, SemIf §117 / 330–336 / #100), merged #38 (be9fe1c, jcr §116 / 309–316 / #99), and merged #36 (7de41ef, NanoJev §115 / 303–308 / #98). Adversarial review + testing gate on this revision: PASS. Do not merge from this agent.

Hard rules

• Do not reopen or amend PR Fold hourly 0042 HIGH: jev-readout, jevify, laya-jev-lab #23–Fold HIGH NiazMorshed2007/jcr: capability-tree lookup, does not execute #38.
• Do not push onto open Post-#35 hygiene: §114 citation, scannable dumps, hop-ECE interleave #39 (§114 hygiene — do not reopen) or Fold hourly 0947 HIGH: jev-as-judge, OneForward candidate_mass, catalogs and RLCD heads #41 (hourly 0947; stale §116 / 309–324 / #99 — Fold hourly 0947 HIGH: jev-as-judge, OneForward candidate_mass, catalogs and RLCD heads #41 remaps, this fold does not).
• IDs kept: this fold is notes.md §118 / composition 322–329 / findings batch #101.
  • merged Fold hourly 0743 HIGH: ywchiu Harbor-jevals, Verdict-open-jev, DecisionOps #31 owns §113 / 273–288 / #96
  • merged Fold hourly 0843 HIGH: hysteresis, instruct-tuning ECE, calibration does not compose #35 owns §114 / 289–302 / #97
  • merged Fold TianyuCodings/NanoJev unified-games-v1 densify #36 owns §115 / 303–308 / #98
  • merged Fold HIGH NiazMorshed2007/jcr: capability-tree lookup, does not execute #38 owns §116 / 309–316 / #99
  • merged Densify SemIf rename + MLX + accuracy ladder #37 owns §117 / 330–336 / #100 (they reserved §118 / 317–329 for this fold)
  • open Post-#35 hygiene: §114 citation, scannable dumps, hop-ECE interleave #39 is hygiene on §114 (unchanged IDs; no collision)
  • open Fold hourly 0947 HIGH: jev-as-judge, OneForward candidate_mass, catalogs and RLCD heads #41 must remap off merged §116 / 309–316 / #99 and off this fold’s 322–324
• Docs/skill only. No wrappers, npm / git clone / TYPESAFE_API_KEY recipes. Skip Archer. invented_signal: false.

What landed

research/notes.md §118, composition items 322–329, findings batch #101, SKILL YAML + protocol triggers, mental-models Apply 0940, FAQ, mixed-architecture fail table, question-design diagnosis, applied-mappings, judgment-class, validation, uniqueness lock as one consecutive substring across 21 overlays, CHANGELOG Unreleased (0940 lock dump in research/changelog-hourly.md). uniqueness_gate.py checks 0843 + 0915 + jcr + 0922 + 0940, headings §114/§115/§116/§117/§118, items 289–316 + 322–329 + 330–336, batches #97–#101, and SKILL protocol 0940 triggers (so rebase cannot glue them onto the jcr/0922 lock). Does not bump 0.4.0. README stays scannable.

Keep-both on rebase onto 501ba28: NanoJev §115 / 303–308, jcr §116 / 309–316, SemIf §117 / 330–336, then this fold §118 / 322–329. Consecutive uniqueness walls remain five separate lines (0843, 0915, jcr, 0922, 0940).

This is the migration / question-design on-ramp: finds where typed judgment belongs inside existing LLM prompts. Companion to the decision-design card. A falsifying experiment is still required. Conversion assistant ≠ equivalence guarantee. Deterministic heuristics ≠ LLM. Soft proposal ≠ production gate. Heuristic conversion ≠ calibrated Noul.

Claim (quote theirs)

• Turn decision-shaped LLM prompts into proposed Jev primitives (Choice, Score, Noul).
• This is a conversion assistant, not an automatic guarantee of equivalent behavior.
• The compiler uses deterministic heuristics, not an LLM or evaluation model. It understands a deliberately small set of common prompt patterns.
• Generated instructions and criteria must be reviewed before production use.
• Suitability theirs: strong / partial / not_a_fit. Compatibility theirs: full / partial / none.
• Prompts requiring open-ended prose are not a fit. Writing new text stays with an LLM.
• Score ranges such as 0 to 1 are translated into ordered Jev criteria. Review the generated Score rubric; Jev scores ordered criteria, not an arbitrary 0-to-1 range.
• Everything runs locally in the browser. No framework, database, account, API, or server-side prompt processing. connect-src 'none'. The key is read from the process environment and is never stored or printed.
• Roadmap theirs (not landed): fixture eval set, editable criteria, multi-step decomposition, compare LLM vs Jev latency/cost.

Namesake: alexwestco/llm-to-jev ≠ altryne/jevify ≠ ryana/jevify ≠ fidecastro/jevify ≠ Mintzs/jevify ≠ gulagala001/jevify ≠ uspraveen/Jevify.

Live REST (review relock this pass — unchanged, no SHA rewrite)

• GitHub alexwestco/llm-to-jev: JavaScript MIT; GitHub id 1377308198; 2★; forks 0; open issues 0; size 29; default main; homepage field null; topics empty.
• HEAD 234058ab372d7754833c6279601755f8fda55d98 (“fix deployment issues on heroku”, 2026-09-20T12:05:59Z).
• README SHA 43cd94fba7527d78635b148628485c0cd1b1ad66; LICENSE SHA 5f3340068a565a3417469985fdbb93631611336b; src/compiler.js SHA fdf235d01247e799fe35003e6c46192124c236f2.
• Pages https://alexwestco.github.io/llm-to-jev/ HTTP 200 (last-modified 2026-09-20 12:06:26 GMT). CSP connect-src 'none' — the demo never calls Jev.
• Star counts ephemeral.

Testing (checklist 1–8) — PASS

python3 .agents/skills/augustus/scripts/evaluate_decisions.py --self-test
python3 .agents/skills/augustus/scripts/uniqueness_gate.py

1. uniqueness-gate ok (0843 chars=2415, 0915 chars=1132, jcr chars=1499, lock0922 chars=1292, 0940 chars=1350, overlays=21).
2. evaluate_decisions --self-test ok.
3. SKILL.yaml parses (name: augustus, version 0.4.0). Protocol line routes 0940 triggers (not glued onto the jcr/0922 lock). YAML 0940 fragments stay inside the quoted description.
4. sources.json valid; last object is kind: github alexwestco/llm-to-jev after SemIf cousins; retrieved 2026-09-20T16:10Z.
5. Live REST PASS (2★ / HEAD 234058ab / Pages 200 / connect-src 'none').
6. ID uniqueness PASS vs merged §113–§117 and open Post-#35 hygiene: §114 citation, scannable dumps, hop-ECE interleave #39/Fold hourly 0947 HIGH: jev-as-judge, OneForward candidate_mass, catalogs and RLCD heads #41 (this fold stays §118 / 322–329 / #101).
7. Honesty locks held: conversion assistant ≠ Harbor; heuristic ≠ calibrated Noul; 0–1 → ordered criteria theirs; generative work stays LLM; soft proposal ≠ gate; mental-model on-ramp beside decision-design card; falsifying experiment still required.
8. No wrappers / npm / TYPESAFE_API_KEY recipes. Does not bump 0.4.0.

CI: Check Pages site / build success on 3c4d222 (run 35522648491). mergeable_state: clean. Draft ready. Head 3c4d222 on 501ba28. Not merged from this agent.