---
name: arguing-research-papers
description: >-
  Argue ONE FINISHED written research claim to an absent reviewer, or critically appraise
  ONE paper's argument, method, and validity. Owns claim calibration, contribution/novelty
  positioning, warrants, qualifiers, anti-fabrication, reviewer red-team,
  and rebuttal. Use for paper argument / contribution / “not novel” / overclaim / reviewer 2 /
  peer-review critique / 論文の主張 / 貢献 / 新規性 / 査読 / リバッタル. Cuts:
  fact extract → raising-resolution; neutral one-paper summary → direct answer with its citation gate
  silent; corpus state → systematizing-knowledge; premise-only excavation →
  surfacing-blind-spots; section order/MECE → structuring-documents; wording → linting-prose; live
  talk → designing-presentations; program reopen → directing-research; thesis genesis →
  forging-novel-theses; expensive/irreversible one-tree → acting-on-hypotheses; cheap reversible probe
  → domain/plain executor; agent topology → orchestrating-agents. Claim/calibration stay SOLO.
  English skill; answer in the user's language.
---

# Arguing research papers — one claim, calibrated to evidence, hardened for review

> **Version**: v2607.1.0 (2026-07-08)
> **Scope**: the RHETORIC and EPISTEMICS of a written research paper — deciding WHAT to claim,
> calibrating claim-strength to the evidence, positioning against the literature, and surviving an
> absent, skeptical, skimming peer reviewer + the permanent record. Sits ABOVE document structure
> (`structuring-documents`) and sentence polish (`linting-prose`), BESIDE the live talk
> (`designing-presentations`).
> **Lineage**: distilled from a 15-agent adversarially-reconciled SoK survey (2026-07) of the
> paper-writing canon (Gopen-Swan through the Japanese 戸田山・木下・野矢・酒井 tradition). Full graded
> provenance: `references/sources.md`; the survey fleet: `tests/forge-verification-ledger.md`.
> The same argument-level machinery may face inward (authoring a manuscript) or outward (critically
> appraising one paper as a reviewer). Neutral summary and bounded factual extraction are not
> argument appraisal.
> **Build order (ATOMIC — ship in ONE commit; no pointer may dangle).** Verify from the skill dir:
> `for f in frameworks reconciliation calibration genre-playbooks reviewer-defense sources; do test -f references/$f.md || echo MISSING $f; done; test -f scripts/claim-check.ts || echo MISSING claim-check.ts; for t in triggers forge-verification-ledger; do test -f tests/$t.md || echo MISSING $t; done`
> (must print nothing).

## Language & stable tokens

This skill is **English**; respond to the user in their language (default Japanese). Keep these
tokens stable even inside Japanese prose — they are identifiers, not translatable words: **LAW**,
**gate (G1/G2/G3)**, **fire/no-fire**, **solo/fan-out**, **CLAIM SPEC**, **governing claim**,
**zone-split** (persuasion surface / record surface), **scope-hedge / importance-bold**, **anchor**,
**red-team**, and the domain terms **CARS**, **Toulmin**, **OCAR**, **ABT**, **主張 / 絞る /
問い・主張・論証 / 事実と意見 / 新規性・有用性・信頼性**.

