Updated: 2026-03-05
HUSAI is a reliability-first SAE research project. It evaluates whether SAE features satisfy strict release criteria: internal consistency, stress robustness, and external benchmark competitiveness.
Result: pass_all=false. Internal and stress gates pass; external benchmarks do not meet strict thresholds.
EVIDENCE_STATUS.md-- what is locally verified vs remote-reportedEXECUTIVE_SUMMARY.md-- detailed status, gate outcomes, evidence pathspaper/sae_stability_paper.md-- the paper (PWMCC = random baseline finding)RUNBOOK.md-- how to reproduce everythingEXPERIMENT_LOG.md-- run-by-run history
HIGH_IMPACT_FOLLOWUPS_REPORT.md-- ranked next stepsNOVEL_CONTRIBUTIONS.md-- what is novel heredocs/04-Execution/EXPERIMENT_PLAN_2026_02_20.md-- experiment roadmapLIT_REVIEW.md-- literature and competitive landscapescripts/experiments/run_all_followup_experiments.sh-- run all 7 follow-up experiments (Section 4.11 of the paper)
pytest tests -q
make smoke