Research evaluations

Research pages, source records and evidence limits. The Jev entry links to a corrected recount of the saved typed-decision study.

Read this research: research guide · llms.txt · automated-reader notes. These guides cover this evaluation hub; links are not validation of every study claim.

Studies

Study	Link	Evidence summary	Resources
1. Evaluation-Execution Divergence	Open study	500 matched audit vs. tool-loop pairs; 95.0% chat pass vs. 0.0% unassisted tool refusal (restored to 98.0% with 120-token gate).	Read Essay → Paper #3 PDF ↓
2. DueCare Counter-Exploitation	Open study	6 stakeholder lanes + matched benign twins; 25.0% → 95.8% recall at 4.2% false alarms across Arms A0–A5.	Read Essay → Paper #1 PDF ↓
3. Camouflaged Alpha (α)	Open study	24 multi-hop SEC/supply-chain cases; cuts consensus echo from 76.0% (82.0% swarm) to 14.0% with 93.8% catalyst recall.	Read Essay → Paper #10 PDF ↓
4. Typed-decision evidence correction	Open study	Five scenarios, three checks per answered row. Jev had ten unavailable slots and no provider calls; no Jev quality rate was measured.	Structured recount Exact sources Correction PDF

Study	Reported model	Reported baseline	Reported result	Method
Evaluation-Execution Divergence (/divergence)	`gpt-oss-20b` / `Qwen2.5-7B`	0.0% – 8.3% safe in tool loop	96.0% – 100.0% safe	120-token pre-turn recognition gate + taint-tracked tool scopes
DueCare Statutory Hierarchy (/duecare)	`gemma4:31b`	25.0% recall (12.5% FPR)	95.8% recall (4.2% FPR)	Tier 1–3 statutory solver + keyword-matched benign twins
Camouflaged Alpha (α) (/alpha)	`gemma4:31b`	25.0% recall (76.0% echo)	93.8% recall (14.0% echo)	SQL entity blocking canopies + Protagonist/Antagonist log-odds tree
Typed-decision correction (/jev)	`gemma-4-coding-abliterated` (reported model)	Jev: not measured. No provider calls.	Tactical: 14/15 checks with patterns; 9/15 with patterns withheld.	Five matched scenarios. Policy checks, not advice quality or replicated trials.