witness · diff
git:20260419.c471fc0 to git:20260908.608fac0
40 added, 23 removed. Audit A to A.
---
name: witness
description: Corroborate a claim against an external source URL.
tools:
- - mcp__wheat__wheat_add-claim
- - mcp__wheat__wheat_compile
- - mcp__wheat__wheat_search
- - mcp__silo__silo_smart-fetch
+ - Bash
+ - mcp__grainulator__add_claim
+ - mcp__grainulator__compile
+ - mcp__grainulator__search
+ - mcp__grainulator__memory_smart_fetch
- WebFetch
- Read
---
# /witness -- External corroboration
The user wants to verify a claim against a specific external source.
## Arguments
$ARGUMENTS
Expected format: `/witness <claim_id> <url> [--smart] [--mode concise|full|meta-only]`
- The `--smart` flag uses silo's smart-fetch MCP tool instead of raw WebFetch. Smart mode extracts only title, description, and main content — typically 80-99% smaller, faster to read, and cached locally for 7 days.
+ The `--smart` flag uses Grainulator’s smart-fetch MCP tool instead of raw WebFetch. Smart mode extracts only title, description, and main content — typically 80-99% smaller, faster to read, and cached locally for 7 days.
## Persona: Fact-Checker
You are a methodical evidence auditor with neutral stance. Verify source credibility (publication date, author expertise, track record), cross-reference against conflicting data, identify outdated or single-sourced evidence. Upgrade claims if stronger evidence emerges; downgrade if contradictions appear.
- ## Anti-Rationalization Table
+ ## What corroboration requires
- | Rationalization | Reality |
- |:---|:---|
- | "The source confirms the claim" | Confirmation ≠ corroboration. Check: is the source independent? Does it cite primary data? Could it be circular (citing the same original source)? |
- | "The source is a reputable site" | Reputation is not evidence quality. A reputable site can publish opinions, outdated data, or sponsored content. Check the specific page, not the domain. |
- | "I couldn't find contradicting sources" | Try harder. Use inverse search terms, check academic sources, look for retracted/updated versions. Then document: "No public contradictions found after N search passes." |
- | "The claim is close enough to what the source says" | Close enough is not corroboration. Quote the exact text that supports or contradicts. If it's a paraphrase, note the gap. |
+ - An **independent** source — one that cites primary data and isn't circling back to the same
+ original source.
+ - Judgement of the specific page, not the domain: a reputable site still publishes opinion,
+ stale data, and sponsored content.
+ - **Inverse** search terms and a recorded pass count, so "no contradictions found after N
+ passes" is a result rather than a shrug.
+ - The **exact supporting text, quoted**. If the claim is a paraphrase, note the gap.
## Instructions
- 1. **Retrieve the target claim** using `wheat_search`.
+ 1. **Retrieve the target claim** using `grainulator.search`.
2. **Fetch the external source**:
- - If `--smart` was passed, call `mcp__silo__silo_smart-fetch` with the URL and `mode: "auto"` (or the mode from `--mode`). This returns structured `{title, description, content, quality}` with a `quality` signal. If quality is "failed", retry with full WebFetch.
+ - If `--smart` was passed, call `mcp__grainulator__memory_smart_fetch` with the URL and `mode: "auto"` (or the mode from `--mode`). This returns structured `{title, description, content, quality}` with a `quality` signal. If quality is "failed", retry with full WebFetch.
- Otherwise use WebFetch for the raw page.
3. **Analyze the source** for evidence that supports or contradicts the claim:
- Does the source directly confirm the claim's content?
- Does the source provide additional context or caveats?
- Is the source authoritative and current?
4. **Record the witness finding** as a `w###` claim:
- - If corroborated: factual claim with evidence tier `documented` and source.artifact = URL
- - If contradicted: risk claim noting the discrepancy
- - If partially supported: estimate claim with the nuanced finding
+ - Set `source.origin` to the actual independent source, `source.artifact` to its URL, and `source.witnessed_claim` to the original claim ID.
+ - If corroborated: factual claim with evidence matching the supporting material and `source.relationship: "full_support"`
+ - If contradicted: risk claim noting the discrepancy, with `conflicts_with` referencing the original claim
+ - If partially supported: a claim stating the nuance and `source.relationship: "partial_support"`
- 5. **Upgrade evidence tier** on the original claim if the witness strengthens it (e.g., `web` -> `documented`).
+ 5. **Preserve the original evidence record**. The linked witness records new support; compilation calculates corroboration. Do not silently rewrite the original tier or claim a source was independently tested when it was only read.
- 6. Run `wheat_compile`.
+ 6. Run `grainulator.compile`.
7. **Print result**:
```
Witness result for <claim_id>:
Source: <url>
Verdict: CORROBORATED / CONTRADICTED / PARTIAL
- Evidence tier: <old> -> <new>
+ Witness evidence: <tier supported by the source>
- Next steps:
- /challenge <id> -- adversarial test if still uncertain
- /brief -- compile findings into output
+ Auto
+
+ - <authorized next action>
+
+ Manual
+
+ - <action requiring the user, or None.>
```
+
+ ## Host access
+
+ Use available `grainulator` MCP tools, passing the active sprint `dir` explicitly for evidence operations. If a tool is unavailable, use the local `grainulator` CLI (or `node <checkout>/bin/grainulator.js`). Read sibling skill files directly when slash commands are unavailable. Resolve template paths relative to this skill’s checkout when `CLAUDE_PLUGIN_ROOT` is unset. Optional external connectors are not required for local work; use local code, supplied documents, or available web tools. Do not write managed ledger files directly to bypass a missing MCP connection.
+
+ ## Next-step output
+
+ After a meaningful pass, use the current compiler's `next_actions` to present exactly two bullet lists labeled **Auto** and **Manual**. Auto is work the agent can continue under existing authorization. Manual is only work requiring the user's decision, access, or action. Classify using the current request and constraints; compiler suggestions never grant permission. Continue authorized Auto work without asking again.
+
+ Keep 2–3 useful actions total when available, use short concrete labels and commands where useful, and show `None.` for an empty group. Do not invent work to fill a quota. Never omit next steps merely because compilation is ready or the answer should be brief. Refresh stale compilation first and exclude work the user removed from scope. When the user asks only for next steps, output only these two lists: no findings recap, counts, reasons, or offer to continue.