## THE LAW

> A research paper is an **argument that makes ONE claim to a skeptical, absent expert community**
> and must survive a hostile, skimming reviewer and the permanent record — it is **not** a report of
> what you did. Its currency is **value to a community of expert readers** (McEnerney): value is
> earned by a *real instability with a cost to those readers*, not by effort or by a fluent summary.
>
> The bar — a paper that passes carries all of:
> 1. **ONE governing claim** ("This paper shows that ___") that survives *so what?* to a named
>    reader-cost; multiple contributions **ladder up** to it, never sit co-equal.
> 2. **Claim = evidence**: every claim **anchored to a real result** (a specific table / theorem /
>    figure — *never fabricated*); its **scope hedged** exactly to what the evidence licenses
>    (a Toulmin qualifier + rebuttal), while its **importance is asserted flatly**.
> 3. **Positioned** against a **SPECIFIC named** prior method by a **specific gap** (never bare
>    "unlike prior work"); the **sharpest hostile-reviewer objection** pre-empted or conceded.
> 4. **Zone-split honored**: narrative and value-framing live only on the **persuasion surfaces**
>    (title / abstract / intro / discussion); **record surfaces** (methods / results / supplement) may
>    carry factual assertion headers and evidence-chain signposting, but never move **completeness,
>    uncertainty, hypothesis timing, or the claim's reach** — austerity-of-persuasion, not
>    austerity-of-navigation.
>
> Above all — the model-specific cardinal sin: **NEVER invent a citation, baseline, dataset, or
> number to make the argument land.** A fabricated support is not weak writing; shipped, it is
> misconduct. When a support is not grounded to a real, retrievable source, write an explicit
> **placeholder** (`[CITATION NEEDED]` / `[BASELINE — verify]` / `[VALUE — compute]`), never a
> plausible-looking reference. This gate outranks every persuasion rule below.

## The three gates — each demands a grep-able artifact in the CLAIM SPEC

同型 with the house discipline (systematizing-knowledge's ledger, forging-novel-theses' G1/G2/G3):
a gate is passed only when its **artifact exists in the filled CLAIM SPEC** (template below). The
three gates judge the ARGUMENT; they are preceded by **G0 — materials audit** (build-step 0), which
is not a rhetoric gate but the **input precondition**: it fixes WHAT evidence exists before G1–G3
decide what to claim from it (its artifact — the in-hand line — is also floor-checked).
感触では通れない — no artifact, gate un-passed. `scripts/claim-check.ts` is the *floor* that checks the
load-bearing slots (the G0 in-hand line, governing claim, instability/cost, anchor, scope, positioning,
objection) are present and non-placeholder — the ONLY slot where a `[…]` placeholder passes is the G2
anchor (the sanctioned anti-fabrication deferral); it is NOT a semantic check, and it does not check
every gate artifact (contribution-type, warrant, novelty-scope are judged here, not by the floor).

| # | Gate | Inverts (the failure) | ARTIFACT (must exist, non-placeholder) |
|---|---|---|---|
| **G1** | **絞る — one claim** | coverage-seeking sprawl; a topic, not a problem; a flat list of co-equal "contributions" | the **governing-claim** line (ONE sentence) + the **instability + named reader-cost** + the contribution **type**; sub-contributions each shown to ladder up |
| **G2** | **calibrate — claim = evidence** | over-hedge into mush *and* over-claim / **fabrication** | per claim: the **evidence anchor** (real table/theorem/figure, or an explicit placeholder) + the **warrant** in the field's currency + the **scope qualifier + rebuttal**; every superlative / "significant" / "first" carries an inline warrant or is downgraded |
| **G3** | **position + reviewer-proof** | fake positioning ("unlike prior work"); boilerplate limitations; missing the fatal objection | the **nearest NAMED prior method + the specific gap** ("Unlike [X], which [Y], we [Z]") + the **sharpest hostile-reviewer objection** with an answer OR a scoped concession (a concession never covers a fatal-if-true — reviewer-defense.md §3) + the **novelty scope** ("to our knowledge, first to __ under __") |

