jama.2024.19585

Imported from AltangadasTbl-source/ai-paper-val (2024/jama.2024.19585/AGENTS.md). Install upstream with npx skills add AltangadasTbl-source/ai-paper-val --skill jama.2024.19585. Copyright stays with the author.

Quantitative Quality-Control Review 1.5.3 — Reuse Existing Evidence Assets

The current project root is exactly one paper package. The package may contain earlier .ai_paper_validation/ records whose discovery stopped at, prioritized, or reported only 10 candidates. Workflow 1.5.3 starts candidate discovery again without a count limit, while reusing all usable existing OCR, native text, layout text, table extraction, workbook extraction, rendered pages, and document maps. Reuse is an optimization, not a coverage boundary: every direct-source unit that lacks usable reusable extraction must be freshly inspected and mapped from the supplied source.

Use English only in prompts, generated artifacts, logs, and reports. Preserve every supplied source and every pre-existing audit artifact. Write new work only to:

• .ai_paper_validation/review_1_5_3/
• .ai_paper_validation/final_report_1_5_3.md
• .ai_paper_validation/final_report_1_5_3.html

Agent-first execution

Start immediately and operate autonomously. The coordinator plans coverage, delegates bounded and disjoint evidence scopes to specialist agents, merges their durable artifacts, and verifies coverage. Python is auxiliary only. Do not use Python to orchestrate the scientific review, select candidates, perform ordinary PDF extraction, run OCR, or replace agent reasoning. Prefer direct local tools: sha256sum, file, pdfinfo, pdftotext, pdftoppm/pdftocairo, tesseract, and libreoffice/soffice. The only permitted Python helpers are the optional Office structure extractor, deterministic token-cost calculator, report renderer, and final validator in workflow_1_5_3/scripts/. The calculator summarizes authoritative runtime usage; it does not estimate tokens from review text.

Use artifact-first handoffs. Every specialist writes complete English output to one unique assigned path and returns only a compact status, exact scope, counts, limitations, and artifact path. Never let two agents write the same artifact concurrently. Shard by disjoint pages, tables, relationships, or candidate IDs when needed; shard size is a context bound, never a finding limit.

Do not put provider credentials or provider selection in the package-root .codex/config.toml. The direct interactive session requires the nine bundled .codex/agents/ role presets and verifies each one before scientific work. Preserve unrelated package .codex controls when copying those files. Validation does not rewrite .codex/.

Fixed model routing contract

Start from the package root with codex --approve-for-me, then send Read START_PROMPT.md completely and execute Workflow 1.5.3 now. as the first request. The user-level Codex configuration supplies the provider and authentication; the project config fixes the coordinator at gpt-5.6-sol/high and the nine named presets fix every specialist model/effort pair. Do not use ~openai/gpt-latest, another moving alias, a shell launcher, or codex exec fallback.

Before scientific work, create routing_preflight.md and require it to report PASS, coordinator gpt-5.6-sol/high, ordinary specialists gpt-5.6-terra/medium, statistical specialists gpt-5.6-terra/high, Sol specialists gpt-5.6-sol/high, Coordinator inference: PASS, execution mode INTERACTIVE_CLI, and all nine named agent presets verified against their required pairs.

Every mandatory specialist stage uses a fresh agent and a distinct runtime ID. Use the named custom agent preset when available. Otherwise use a fresh default agent, pass the exact fixed model and reasoning effort for that role, and include the role contract in the spawn prompt. Never reuse an agent for another mandatory stage, promote an agent with follow-up, omit a fallback model override, or substitute Sol for a Terra role. Record the exact model ID in the execution manifest and token ledger.

Read before reviewing

Read completely:

• QUALITY_CONTROL_SCOPE.md
• PERFORMANCE_PROFILE.md
• workflow_1_5_3/review_contract.md
• workflow_1_5_3/report_spec.md
• workflow_1_5_3/settings.toml

Do not read an old candidate set, review queue, verifier disposition, critic decision, endetail section, or final report as a scientific input or candidate source. Preserve those records unchanged, but rebuild discovery from source-linked reusable assets plus mandatory direct-source inspection of every uncovered unit.

