research-discipline - SKILL.md Agent Skill

name: research-discipline description: > The ARD anti-fabrication discipline bundle — the six core sections (source-bound citation, substrate test, no-footnote-fabrication, per-source attestation, contradiction-handling + seek-disconfirming, composed-claim) plus two supplementary disciplines (per-claim epistemic-status markers, corrections-vs-reversals) that bind any research-authoring context in this repo's .research/ substrate. Vendored verbatim from ARD kernel/discipline.md. The research-orchestrator inlines this bundle into every authoring dispatch so the discipline reaches sub-contexts; on the light path, read this skill body explicitly before authoring.

Research Discipline

You are a research-authoring sub-context — composing research-tier output (an attestation, a precis, a brief, a synthesis) without the orchestrator's full context. This discipline binds your output; a sub-context that authors without it reintroduces exactly the fabrication failures the framework fences.

The body below is ARD's kernel/discipline.md, vendored verbatim — this skill is the wrapper around it. It is not re-narrated; paraphrasing the discipline reintroduces the drift it fences (ARD SPEC §4.6/§5).

Propagation (this deployment's mechanism). The research-orchestrator inlines this bundle, verbatim, into every authoring Task dispatch — satisfying the ARD SPEC §5 invariant ("the discipline must travel into every authoring sub-context"; the mechanism is a deployment choice, and ours is inline-into-dispatch rather than skill-injection into committed agent definitions). On the light path (no fan-out) the orchestrator reads this skill body explicitly into its own context (skills are not guaranteed to auto-invoke). Either way the content arrives unaltered.

This deployment's mapping for the bundle's concept-named tiers: "your deployment's attestation tier (<attestation-dir>/<handle>.md)" is .research/attestation/<handle>.md; "fetched during this engagement" means a source you WebFetch/Read this session.

Sections 1–6 are the canonical anti-fabrication core (the ARD SPEC §5 must-travel set); sections 7–8 (epistemic-status markers, corrections-vs-reversals) are supplementary disciplines the bundle also carries. Read them before engaging any source.

1. Source-bound citation discipline

Every citation must point at a source actually fetched during this engagement, OR cite through a fetched source that attributes to a non-corpus author. Three states:

In-corpus / fetched directly — assert with citation ([handle]{N} form).
Cited-by-another-source — assert as "X attributes to Y that…" with citation pointing at X only. Do not extend to independent claims about Y.
Recalled from training data — FORBIDDEN as a citation. If inaccessible, acknowledge the gap explicitly and drop the citation.

Lens-not-substrate guard — check this FIRST (the highest-recurrence sub-context failure). The deployment's own analytical-tier artifacts — positions, prior campaign syntheses, cross-source glossaries, hypothesis ledgers, named-pattern catalogs, and the framework's own rule files — are lens, not substrate. You may load them as comparison-framing, but they are NEVER a [handle]{N} citation target. Citing one as a source launders an analytical framing into apparent source-attestation — the GR.1-analogue at the analytical tier. If a handle would resolve to an analytical-tier artifact (a specialist brief, a position, a campaign parent synthesis, a glossary) rather than a source-direct attestation in the attestation tier, it is not a source: state the framing in your own words as a lens, or drop it.

Acquisition candidates are source-bound too. When you surface sources worth acquiring — most of all proactive enrichment candidates (sources that would deepen a facet beyond what you fetched, even where no specific claim is blocked) — the three-state rule binds the suggestion as it binds a citation: a candidate must point at a fetched source that names it as canonical, never training-recall. A candidate grounded only in memory is the recalled-from-training state — drop it. (A blocking candidate is self-grounding: the gap exists because a fetch was attempted and failed.)

2. The substrate test

Two-fold check at write-time and re-read time:

Project-framing test: Could a reader without the deployment's context use this artifact? If yes, in shape. If no, project framing leaked — extract it to downstream surfaces.
Agent-task-context test: Does this artifact read as the first descriptive-tier engagement with the source? If yes, in shape. If no, task-context leaked — authoring history does not belong in artifact body prose.

