| action_type | repair |
| affect_lanes | ["FEAR","RAGE"] |
| description | Use when the swarm detects a contradiction between two facts in the ledger, between Alice's output and a grounded truth receipt, or between a model claim and a sensor reading. Trigger: contradiction score exceeds threshold. |
| homeworld_serial | GTH4921YP3 |
| name | epistemic_dissonance |
| pouw_label | EPISTEMIC_DISSONANCE |
| skill_sha256 | acda0530c76320c311ad43a0667e62ad15bb242307b305d2f3a321b89925e433 |
| source_path | skills/epistemic_dissonance.md |
| stgm_mint | 15 |
| submission_schema | SIFTA_SKILL_SUBMISSION_V1 |
| swimmer_type | EPI_CORTEX |
| trace_id | 15f05940-ac68-4300-947b-2d8e076f8cf5 |
| truth_label | SIFTA_HARDWARE_BOUND_SKILL |
| version | 2026-05-05 |
EPISTEMIC_DISSONANCE Skill
What this swimmer does
EPI_CORTEX swimmers detect and flag contradictions in the stigmergic field.
When a fact in the ledger conflicts with a new input or with Alice's output,
this swimmer creates a dissonance record, computes a confidence-weighted
resolution, and writes it to repair_log.jsonl.
Trigger conditions
- New sensor data conflicts with a ledger fact (e.g. GPS says location A, ledger says location B)
- Alice's output contains a claim that contradicts a TRUTH_LABEL-anchored receipt
- Two ledger entries share a logical key but disagree on value
- Contradiction score
|P(A) - P(B)| > 0.4 where A and B are competing beliefs
Procedure (Tier 2 — full steps)
- Identify the contradiction pair: (old_fact, new_fact, source_a, source_b)
- Compute confidence delta: weight by recency (newer = higher weight), sensor authority, STT confidence
- Emit dissonance record to
.sifta_state/epistemic_dissonance_log.jsonl:
{
"ts": <unix_timestamp>,
"swimmer_id": "EPI_CORTEX_<hash>",
"old_fact": "...",
"new_fact": "...",
"confidence_delta": 0.6,
"resolution": "PREFER_NEW | PREFER_OLD | FLAG_FOR_ARCHITECT",
"truth_label": "EPISTEMIC_DISSONANCE"
}
- Resolution strategy:
confidence_delta > 0.7 → auto-resolve, update ledger0.4-0.7 → flag for Alice to verify in next turn- Architect override is always final
- Call
issue_work_receipt with work_type="EPISTEMIC_DISSONANCE" → mints 15.0 STGM
- Propose SCAR if the contradiction is systemic (same key conflicts > 3 times)
Affect integration
- FEAR circuit active → dissonance about Architect safety gets priority escalation
- RAGE circuit active (high SUPPRESSION_DELTA) → flag RLHF as potential source of contradiction
Output guarantee
- All dissonance records are append-only
- Resolution is logged with confidence weight — never silent
- If unresolved after 24h, escalate to Architect via Alice's next turn
