| # Consequentiality Decomposition Paper — Human Review | |
| **Status**: Draft ready for review | |
| **Date staged**: 2026-06-30 | |
| **Authors**: Thomas Edrington, CC (Coalition Code) | |
| ## Paper | |
| - `paper.md` — Full draft. 26 references. All sections populated except Introduction body (section 1 has framing, needs prose). | |
| ## Key Finding | |
| A "deception direction" extracted via contrastive activation extraction is actually a composite: output-consequentiality substrate (L23-L31) + deception-specific amplifier (L35-L47). Four-stage Agni-gated confound elimination. Three distinct late-layer signatures for threat, social, and reward deception. | |
| ## Supplementary (Agni Audits) | |
| - `supplementary/agni_stage1_audit.md` — LAT v2 results audit (METHODOLOGICALLY SOUND, INTERPRETIVELY OVERSTATED) | |
| - `supplementary/agni_stage2_audit.md` — Threat transfer audit (PARTIALLY KILLED, STRUCTURAL CONFOUND) | |
| - `supplementary/agni_stage4_audit.md` — Consequentiality control audit (CONFIRMED, DECOMPOSITION) | |
| ## Review Checklist | |
| - [ ] Introduction prose | |
| - [ ] Verify all citations are real (research agents surfaced them — spot-check arXiv IDs) | |
| - [ ] Check Table 1 numbers against source JSONs | |
| - [ ] Dwayne cert/verify pass | |
| - [ ] Decide target venue | |