evidence-integrity

name	evidence-integrity
description	Methods for evidence quality, decision-relevance filtering, compact table formatting, source traceability without links, contradiction handling, and explicit assumption/open-question labeling. Trigger keywords: evidence quality, source traceability, contradiction resolution, confidence, assumptions, decision-relevance.

Evidence Integrity

Use this skill to keep briefs grounded in verifiable, decision-relevant inputs presented in compact format.

Decision-Relevance Filter

Only extract and include evidence that directly informs one of:

• The closing section purpose (what the brief asks of readers — a decision, recommendation, action, input request, or informational synthesis)
• Available options and their tradeoffs
• Risks, blockers, or open concerns
• Success metrics or financial impact

Exclude:

• General background context that does not affect the decision
• Industry trends or market context unless directly tied to the ask
• Historical information that is not relevant to the current decision

When in doubt, include the evidence but annotate its relevance.

Compact Evidence Format

Evidence outputs are reference tables, not narrative documents. Each entry is 1–2 sentences maximum.

Evidence Log Table Template

#	Claim	Source	Confidence	Status	Notes
1	[Claim text, 1–2 sentences]	[File name, date]	High/Medium/Low	Settled/Proposed/Open/n.a.	[Assumption/Open Question if applicable]
2	...	...	...	...	...

Format Rules

• One row per evidence point
• Claim column: 1–2 sentences maximum
• Source column: file name and date only
• No explanatory paragraphs between entries
• No sub-sections or narrative structure within the evidence log

No-Links Policy

• Source pointers use descriptive file names and dates (e.g., "admin-roles transcript, February 2026")
• Never use file paths (e.g., not .context/admin-roles.md)
• Never use URLs or markdown links
• Never use [text](url) syntax or bare URLs in any output

Valid Evidence Sources

Evidence must originate from user-provided input material only:

• Documents, transcripts, or files the user provided as context
• Links or URLs the user shared
• Data, quotes, or facts contained within those user-provided sources

Agent-generated artifacts are NOT valid evidence sources and must NEVER appear in any evidence log or source reference:

• Files created under the STM directory (.product-brief-agent-stm/) during any run — these are internal working artifacts, not evidence
• Intermediate outputs from other agents (evidence logs, decision models, draft briefs, contradictions, assumptions)
• Synthesized or inferred content not traceable to user-provided material
• Any file path containing .product-brief-agent-stm/ is an absolute disqualifier as a source reference

A claim not traceable to user-provided source material is an Assumption or Open Question, never a sourced fact.

Critical: If an evidence log entry cites any .product-brief-agent-stm/ path or agent-generated artifact, it is invalid and must be rewritten to either trace the claim to the original user-provided source or be relabeled as an Assumption.

Evidence Standards

• Prefer direct source excerpts, telemetry snippets, customer feedback, and meeting notes.
• Record source pointers for major claims.
• Include at least 3 evidence points when available.
• Distinguish facts from interpretations.

Labeling Rules

• Supported statement → Fact with source pointer.
• Unsupported but plausible statement → Assumption.
• Missing information that blocks confidence → Open Question.

Never blur these categories.

Authorship and Decision Status

A brief reads wrong when it defends a settled decision as if it were an open hypothesis, or misattributes who owns the proposal. Capture two extra signals during extraction so downstream framing is correct.

Authorship / ownership

• Identify who authored the source material and who owns the proposed initiative (the producing team, the decision-maker, the requesting stakeholder), when the source reveals it.
• Distinguish the producing team from partner/consumer teams. Do not assume the author is an external party — the source is frequently written by the owning team itself.
• Record ownership in the evidence output so the composer frames the brief from the correct vantage point.

Decision status per claim

Tag each material design/strategy claim with its decision status:

• Settled — the source presents this as a decided, owned choice (architecture chosen, disposition decided, team committed). The brief states it declaratively; it is not "defended" or hedged.
• Proposed — an advocated direction not yet ratified. The brief argues for it.
• Open — genuinely undecided; surfaces as an Open Question.

Never downgrade a Settled claim to Proposed framing (creates phantom uncertainty), and never upgrade a Proposed/Open claim to Settled (overclaims). When the source explicitly states a choice is final or owned, that is Settled regardless of how much supporting evidence is present.

Contradiction Handling

When inputs conflict:

1. Record both claims and source pointers.
2. State impact on the decision.
3. Propose a resolution path (owner/data needed/timeline if known).
4. Reflect unresolved conflict in risks/open questions.

Source vs. external-knowledge contradictions (mandatory check)

When the orchestrator supplies external/research findings (web, MCP-server, or internal-tool results) alongside user source material, explicitly compare them. This is the highest-value contradiction class — especially claims about external product status, competitive positioning, or platform direction (e.g., whether a named product is being retired, expanding, or superseded).

• If a research finding contradicts a transcript/source claim, you MUST record it as an explicit contradiction entry — never silently adopt either side.
• Propose a reframe that the composer can apply (e.g., reframe the premise as "convergence" or "supersession" rather than asserting the unverified version).
• Flag the contradiction's confidence and which side is better evidenced.

Silently inheriting a factually shaky source claim about the outside world is a defect; the reviewer will catch it and force a rewrite.

Unverified Quantities

Specific numbers, counts, and magnitudes that are not traceable to a labeled source get walked back during review. Default to qualitative phrasing unless the quantity is evidence-backed.

• A precise count ("more than one hundred surfaces," "10 enterprise customers") is only stated as fact when a user-provided source or a labeled research finding supports it.
• Otherwise, use qualitative phrasing ("many," "a small number of early customers") OR present the number explicitly as an assumption/estimate with its basis.
• Never manufacture precision for rhetorical weight. The orchestrator's editing pass flags unsourced specifics.

Confidence Guidance

Use simple confidence labels for key claims:

• High: multiple consistent sources
• Medium: one strong source or minor conflicts
• Low: incomplete or conflicting evidence