evidence-lock

Evidence Lock Skill

For most drafts, plausible is enough. This mode is for the documents where someone will check: every substantive sentence either cites the exact passage in the user's sources that supports it, or wears an explicit [UNSOURCED] flag. No third state.

What This Skill Produces

• The document, with numbered footnote markers on every substantive claim
• A source map: each footnote → source name + the supporting passage quoted verbatim
• An unsupported-claims register: every [UNSOURCED] item with what evidence would resolve it
• A coverage score: % of substantive claims that are locked

Required Inputs

Ask for (if not already provided):

• The sources — pasted documents, files, or excerpts. This skill cannot run without them; general knowledge is not a source here.
• The task — either a draft to lock (rewrite mode) or a brief to write from scratch (compose mode)
• Strictness — hard lock (unsupported claims are removed to the register) or soft lock (they stay in the text, flagged [UNSOURCED]). Default: soft.

Locking Method

1. Index the sources first. Skim all provided material and note what each source can support. If sources are thin relative to the task, say so up front — don't compensate with fluency.
2. Classify each sentence while writing: substantive (factual claim, number, attribution, causal statement) → needs a lock; structural (transitions, headings, statements of the document's own intent) → exempt. When in doubt, it's substantive.
3. Lock = quote-level, not document-level. A footnote cites the source and the passage: [3] → "Q2 churn analysis, §4: 'logo churn concentrated in accounts under $10k ACV (71% of losses)'". Citing a whole document is not a lock.
4. No stretching. The passage must actually support the claim as written — not a weaker cousin of it. If the source says "churn rose in Q2" and the draft says "churn is accelerating", that's [UNSOURCED] (or the sentence gets weakened to what the source supports — prefer weakening).
5. Conflicts surface, never average. Two sources disagreeing produces both citations and a visible note, not a blended number.
6. Inference is allowed but labelled. A conclusion derived from cited facts gets [inference from 2,5] — distinguishing sourced, inferred from sourced, and unsourced.

Output Format

[Document title] — evidence-locked · coverage: [n]% ([x] of [y] substantive claims)

[The document. Substantive claims carry [n] markers; unsupported ones carry [UNSOURCED] (soft) or are absent (hard). Inferences carry [inference from n,m].]

---

Source map

| # | Source | Supporting passage (verbatim) |

|---|---|---|

| 1 | [doc, section] | "[exact quote]" |

Unsupported claims register

| Claim | Why it's unsourced | What would resolve it |

|---|---|---|

Conflicts noted: [source A says X; source B says Y — surfaced at footnote n]

Quality Checks

• [ ] Every substantive sentence has a footnote, an [UNSOURCED] flag, or an [inference from …] label — zero unmarked claims
• [ ] Every footnote quotes the passage verbatim; spot-checking any quote against the source succeeds
• [ ] No passage is stretched — each supports the claim as written
• [ ] The coverage score is computed by counting, and low coverage is stated plainly, not disguised
• [ ] Source conflicts appear in the text, not silently resolved

Anti-Patterns

• [ ] Do not fill source gaps with general knowledge — in this mode the provided sources are the entire universe of evidence
• [ ] Do not cite documents wholesale — a lock names the passage
• [ ] Do not launder inference as citation — derived conclusions are labelled as derived
• [ ] Do not quietly drop claims that can't be sourced in soft mode — the register exists so the author sees what's resting on air
• [ ] Do not proceed without sources "just this once" — without sources this is a normal draft, and other skills do that better