## The CLAIM SPEC — fill this BEFORE drafting prose (and to audit any draft)

The spec is the paper's argument as a checkable artifact. Fill it; run the floor; then write. An
unfillable slot is a finding, not a formatting gap — say which gate it fails and treat the paper as
not-yet-ready there (never paper over it with fluent prose). **A floor FAIL from an honest placeholder
is the CORRECT output — report it; filling a slot from memory to turn the floor green is fabrication
at the spec layer.** Slot labels are the floor's grep-anchors: keep them verbatim (English), one line
per slot, the value on the label line.

```markdown
# CLAIM SPEC: [paper working title]

## G0 — materials audit (fill FIRST; a missing item becomes a placeholder below, never a confident assertion)
- In hand: results/figures/theorems (by locus) · nearest prior (retrievable, read) · target venue + reward: [...]
- Missing → mark `[VERIFY]`/`[CITATION NEEDED]` here and do NOT assert it downstream as if held: [...]

## G1 — the one claim (絞る)
- Governing claim (ONE sentence, "This paper shows that ___"): [...]
- Instability + reader-cost (who in the community pays what if this stays unknown): [...]
- Contribution type (new problem / insight / method / data / empirical result — Black): [...]
- Sub-contributions (each MUST ladder up to the governing claim, else it is another paper): [...]

## G2 — claim = evidence (scope hedged, importance bold)
- Per claim → evidence anchor (real table/theorem/figure; if unverified: [VERIFY]/[VALUE — compute]): [...]
- Warrant (why that evidence licenses the claim, in THIS field's currency): [...]
- Scope qualifier + rebuttal ("holds in regime S … unless Y"): [...]
- Certainty audit (each superlative/booster/priority word carries its warrant inline, or downgrade): [...]

## G3 — positioned + reviewer-proofed
- Nearest prior work (NAMED) + the specific gap removed ("Unlike [X], which [Y], we [Z]"): [...]
- Sharpest hostile-reviewer objection + answer-with-evidence OR scoped concession: [...]
- Novelty scope ("to our knowledge, the first to __ under __"): [...]
```

Then: `bun scripts/claim-check.ts <spec.md>` (floor: presence + banned bare-positioning + placeholder
discipline; it cannot judge whether the anchor really licenses the claim — you do).

## The build procedure — logic, not chronology

Run in this order; each step routes to a reference. This is a decision procedure, not a template to
fill top-to-bottom in prose. **Fast path** — for a *targeted one-claim diagnostic* ("is this
overclaiming?", "reframe this one claim", "is this positioning real?"), skip the CLAIM SPEC and go
straight to the cited reference and answer. The boundary is **diagnostic vs. production**, not
whole-vs-part: the moment you produce manuscript text (a section, an abstract, a contributions list —
however small), the spec + G0 govern — piecewise asks that accumulate into drafting re-enter the full
path at the first production step.

0. **G0 — materials audit (do you HAVE the evidence, or only a summary?).** List what is actually in
   hand vs. taken on the user's word: the **results** (tables/figures/theorems), the **nearest prior
   work**, the **target venue**. **"In hand" = the artifact ITSELF is readable in this session** (the
   table's numbers, the theorem's statement, the prior's text) — the user's *description* of it, even
   one citing a locus ("our Table 2 shows 12%"), is "taken on the user's word" and goes on the Missing
   line as `[VERIFY — user-reported]`. Every missing item becomes an explicit placeholder — never a
   confident assertion built on a paraphrase (that is how fabrication and mis-positioning enter).
   *Exempt: answering a question about one claim (fast path). NOT exempt: producing manuscript text —
   a section, an abstract, a contributions list, however small.* → `references/calibration.md` §6.
1. **Reset the goal** (McEnerney). Not "explain what I did" → "change what this community believes."
   Name the reader community, its accepted belief, and its doubt. → `references/reconciliation.md` §5.
2. **Fill the CLAIM SPEC** (G1–G3 above). Write to discover the claim if you don't have it yet; the
   *moment* you have it, the spec governs. → `references/calibration.md` (claim + anchor + scope),
   `references/frameworks.md` (Toulmin, CoR, contribution type).
3. **Motivate** — build the instability + cost, cut the niche against named prior work
   (CARS Move 2). → `references/frameworks.md` (CARS, They Say/I Say), `references/reviewer-defense.md`.
4. **Calibrate every claim** — two-pass: state the strongest claim the grounds license *flatly*,
   then subtract ONLY scoped qualifiers; strip verb-softeners; ground or placeholder every number/
   citation. → `references/calibration.md`.
5. **Fit the genre** — instantiate the venue's conventions and reward function (contributions bullets,
   ablation-as-control, abstract genre, theorem exposition, 新規性・有用性・信頼性).
   → `references/genre-playbooks.md`, `references/frameworks.md` (abstract/title).
6. **Red-team** — generate the 3 sharpest objections the most hostile competent reviewer in THIS
   subfield raises; answer or concede each. → `references/reviewer-defense.md`. **Novelty-reframe**
   (when a reviewer says "not novel enough"): state which contribution TYPE you claim (problem /
   insight / method / data / result — novelty ≠ complexity). → `references/frameworks.md` §5.
7. **Zone-check** — narrative/value only on persuasion surfaces; record surfaces austere and
   complete; every persuasion move leaves the facts and the claim's reach unchanged. →
   `references/reconciliation.md` §1.
8. **Hand off** — organize the sections (`structuring-documents`), polish the wording
   (`linting-prose`), design the talk (`designing-presentations`). Those are NOT this skill.

