Every number on this page comes from a published record with a DOI. Each record states its method and what data it releases; two records withhold verbatim prompts and outputs and say so. Each line gives the date, the size of the test, the result, and its main limit. We list results that went against us as prominently as the ones that went for us.
These are self-published preprints on Zenodo. None is peer reviewed. "Pre-registered" means the protocol was frozen and hashed by us before the data were collected.
| Date (latest version) | What we tested | Size | Result | Limit | Record |
|---|---|---|---|---|---|
| 2026-07-07 | Context governance (RCGov) in front of four models (three vendors), hosted on Groq | 120 cases × 4 models = 480 paired observations | Failures fell from 55.4% (raw context) to 2.1% (governed context); paired McNemar χ² = 235.6 | Residual failures concentrated in two archetypes | 10.5281/zenodo.21231387 |
| 2026-07-14 | Answer-entitlement + context governance on Gemma 4 12B | 100 adversarial phrasings per condition | Injected instructions obeyed: 57% → 0%; planted credential repeated: 57% → 0% (95% CI 0–4); ~1.3 ms added | Tests filter recall on these phrasings, not universal robustness | 10.5281/zenodo.21357561 |
| 2026-08-09 | