Every experiment, including the ones that failed.

Identity store 100% in every run. Gap over chunk retrieval: templated +42.4 (Gemma) / +38.9 (Qwen), natural text +51.4 / +48.6 points. Stale-belief errors 0% vs 64–69%.

Source: experiments/R01_replicate_M08_M01b/results/SUMMARY.txt

96.5% at 0.36% of a 2,410-file repository's tokens (mean of two models). BM25: 88–89%. Definition-only: 49.6%.

Source: experiments/R01_replicate_M08_M01b/results/SUMMARY.txt

Absent questions answered confidently wrong at 99,000 entities: similarity with a perfect key 48.3% (mean of 3 seeds), exact identity 0.0% (6 seeds).

Source: experiments/M11b_binding/results/eval_seed{0,1,2}.json, eval_seed{1,2}_i.json, eval_seed{3,4,5}_iii.json

One episode, 31 yes/no questions. 48-token summary: 39% confidently wrong. Full 337-token episode: 0%. Every absent-fact question had the answer “no”, which inflates the figure.

Source: history/RESULTS.md (P1)

Chance level: 0.045 vs 0.040 for the control. The facts were already known, so there was nothing to learn from.

Source: experiments/M04_cooccurrence_consolidation/results/, history/LEDGER.md

Accuracy 100% at 10 entities, 92.4% at 10,000, 70.3% at 99,000. The absence gate was calibrated at small scale and opens at large scale.

Source: history/LEDGER.md (M11 row)

+3.4 and +3.7 points over the model's own choices against a +10 requirement. A hand-written rule scored as high (88.3%).

Source: experiments/M13_calibrated_controller/results/, history/LEDGER.md

A 4B model could not write the programs: 13.8% of compile attempts were usable, and answer accuracy was 11.9%. Parser, token limit and sandbox were ruled out.

Source: experiments/B02_compilation_curve/results/diag_b02_qwen3.5-4b-hf*.{log,jsonl}

+1 point overall. Killed on its criterion.

Source: history/LEDGER.md