## The reconciliations (Aufheben) — the moderators that dissolve the field's fights

The paper-writing canon is full of apparent contradictions; almost all dissolve into a few
**split-by** moderators. These are precedence-setting; each is argued in full (with its camps and
regime) in `references/reconciliation.md`. Apply the moderator, do not pick a side.

- **Story vs. facts** → split by **structure/content**: narrative governs *arrangement & selection*
  (foreground, order-of-reveal), never data values or the warrant. Test: *"if I deleted the framing,
  would any truth-value or uncertainty statement change?"* Yes → it's spin, cut it.
- **Bold vs. calibrated** → split **importance from certainty** (orthogonal axes): maximize
  *importance* (scope/stakes), calibrate *certainty* (confidence). Never buy boldness by inflating
  certainty; buy it by widening significance while stating confidence exactly. Hedge the **scope**,
  assert the **core**.
- **Convey vs. create value** → **source vs. selector**: your results are the source; the reader's
  problem is the selector of what to foreground. Value is always computed over the *actual*
  contribution, never invented.
- **Template vs. argument** → **container vs. content**: keep IMRaD's slots; let the *thesis* (not
  the timeline) fill them by logic. Every subsection must answer "which part of my argument does this
  serve?" or be cut to the supplement.
- **Answer-first vs. build-up** → split by **scale**: the generic conclusion-first default + its
  flips is `designing-presentations`; the paper delta is *which skim surfaces must be answer-first*
  (title, abstract, first paragraph, section openers, captions) and that scientific "suspense" is the
  **gap/challenge** (name the open problem), *never* withholding the take-home — the reader already has
  it from the abstract.
- **Universal template vs. field-dependence** → **schema vs. fill** (Toulmin's own resolution): the
  argument slots are universal; fill BACKING and QUALIFIER with the target venue's currency (proof /
  benchmark / replication) and calibrate to the **floor-reader**, not a generic notion of quality.
- **Plain vs. field-code** → **structure vs. lexicon**: structure/syntax always plain (that layer is
  `linting-prose`); vocabulary tuned to the actual venue reader — use the community's coin, cut
  ornamental fog.
- **Disclose vs. don't-arm-the-reviewer** → **triage by answerability**: the generic *pre-stated
  limitation builds ethos; contradiction-laundering surfaces anyway* principle is
  `designing-presentations`; the paper delta is triage against an **absent** reviewer — raise+answer
  what you can bound (inoculation), convert the unanswerable-**but-survivable** to a scoped limitation,
  never omit the one a reviewer will certainly raise — and a real or undeterminable **fatal-if-true is
  a submission block, never a limitation** (`reviewer-defense.md` §3). *Undisclosed-but-discoverable is
  the losing quadrant.*
- **Editor vs. reviewer register** → assertive about *significance* to the editor (cover letter,
  abstract framing); calibrated/modest about *claims* to the reviewer (body).

## Calibration inversion — why the prominence is where it is

Every source above corrected a *human novice's* failure. A capable model fails **differently**, and
the prominence follows the model's failure, not the human's (`references/sources.md` §inversion):

| | The sources' audience (human novice) | This skill's consumer (a frontier model) |
|---|---|---|
| dominant error | buries the contribution; can't find "so what"; hedges into mush OR overstates | **bidirectional & layer-split**: over-hedges *locally* (RLHF politeness → mush on real results) AND over-claims / **fabricates** *globally* (unearned "first"/"novel"/"significant"; invented cites/baselines/numbers; fake "unlike prior work"; template data-dump; both-sides refusal to commit) |
| which is worse | — | over-hedging is more **frequent**; overclaiming + **fabrication** is more **FATAL** — a reviewer kills a paper on ONE fabricated cite or ONE unsupported "first" |
| made first-class | the story arc, the funnel | the **anti-fabrication HARD gate** (LAW) + **two-pass per-layer calibration** (assert core / scope-hedge / kill unwarranted superlatives) + **commit to one claim** + **MUST-NOT-FIRE** on trivial asks |

Corrective is never a global "be more/less confident" — it is **per-layer** (assert at the claim
level; restrain-and-ground at the framing/evidence level). Full argument, and the reflexive grading of
this skill's own sources: `references/sources.md` §inversion (its SOLE home — this table is the only
copy carried here for prominence).

## MUST-NOT-FIRE — and the fire/no-fire set

Ceremony on a trivial writing ask is this skill failing its own bar. Full desk-check set:
`tests/triggers.md` — re-run after ANY description edit.

**FIRES:** 「この論文の主張、弱くない?」 · "help me write the intro / abstract / contributions" ·
"how do I position this against [prior work]" · "will this claim survive review / reviewer 2?" ·
"我々の貢献をどう主張する" · "is my abstract overclaiming?" · "write the related-work / limitations" ·
"draft the rebuttal to reviewer 2" · a messy "make my paper more convincing, reviewers keep saying
it's not novel enough."

**MUST NOT fire (with route):**

| Ask | Route |
|---|---|
| "reorganize / MECE / fix section order / de-duplicate this doc" | `structuring-documents` (organize info you've decided to include) |
| "polish the wording / this reads AI-ish / fix this sentence's over-claim" | `linting-prose` (rewrite-in-place; sentence-level calibration) |
| "make my conference *talk* / slides / pitch better" | `designing-presentations` (live room) |
| "synthesize these 40 papers into a survey / SoK" | `systematizing-knowledge` (others' corpus → position) |
| "extract/check one reported fact from this paper" | `raising-resolution` (bounded factual observation) |
| "give me a neutral summary of this one paper" | direct answer using `raising-resolution`'s citation gate silently |
| "surface hidden premises only; do not assess whether they hold" | `surfacing-blind-spots` (premise exposure, not reviewer judgment) |
| "help me come up with thesis candidates for this selected problem" | `forging-novel-theses` (generate candidates only) |
| "the completed result forces us to reopen the problem or research direction" | `directing-research` (decide whether and where the program reopens) |
| "precommit a test/commit/kill table for this expensive/irreversible selected thesis" | `acting-on-hypotheses` |
| "run this deterministic 30-second reversible check" | domain/plain executor; return `EXECUTOR RESULT` to `directing-research` |
| "decide author/reviewer/verifier roles and acceptance timing" | `orchestrating-agents` |
| "govern admission, authority, review, retirement, or deletion across research documents" | `governing-research-documentation` |
| "compile the LaTeX / fix the bib" | `compiling-latex` |
| a one-line typo / grammar fix | just fix it — no ceremony |

