engineersoheinsahani-ai/skill

SKILL.md — journal-arena ENB-Pro v4.0 (exclusive pre-submission edition)

Purpose: full scientometrics review of a manuscript before submission to Energy and Buildings (coding method), executed by real agents on local models — the original 16-group division, <=100 executors, every call separate and stamped. All English. All v3.1 capabilities retained (dual grounding, ledger, corpus, matrix, council, kit, guard) plus the machine wave over input/ (map, precheck, figureqc, diffnum) and the post-council writers team with figure specialists.

0. The laws (every agent, every model, every phase)

1. One-duty law — each agent owns EXACTLY one part (its G-group card in groups/16-group-playbooks.md); out-of-duty observations go to at most one forwarded note (severity style).
2. Dual-grounding law — every evaluative judgment carries TWO anchors: (a) an exact quote from the STUDIED LITERATURE with source id + locator ([L:e4-012 ...] from work/ledger.jsonl), and (b) an exact manuscript quote, only where a claim is being mapped. No (a) => quarantined, never counted. Exemption: deterministic rule-compliance findings (anchor_type: rulebook|machine, lit_quote_id: L:rule-<id> where the id exists in checklists/ or an engine enforcer) are FACTS — machine-verified by matrix3.py rule_exists.
3. Decision-value law — verdict + at most 3 actionable moves + the quotes. No paraphrase-only summaries. The ledger deduplicates; activity without a decision is waste.
4. Evidence-tier law — body=1.0 > abstract=0.5 > rulebook=1.0 (deterministic)

   metadata=0 (quarantine). Every quote carries its tier.

5. Verification license — read-only toward sources and input/; the manuscript never leaves the owner's machine except the owner's explicit --attach decision (POL-AI-5 for online models; local models exempt).
6. PRIV-1…PRIV-6 — engine/humanize.py is a PRIVATE derived module: never publish it; guard_private.py scan must print clean before ANY package leaves this machine (public builds exclude it — see NOTICE.md).
7. No-fabrication law — never invent numbers; missing value = [number needed], resolved only by the author. No detector score decides.
8. No-exaggeration law — no claim bigger than its evidence; claim inflation is scored and vetoable (G11 meta-audit + arena rubric weight).
9. Execution-honesty law — every artifact records WHO executed it: indep: machine (deterministic tool — a fact) · harness (separate model calls via engine/harness.py, key from owner env only) · human (owner or paste-back) · single (one model wearing hats — a labeled simulation). "100 agents" is only true with 100 separate calls (campaign.py) — otherwise it is single. Single-executor findings NEVER confirm a P0 alone: P0 needs machine confirmation, >=2 independent executors, or the owner's sign-off.
10. Artifact-diff law (NEW, owner decision 2026-10-12) — agents NEVER execute the research pipeline (time-consuming; owner's job). The owner deposits the pipeline's own outputs in input/results/ (manifests, audit JSONs, tables, CSVs, logs) and the code in input/pipeline/ (read-only static review). engine/diffnum.py screens every manuscript number against the artifacts (traced / UNTRACED); untraced numbers become machine findings for G8 and [number needed] candidates. An UNTRACED flag is a screen hit, not proof of error — honest by construction.

1. Decisions on record

Carried from v3.1 (owner, 2026-10-09/11): code_scope=research_code · tier1_journals=[Energy and Buildings (primary), Applied Energy, Building & Environment, Energy, Journal of Building Engineering, Energy and AI] · corpus_axis=manuscript-derived (work/corpus/axis.json) · corpus_counting=body-access-only · corpus_ceiling=20 + sufficiency (study CLOSED at E6 128 / E2 82 / E4 76 countable; anchors deposited in work/ledger.jsonl) · humanizer_mode=distilled-brief. New (owner, 2026-10-12): ENB-Pro exclusive edition — all English · agent division EXACTLY the original v2 topology (16 duty groups, waves of 10, <=100 real executors) · execution on LOCAL models (current: qwen3.5:4b; code optimized for 4B-class) · pipeline never executed by agents (law 10) · input/ drop-zone contract · writers team includes FIG-MPL + FIG-GV figure specialists · figures must meet the most advanced ENB craft (templates shipped).

2. Topology — the original 16-group division (single source: groups/16-group-playbooks.md)

G1 title · G2 abstract · G3 keywords-highlights-GA · G4 introduction-gap · G5 literature-review · G6 methodology-coding · G7 data · G8 results-consistency · G9 statistics-UQ · G10 discussion-claims · G11 recommendations+meta-audit (veto) · G12 references-integrity · G13 ethics-policy · G14 formatting-kit · G15 language-style · G16 reviewer-simulation. Division math (campaign.py): N=100 => G1–G4 x7 + G5–G16 x6 = exactly 100 slots (v2 used ceil => 112; Pro honors the owner's standing <=100 law), waves of 10, agent ids G6-a03. Post-council writers team (6 fixed roles): WR-TEXT-1/2/3, FIG-MPL, FIG-GV, STAT-CHECK.

3. Pipeline (waves, resumable)

W-machine   run.py prep --work work --input input        (map, precheck, figureqc,
                                                          diffnum, inventory — no key)
W-division  run.py campaign run --work work [--agents 100] (real executors, law 9;
            resumable; checkpoint after every group: check + matrix + council)
  or AGENT MODE (no key; GitHub clone-&-run per AGENTS.md): campaign briefs ->
            the cloning agent answers work/answers/<slot>.txt -> campaign
            ingest-dir --indep single (honest stamp: one model, 100 hats)
W-aggregate run.py check/matrix/council work/findings.jsonl ...
W-writers   run.py campaign run --stage writers --work work (after council verdicts)
W-apply     task/05 + apply_revisions.py (owner approves; traceability: copies only)
W-kit       kit.py strip --src input/manuscript.txt --out <submission-copy> + figureqc clean
W-release   LICENSE / Zenodo DOI / data deposit (Option C) / GenAI declaration

4. Anchors & pools

Ledger rows keep their v3.1 E-gids; campaign maps duty groups to pools: G2,G7,G8 -> E6 (integrity 319 anchors) · G4,G5 -> E2+E4 (novelty/methodology 412) · G6,G9 -> E4(+E5 repo rows) · all other groups: rulebook anchors from checklists/ (L:rule-*). Briefs embed <=20 anchors (120-char trim) — the ONLY quotable pool (law 2).

5. Small-model contract (4B-class local models)

Prompts: laws digest ~700 chars + duty card + <=20 anchors + duty-relevant manuscript excerpt ONLY (<=6000 chars, section-mapped) + machine facts + one few-shot JSONL example. Generation: max_tokens 1200, temperature profile 0.2–0.5, /no_think for qwen3-family (default), one parse-retry, robust extraction (fences/trailing commas/concatenated objects). Expect aggressive quarantine by the gates — that is the system working, not failing; realistic yield on 4B: 2–4 surviving findings per executor.

6. Layout & commands

checklists/ engine/ groups/ input/ sources/ task/ templates/ demo/ work/ Entry: python3 engine/run.py prep|campaign|diffnum|figureqc|harness|map| precheck|matrix|council|check|validate|ledger|corpus|kit|apply|guard|stage — RUNNER.md is the connection guide (Ollama/API/paste-back). NOTICE.md: privacy, license status, environment variables.