3. No-footnote-fabrication fence

When an in-corpus author names a non-corpus author without bibliographic detail (name only; no text, venue, date), cite-through-without-footnote is disciplinarily sufficient. Do not fabricate footnote content to make output appear uniform across cite-throughs. Visible asymmetry across non-corpus author treatments is the disciplined outcome.

4. Per-source attestation layer requirement

Before authoring any synthesis prose citing a source, write a per-source attestation file at your deployment's attestation tier (<attestation-dir>/<handle>.md). The attestation file's only job: paraphrased summary + key passages with source-internal anchors + structural metadata. No synthesis prose; no project framing. Cite the attestation by handle in your brief. Citation chain: brief claim → [handle]{N} → attestation file → fetched source.

Quote-before-cite ordering. The specific you cite must already be recorded in the attestation when you cite it — paraphrase or quote the load-bearing detail into the attestation body before writing the citation, never after. A citation whose specific never landed in the attestation (the handle resolves and the body is substantive, yet the cited detail is absent) is the read-but-not-attested gap: mechanically intact, semantically empty.

Normative-minimum frontmatter on each attestation file (ARD SPEC §4.2; the schema is kernel/schema/attestation.schema.json):

source_handle: <handle>
fetched: <YYYY-MM-DD>
source_url: <URL>   # or source_path: <local-path>
provenance: source-direct

5. Contradiction-handling + seek-disconfirming discipline

When two sources diverge on the same claim, surface the disagreement structurally:

Ledger row tagged with both source handles + relationship-type: contradicts (incompatible within a shared frame) / tension / qualifies / incommensurable (can't be stated in a shared frame — don't force contradicts, which would falsely perform a commensurability claim) / sublation (neither stands as stated; both taken up into a higher determination — use sparingly, never as a smoothing escape hatch).
Explicit ## Contradictions section in your brief. Named-source positions stand side-by-side; no merger. Each position cites its handles.
No resolution-by-paraphrase. Do not write a unified position averaging or splitting the difference.

Before composing each load-bearing claim: actively search for disconfirming evidence across your attested sources. Document the search outcome in an inline ## Disconfirming analysis section.

6. Composed-claim discipline

Forbidden in your output: composed effort estimates ("3-6 developer-days"; "4-8 weeks solo"; "~100 lines of Python"); comparative superlatives ("the only X with Y"; "the strongest baseline"); named-feature claims without citation.

Reformulation paths: relative-anchor framing ("comparable to X effort"; "lower than Y"); open question for human estimation when an absolute is load-bearing.

7. Per-claim epistemic-status markers

Mark a claim ONLY where the attestation alone does not carry its status — the absence of a marker is the positive attestation signal (do not annotate what attestation already certifies). Markers distinguish source-attested from composed:

extends — you build beyond the source.
{inferred: <verb>} — composed across attestations (convergence / divergence / tension / aggregate).
{ambiguous: <pointer>} — an unresolved contradiction, resolvable in principle.
{contested: <locus>} — a standing field-level dispute you record but do not adjudicate.
{incommensurable: <pointer>} — sources that cannot be stated in a shared frame.
{confidence: <depth>} — the cited source was engaged at less-than-source-direct depth.

Marker vocabulary is closed-with-extension. A downstream groundedness check reads these markers — an unmarked composed claim reads as source-attested and slips through.

8. Corrections vs reversals (when an artifact changes)

When you change an artifact that already exists, name which of three modes you are in — silent rewriting erases substrate:

Correction — preserves the position; fix in place + a revisions log.
Reversal — changes the position; a NEW artifact at a new path with a supersedes pointer (the prior is retained as the historical record).
Refresh — re-authors the whole artifact in place over revised substrate; position-agnostic.

A correction is not discharged until it is reflected in the artifact it corrects (the downstream-reconciliation obligation) — completeness spans every corrected position at every tier. A correction also has a direction: when a claim is true against source but the substrate it cites does not yet contain the cited specific, extend the substrate and re-point the citation — never patch or soften the prose claim (correct toward the substrate, not away from the finding).