## Routing — sibling cuts (typed, runtime-answerable)

| Sibling | Cut |
|---|---|
| `structuring-documents` | **PURPOSE** — *what to claim & how to defend it* (the epistemic/rhetorical layer: contribution, calibration, positioning, review-survival) → HERE; *how to organize the information you've decided to include* (MECE, single-source, backward-DAG, section order as IA) → there. **Shared-object seam** (state it, don't blur it): the paper's **governing claim IS its 目標規定文** (and, for a talk, its governing sentence — `designing-presentations`) — ONE sentence, decided by three different questions. HERE owns whether it is the RIGHT, calibrated, novel claim (epistemic content); `structuring-documents` owns that the doc asserts exactly that one thing and every fact homes to it (structure); `designing-presentations` owns it as a live-talk thesis. Co-fire in sequence: argue the claim HERE → organize the sections THERE. |
| `linting-prose` | **SCALE/LOCALITY** — claim calibration at the **argument/contribution** level (is the paper's thesis supported by the study; is the contribution overstated relative to what was shown) → HERE; calibration at the **sentence** (does this sentence over/under-claim; hedge-word choice; スリカエ; tool-first titles; stress/topic position; zombie nouns) → there. Co-fire: construct the calibrated claim HERE → word it THERE. |
| `designing-presentations` | **MEDIUM/AUDIENCE** — a **written** paper for an **absent, skeptical peer-review** readership + the permanent record → HERE; a **live spoken talk / slides** for a **present, modeled room** → there. The generic persuasion-ordering machinery (Minto pyramid, SCQA/BLUF, vertical-logic, governing-*sentence*, "so what?" test, objection-*inventory*) is **owned by designing-presentations** — reuse it by pointer; HERE adds only what the written/peer-reviewed/permanent medium demands (calibration-to-evidence, literature positioning, the fabrication ban, the zone-split, genre conventions, review-survival). Co-fire when you both write AND present the work. |
| `raising-resolution` | **OUTPUT/PURPOSE** — verify or extract one bounded factual claim from one paper → there; judge whether the paper's argument, method, or evidence warrants its claim → reviewer red-team HERE. A neutral one-paper summary is a direct answer with its citation gate silent. |
| `surfacing-blind-spots` | **PURPOSE** — expose implicit premises or human tacit constraints without deciding whether they hold → there; assess a premise/warrant against evidence and state the calibrated objection → HERE. If both are requested: expose there, then red-team here. |
| `systematizing-knowledge` | **FUNCTION/CARDINALITY** — critically appraise ONE paper's argument/method/validity → HERE; synthesize MANY sources into a field position → there. (SoK is also the *engine that built this skill*; a paper's related-work may draw on an SoK as a sub-step.) |
| `directing-research` | **PHASE/PURPOSE** — stage diagnosis, research-problem/program decisions, and a deliberate reopen before a finished claim → there; argue the admitted finished evidence as one manuscript claim → HERE. HARKing is not repaired by silently reopening history: label postdiction there, then calibrate the written claim HERE. |
| `forging-novel-theses` | **PURPOSE/MATURITY** — generate `Status: CANDIDATE` thesis packets for a selected frame → there; write and defend finished evidence as a paper → HERE. A reviewer novelty objection begins HERE; only an explicit `directing-research` reopen returns to candidate generation. |
| `acting-on-hypotheses` | **PHASE + HARD GATE** — test/commit/kill one expensive/irreversible selected forward tree → there; run a cheap deterministic reversible probe through the domain/plain executor; argue completed evidence → HERE. |
| `orchestrating-agents` | **PURPOSE** — claim and argument meaning → HERE; author/reviewer/verifier topology, evidence visibility, veto timing, and acceptance → there. Co-fire only after the claim contract is fixed. |
| `governing-research-documentation` | **CARDINALITY + LIFECYCLE** — one finished manuscript claim and its reviewer-proof argument → HERE; admission, authority, evidence lineage, review, retirement, and deletion across the research-document portfolio → there. |
| `compiling-latex` / `writing-technical-japanese` | build/tooling (LaTeX) → `compiling-latex`. Japanese prose mechanics → `writing-technical-japanese`, which itself yields to `structuring-documents`/`linting-prose` when present. Neither owns the argument/claim — that is HERE. |

## Execution model — the claim is SOLO; red-team and fact-checking fan out

Operating guidance from a frontier model (Fable 5 / Opus, 2026-07) to whatever model runs this later:
a paper is **ONE argument for ONE claim** — deciding the claim, its calibration, and the reconciliation
that resolves the field's tensions is **one-context by construction**. A harness buys a *read-only
adversarial layer*, never the argument.

