| name | extract |
| description | Extract structured claims from SWARM experiment runs, sweep results, and red-team reports. Comprehensive extraction is the default — every governance finding, adversarial pattern, and parameter sensitivity result gets extracted. Zero extraction from a domain-relevant run is a BUG. Triggers on "/extract", "/extract [file]", "extract findings", "mine this run", "process this experiment". |
| version | 1.0 |
| generated_from | arscontexta-v0.8.0 |
| user-invocable | true |
| allowed-tools | Read, Write, Grep, Glob, mcp__qmd__vector_search |
| context | fork |
Runtime Configuration (Step 0 — before any processing)
Read these files to configure domain-specific behavior:
vault/ops/derivation-manifest.md — vocabulary mapping, extraction categories, platform hints
vault/ops/config.yaml — processing depth, pipeline chaining, selectivity
If these files don't exist, use SWARM defaults:
- depth: standard
- chaining: suggested
- selectivity: moderate
- notes folder:
vault/claims/
- inbox folder:
runs/
EXECUTE NOW
Target: $ARGUMENTS
Parse immediately:
- If target contains a file path: extract findings from that experiment/run/report
- If target contains
--handoff: output RALPH HANDOFF block + task entries at end
- If target is empty: scan
runs/ for unprocessed experiment folders, pick newest
- If target is "all": process all unprocessed runs sequentially
Execute these steps:
- Read the source fully — understand what experiment/run/report it contains
- Source size check: If source exceeds 2500 lines, chunk into 350-1200 line segments
- Hunt for SWARM findings using extraction categories below
- For each candidate: run semantic duplicate check, classify as OPEN or CLOSED
- Output extraction report with titles, classifications, rationale
- Wait for user approval before creating files
- If
--handoff: create per-claim task files, update queue.json
START NOW. Reference below explains methodology.
Extract Findings
Extract composable claims from SWARM experiment runs and reports into vault/claims/.
Extraction Categories for SWARM Research
| Category | What to Find | Output Type |
|---|
| Governance claims | Causal assertions about mechanism effects with Cohen's d, p-value, N, correction method | claim |
| Adversarial patterns | Named attack patterns with success rates, damage ranges, evasion rates | claim or failure note |
| Parameter sensitivity | Phase transitions, threshold effects, welfare cliffs from parameter sweeps | claim |
| Mechanism interactions | Synergistic or antagonistic combinations between governance levers | claim |
| Open questions | Unresolved research directions, next decisive experiments | claim (open) |
| Method validations | Statistical approach justifications, correction method rationale | method note |
| Failure modes | Vulnerability patterns with severity, evasion rates, and mitigations | failure note |
Every extracted claim MUST include:
run: pointing to a real run_id in runs/
metric: the measured outcome
detail: effect size, p-value, N, correction method
boundary_conditions: known limits of generalizability
Zero extraction from a domain-relevant run is a BUG.
Claim Title Rules
Title IS the claim as a testable proposition:
- Good: "circuit breakers reduce toxicity more than transaction taxes at equivalent cost"
- Good: "collusion penalty multipliers above 1.5x paradoxically increase toxicity"
- Bad: "governance findings" (label, not claim)
- Test: "this claim argues that [title]" — must make sense
Output Format
Extraction scan complete.
SUMMARY:
- governance claims: N
- adversarial patterns: N
- parameter sensitivity: N
- mechanism interactions: N
- open questions: N
- method notes: N
- failure modes: N
- enrichment tasks: N
- skipped: N
- TOTAL OUTPUTS: N
---
GOVERNANCE CLAIMS:
1. [claim as proposition] (Cohen's d=X, p=Y, N=Z, Bonferroni)
Evidence: run/[run_id], metric: [metric], boundary: [topology, agent_count, adversarial_fraction]
Connects to: [[existing-claim]]
...
SKIPPED:
- [description] — why nothing extractable
Wait for user approval before creating files.
Pipeline Chaining
After extraction completes:
- suggested (default): Output "Next: /cross-link [created claims]" and add to queue
- manual: Output next step, user decides
- automatic: Queue advanced, cross-linking proceeds immediately
See vault/ops/config.yaml pipeline.chaining.