argument-governance

/argument-governance - Argument System Governance

Purpose

Make the paper's argument inspectable as a system:

intent -> gap -> contribution
data -> result -> claim -> contribution
prior-work evidence -> innovation -> contribution
claim/contribution -> boundary -> reviewer risk

Use this before major manuscript revision, before /peer-review, before /self-review, or after /manuscript-reframe when the draft needs stronger gap-contribution and claim-evidence control.

Codex-Only Baseline

Complete this workflow with Codex, local files, and the bundled checker. External reviewers, subagents, Gemini, or other model calls are optional advisory inputs only; do not require them and do not stop because they are unavailable.

Enhanced Advisory Mode

When the user explicitly enables an API-key-backed reviewer, Codex may request an external advisory pass after the local argument packet exists. Use this only as a second opinion:

• keep API keys in environment variables, never in CSV, Markdown, logs, or manifests
• send only the source subset the user or manifest permits
• record the provider, model, source subset, prompt purpose, and output path
• place results in advisory notes or reviewer-risk fields
• never treat external output as evidence, source support, or a final gate

Core Rules

1. Every contribution must answer a named gap.
2. Every main claim must belong to a contribution or the paper-level thesis.
3. Every lower-level claim must support a parent claim.
4. Every claim must have the right type and amount of evidence.
5. Background, methodological, workflow, or governance evidence must not be promoted into empirical, outcome, causal, or deployment validation.
6. Important claims must not be evidence-thin; peripheral claims must not be citation-padded.
7. Reviewer risks must target a specific intent, gap, contribution, claim, evidence cluster, or limitation.
8. Do not fill missing argument links from memory, prior chat, or unstated assumptions. Mark them as gaps.
9. Treat external model reviews as advisory notes, not evidence or required gates.
10. Keep data and results separate: data are analysis inputs; results are derived outputs with a method, scope, and provenance.
11. Keep contribution and innovation separate: contribution states what the project delivers; innovation states how it differs from an explicit comparison set.
12. Judge over- or under-balance by missing, duplicated, orphaned, conflicting, or scope-mismatched functions. Do not infer quality from raw counts, word count, citation count, dataset count, or one composite score.
13. Treat contribution-focus changes as advisory candidates. Require an explicit author decision before changing the title, abstract, research question, contribution order, or manuscript spine.
14. Keep the field-level gap, paper-level contribution, data-level finding, and extrapolative claim separate. A true field gap does not prove that this paper closes it; require evidence for every upward inference.
15. Treat boundary as a paper-wide evidence licence across intent, population or denominator, estimand, evidence scope, title, abstract, contribution list, result headings, and conclusion, not as a limitation-only disclaimer.
16. Do not treat tables, ledgers, guardrails, qualifiers, or appendices as substitutes for missing construct validity, reliability, sampling, baseline, power, or transfer evidence. Record evidence_needed or narrow the claim.
17. Keep the real-world problem, intended use, present method task, scientific question, empirical finding, and evidence boundary separate. Missing deployment or outcome validation narrows the licensed claim; it does not erase the motivating application.
18. Admit experiments and analyses by argumentative function, not by completion. Record whether each item directly answers the core question, whether the main conclusion survives its removal, its role, destination, and author decision before changing its prominence.