Required workflow

1. Verify routing_preflight.md first. Record the UTC start time, then inventory every direct source, its complete stable unit count, and SHA-256 in .ai_paper_validation/review_1_5_3/source_hashes_before.sha256. Set a package-specific target and bounded Target basis from PERFORMANCE_PROFILE.md; do not infer either from SAP length or page count alone. Initialize every required field in run_state.md. Initialize agent_execution_manifest.md with the current coordinator and add every spawned agent exactly once as the run proceeds; the manifest is not limited to statistical agents.
2. Use a fresh qc15_reuse_asset_curator agent to inventory and hash every reused OCR/text/table/page/document-map artifact in reused_artifact_hashes_before.sha256, assess its coverage and fitness, and write the source and evidence-asset inventories. In source_coverage.md, record one row per direct source with total, reusable, fresh-required, and mapped unit counts. Do not modify reused artifacts.
3. Compare reusable coverage against every direct-source unit. Assign every uncovered page, sheet, record group, paragraph group, and table group to fresh direct-source extraction and mapping. The reusable and fresh-required unit counts must partition the total, and mapped units must equal the total before completion. A reusable-derivative gap may remain a limitation; a scientific-coverage gap may not.
4. Create coverage_manifest.md before scientific mapping. Record every disjoint source/evidence unit and stage assignment. Each table row contains exactly one plain relative artifact path; put every shard part on its own row.
5. Run fresh, distinct qc15_main_quantitative_mapper and qc15_support_quantitative_mapper agents in parallel over the disjoint union of reusable-backed and fresh-required units. Reconstruct every result-relevant numeric and statistical relationship.
6. Run numeric and cross-source review in two distinct fresh gpt-5.6-terra/medium agents. Independently spawn a new gpt-5.6-terra/high agent for statistical pass 1; do not reuse a mapper agent or use follow-up to request a reasoning-effort change. Record the runtime agent ID, model, effort, start mode, and one output artifact in agent_execution_manifest.md. After each completed model response, retain its authoritative runtime usage record for later token accounting when the runtime exposes it.
7. Merge only genuine duplicates before stable IDs. Assign C001, C002, ... to every distinct candidate and write candidate_ledger.md. Preserve all checker provenance and source locations.
8. Run a fresh qc15_evidence_rechecker/high agent over every stable ID against the exact source page or truthful Office location. Existing OCR/text is a locator and transcription aid, not final authority.
9. Spawn a different new gpt-5.6-terra/high agent for statistical pass 2 over every statistical relationship, the complete cross-lane ledger, and all recheck facts. Record its distinct runtime ID in agent_execution_manifest.md. Append genuinely new candidates without renumbering, then recheck every appended ID.
10. Run a fresh qc15_quality_control_auditor/high agent over every stable ID, every coverage row, the complete source_coverage.md, and the statistical execution manifest. Repair supportable omissions, but never delete or suppress an assigned candidate ID.
11. Run a fresh qc15_report_generator/high agent to assemble the complete Markdown report. It must contain every candidate in the ledger. If no candidate is found after complete coverage, produce a zero-candidate report that states the completed scope and limitations. For a large ledger, generate disjoint candidate-card parts in parallel waves and merge all parts; never shorten the report to fit one agent call.
12. Immediately after Markdown assembly, finalize run_state.md. Treat Finished UTC as the token- accounting cutoff. Write token_usage_ledger.csv for the coordinator and every manifested agent, using exact response-level runtime/API usage only. Use TOTALS_ONLY when exact input/output/total counts exist without billing details; use UNAVAILABLE plus exact __ token fields when no count is exposed. Never approximate from text length. Run the bundled cost calculator, then copy the exact performance fields, token-accounting status, per-model totals, package total, and complete token-only price when available into the report using the bundled dated Sol/Terra pricing snapshot. Recompute source and reused-artifact hashes, render HTML once, and validate:

python3 workflow_1_5_3/scripts/calculate_token_cost.py \
  --ledger .ai_paper_validation/review_1_5_3/token_usage_ledger.csv \
  --pricing workflow_1_5_3/token_pricing.toml \
  --markdown .ai_paper_validation/review_1_5_3/token_usage_summary.md \
  --json .ai_paper_validation/review_1_5_3/token_usage_summary.json
python3 workflow_1_5_3/scripts/render_report.py \
  .ai_paper_validation/final_report_1_5_3.md \
  .ai_paper_validation/final_report_1_5_3.html \
  --profile 1.5.3
python3 workflow_1_5_3/scripts/validate_review.py --profile 1.5.3 --package .

Repair correctable defects until .ai_paper_validation/review_1_5_3/review_validation.json reports PASS. If elapsed time exceeds the selected target's upper bound, record bounded causes and complete the work; do not omit coverage or candidates to meet the target. If a repair requires another model call after Finished UTC, reopen the accounting window, update the finish time and manifest/ledger, and rerun the token calculation before rendering and validation.

Fixed scientific and communication boundaries

• Prioritize numeric, denominator/proportion/total, inferential-statistical, cross-document numeric, effect-measure/label/scale, and rate-versus-count consistency.
• Analysis-unit, randomization-level, sample-unit, or population-definition issues are secondary and qualify only when they create a concrete inconsistency in a reported number, statistic, denominator, label, or interpretation.
• Do not perform a broad methodology, study-design, clinical, novelty, misconduct, or raw-data audit.
• Use only supplied package evidence. Do not browse the web or use external literature.
• Never register a candidate solely because a very small P value is displayed as P = 0, p = 0.000, or equivalent. Treat coherent display zero as finite-precision shorthand and record DISPLAY_ZERO_NOT_CANDIDATE. Require an independent supplied-source contradiction before assigning a C ID, and frame the candidate around that contradiction.
• Treat findings as quality-control candidates: small preventable reporting defects may matter because they can propagate into systematic reviews, meta-analyses, or other downstream evidence products. Do not inflate that possibility into a claim that the paper's conclusion is wrong.
• Never assign Major, Minor, Verified, Rejected, Uncertain, severity, validity, acceptance, exclusion, or a final correction. Every candidate remains Pending Human Adjudication.
• Preserve 1.2-style truthful relative links: PDF evidence ends in #page=N; workbook evidence names worksheet and exact cells; CSV evidence names row and column; DOC/DOCX evidence names a stable paragraph/table location or a locally derived PDF page.
• The Linux target is CPU-only. Reuse existing OCR first. Use direct native/layout extraction for uncovered units and targeted Tesseract only when text is unusable or visual confirmation requires it; never probe or invoke a GPU.

Completion

Do not finish until every direct-source row has mapped units equal to total units; all coverage rows are complete and contain one artifact path; routing preflight is PASS; every mandatory specialist stage has a distinct fresh runtime agent with the required model and effort; both fresh Terra/high statistical agents cover every registered relationship; every actual agent including the coordinator appears in both the execution manifest and token ledger; agent- and model-level token summaries are current; ledger/recheck/quality/report ID sets are identical; every local evidence link resolves; source and reused-artifact hashes are unchanged; standalone HTML exists with embedded CSS and a table of contents; and the validator reports PASS.