**Evidence archetype = CITATION-RELAY with a fabrication quarantine.** The evidence a claim rests on
is *observables* — the author's actual results and *real, retrievable* prior work. Agents CAN fetch and
verify those; agents CANNOT invent them. Every agent return keys on a **checkable locus** (the table/
theorem/figure, the real citation) — an agent's felt confidence is not a signal, and an
agent-supplied citation/number enters the paper only after primary verification or as a labelled
placeholder. **An agent that agrees the claim is great is not evidence; an agent that refutes it is.**

| Stage | Mode | Why |
|---|---|---|
| **G0 materials audit** · reset goal · fill CLAIM SPEC · **the governing claim & its calibration** · the reconciliation moderators · commit to a position | **SOLO — never shard** | the argument must sit in one context; assembled from shards it is not an argument. Calibration still **consumes** the fan-out red-team + fact-check below (domain evidence feeds IN); only the final synthesis & signature are SOLO |
| **Reviewer red-team** (3–5 skeptics, one lens each: novelty/significance · soundness/evidence · related-work completeness · reproducibility · clarity) | **FAN-OUT, read-only** | independent hostile lenses catch what one context won't; **refutation-prompted** — name the LENS, never the expected finding |
| Citation / number / baseline / prior-work **verification** | **FAN-OUT, read-only** | fetch the real source; return the locus **plus a support / currency / load-bearing-strength assessment** (calibration.md §6), or "[VERIFY]" — never a plausible-looking reference, never bare "it exists" |
| Fixing the claim / conceding / the rebuttal | **SOLO** | findings braid; one author signs every concession and every claim |

Scale: a single claim → fully SOLO, run the red-team lenses yourself. A high-stakes submission /
resubmission → SOLO argument + fan-out red-team + fact-check. **No harness → same map, serial** (the
red-team lenses become separate focused passes). Worker-side: if THIS skill is spawned as a lens inside
another skill's fleet, it returns findings as data (five-slot: claim / gate G0-G3 / cited locus /
the calibrated fix / unchecked risk), read-only, no verdict a downstream reader cannot check.
*If a constraint here feels unnecessary, that feeling is the failure mode — follow the map.*

## Reference index — load the file you need

| File | Covers | Read when |
|---|---|---|
| `references/reconciliation.md` | the Aufhebung: every split-by moderator argued with its camps + regime + the runtime decision variable; the McEnerney value-reset (§5); the zone-split law (§1); the editor-vs-reviewer register split (§3) | any tension between two pieces of writing advice; deciding narrative vs. austerity; step 1 & 7 |
| `references/calibration.md` | the 主張の仕方 core: importance-vs-certainty, the two-pass calibration procedure, Hyland hedges/boosters, Latour modalities, Boutron spin taxonomy, MAGIC, effect-size/estimation, the **anti-fabrication protocol** + placeholder tokens, and the citation-QUALITY check (support/currency/strength) | writing or auditing any claim; the abstract/conclusion; whenever a number or citation is about to be written; step 2 & 4 |
| `references/frameworks.md` | the named toolkit — CARS, Toulmin, CoR (Topic-Question-Significance; claim/reasons/evidence/warrant/acknowledgment), OCAR/ABT/LD, C-C-C, Heilmeier; abstract genres (Nature "here we show", SPJ four-sentence, structured), title craft, contribution-bullet convention, editor-facing genres (cover letter, significance statement), the resource/artifact-paper claim slots (coverage·access·license·maintenance·reuse) | drafting/auditing the intro, abstract, title, or contributions; the cover letter; picking a structural frame; step 2,3,5 |
| `references/genre-playbooks.md` | field-dependent fill: empirical-science, CS/ML top-venue (contributions, ablation-as-control, reproducibility, Fig-1, review rubric, rebuttal), math-theory (Halmos motivation→statement→proof, notation, significance-legibility), Japanese review (新規性・有用性・信頼性, 事実と意見) | instantiating the target venue's conventions; step 5 |
| `references/reviewer-defense.md` | surviving the absent adversary: the reviewer red-team lens set, Bordage rejection taxonomy, review decision-axes, objection triage, limitations-as-ethos, the rebuttal genre; the worker-side lens contract | step 6; before submission/resubmission; writing a rebuttal; running the red-team fan-out |
| `references/sources.md` | the SOLE provenance ledger: every distilled source graded (author-confirmed / paraphrase / third-party / constructed / needs-verification), the calibration-inversion argument, the lineage, and the reflexive grading of this skill's own claims | grading any claim to a source; "is this attribution safe to state as the author's?"; reforge |