Skip to main content
Paper-synthesis is the resolve skill at the paper level: it drives one committed paper from its accumulated hypothesis evidence, through a defensible position against prior work, to a publish / no-go decision and an assembled manuscript. It is the exact mirror, one level up, of hypothesis-testing (../hypothesis-testing/SKILL.md): where that skill takes a hypothesis to a findings verdict, this one takes a paper to a decision verdict, and the same firewall holds — the skill drafts, keeps the accounts, and advises; the author authors and decides (see ../../docs/design/00-meta-spec.md §2.1). You cannot “run” this skill to produce a paper; you drive it.

​
When to use

  • After a paper has been promoted from portfolio-backlog.md (by paper-exploration, ../paper-exploration/SKILL.md) and registered in docs/research/papers.md — it is a committed deliverable, not a candidate.
  • When its constituent hypotheses have accumulated enough evidence (findings.md verdicts, backed by run-refs) that a publish decision is in reach, or to develop the manuscript incrementally as they resolve.
  • Not for generating paper ideas — that is paper-exploration (generate proposes; resolve disposes). Do not use this skill to adjudicate a hypothesis; that is hypothesis-testing one level down.

​
Staged documents

All staged docs live under docs/research/<paper>/paper/ in the consumer repo (../../docs/design/00-meta-spec.md §5). Each carries a status frontmatter block feeding the progress roll-up (../progress/SKILL.md). The pipeline, in order: The stages are a flow, not a gate sequence you can automate through: each is a document the author fills, drafted as a proposal by the skill. Positioning may be revisited as hypotheses resolve; the decision is gated on the whole.

​
How it works

  1. Pitch. Carry the promoted backlog row into pitch.md: the central claim, the intended contribution, the target venue and its bar. Confirm which hypotheses (docs/research/<paper>/hypotheses/*) are load-bearing for this paper.
  2. Positioning. Invoke literature position --level paper (../literature/SKILL.md, ../../docs/design/02-literature.md §3) to produce positioning.md. At paper level this is the full treatment, not the hypothesis-level rapid review: a taxonomy of the field, a concept-centric matrix (rows = methods, cols = the attributes the paper’s delta turns on), a PRISMA-style include/exclude log (the anti-cherry-picking audit trail, sourced from the triage sidecar), a per-branch delta paragraph, and a derived baseline list (one strong tuned representative per branch + current SOTA + the most-likely-cited-against
    • a simplest floor). The closest-prior-work paragraph and the isolating ablation guard against overclaimed novelty.
  3. Outline / plan — delegated. Structuring the manuscript is engineering, so hand it to the engineering backend bound in .honest-scholar/config.yml (engineering_backend:) — its design → plan capabilities — exactly as hypothesis-testing delegates design.md/plan.md. Store the resulting outline.md / plan.md under docs/research/<paper>/paper/. This skill owns the scientific framing; it does not reimplement engineering planning.
  4. Decision. When the evidence and positioning are in, draft decision.md as a publish / no-go proposal: does the accumulated hypothesis evidence support the contribution the pitch claims, and does the positioning show a defensible delta? This is a material decision — see Guardrails. The skill marshals the evidence and recommends; the author decides and signs.
  5. Sections. Assemble sections/ from the claim→evidence ledger (below). Draft prose as proposals; every reported number is written by the backend tables capability, never typed by hand.
  6. Disclose (finalize, publish only). Once the decision is publish and the sections are assembled, proactively propose an AI-use disclosure + a citation for the manuscript’s Use of AI / Acknowledgments — drafted from the provenance record (who signed off which decisions, which run-refs back which results, what the skills drafted), so it is truthful, not boilerplate. This is opt-in and author-owned: the author edits, adopts, or declines; never auto-insert. Surfacing it here — automatically, at the moment it matters — is deliberate (see ../../DISCLOSURE.md, ADR-0025). Keep it humble: it discloses what was done and links the record; it does not certify honesty. On a no-go decision there is nothing to disclose — the paper is not submitted — so this step is skipped.

​
The claim→evidence ledger

docs/research/<paper>/paper/ledger.md is the backbone of the manuscript: each row is one paper-level claim recorded as a Toulmin sextet, so that every assertion the paper makes is traceable to the evidence and its limits are explicit. The six fields: Rules for the ledger:
  • Grounds cite run-refs, not numbers. A run-ref is the citable unit of evidence from the experiment-backend contract (../../docs/design/04-substrate-and-contract.md §3). The backend tables capability is the only writer of result numbers into sections/ — it renders managed, regenerable result blocks from the run-refs a claim cites. Hand-copied numbers are forbidden: they cannot be re-verified and break when evidence is re-run.
  • Every load-bearing claim must resolve to grounds that resolve to exact bytes via the provenance stamp (config hash, code/symbol provenance, dataset id+version+sha256). A claim with no run-ref backing is a claim with no evidence.
  • Staleness is honest. Before the decision and before assembly, check is-current on the cited run-refs; a stale run-ref surfaces the gap but the human decides whether to re-run (the backend never decides for you).
  • Backing draws on the literature registry so the defend cited-work target can verify that a cited source actually supports the sentence it backs.
A worked row (domain-neutral shape):
claim Method M improves the primary metric over the strongest baseline. grounds run-refs rr-a91f (M) and rr-7c02 (baseline), matched split/seed. warrant paired improvement under matched conditions isolates the method as the cause. backing the matched-tuning protocol from positioning.md’s baseline list; comparison discipline per [Webster & Watson]. qualifier holds on the two in-distribution benchmarks tested, not out-of-distribution. rebuttal fails if the baseline was under-tuned — hence the isolating ablation and the shared tuning budget recorded in the provenance stamp.
Every number in the assembled section for this claim is rendered by tables from rr-a91f / rr-7c02; the author never types them.

​
Composition

  • Feeds from: paper-exploration (the promoted pitch) and the paper’s resolved hypotheses (hypothesis-testing findings.md verdicts + their run-refs).
  • Calls: literature position --level paper for positioning.md; the engineering backend bound in .honest-scholar/config.yml (engineering_backend:) for outline.md/plan.md; and the experiment backend bound in docs/research/papers.md (backend:) — tables / evidence / is-current — for result blocks and staleness.
  • Examined by: defend (../defend/SKILL.md), whose paper-synthesis preset targets positioning (novelty vs. prior work) and cited-work (do the cited sources support the claims), plus the claim target over the ledger. The guardrail fires automatically before the decision.md sign-off.
  • Reported by: progress (../progress/SKILL.md), which reads the decision and section frontmatter; a paper is done when its hypotheses are resolved and it is submission-ready. A no-go decision reads as done, not failed — verdict and readiness are distinct axes.
  • Feeds up to: the thesis level, where a paper becomes a chapter mapped to the aims (../thesis/SKILL.md).

​
Guardrails

Hard rules — the load-bearing constraints, not preferences.
  • Drafting is assistive; the author authors. The skill drafts pitch, positioning, ledger rows, and section prose as proposals. The scientific claims and their wording are the author’s. Anywhere a scientific judgement is required, stop and ask rather than deciding (../../docs/design/00-meta-spec.md §2.1). You cannot produce a paper by “running” this skill.
  • The publish decision is material and human-signed. decision.md names its human decision-maker and date; it is gated on accumulated hypothesis evidence + positioning, the paper-level mirror of a findings verdict (../../docs/design/01-lifecycle.md §8). The defend guardrail fires before the sign-off (positioning + cited-work + claim): it stops, surfaces gaps, offers to examine/teach, and records. The human may override; the override is logged — a stop-and-confirm, not a hard block. The AI never adjudicates publish-worthiness.
  • Numbers come from the backend, always. Only tables writes result blocks into sections/; the ledger cites run-refs. If a number cannot be traced to a run-ref, it does not go in the paper.
  • Understanding gates the claims. The author must be able to defend the positioning’s novelty argument and each ledger claim to a reviewer’s standard; unmet, defend surfaces it (../../docs/design/00-meta-spec.md §2.2).
  • Firewall. Explore proposes, resolve disposes, synthesize reports. This skill develops and assembles a committed paper; it does not generate paper ideas and does not adjudicate its own decision.
  • Follow-ups become issues, not TODOs — a deferred check or a known gap in the record is captured as a self-contained GitHub issue, not left in a doc’s margin.

​
Commit attribution

When you commit artifacts produced by this skill, add these git trailers — discovery + provenance (see ../../resources/commit-attribution.md):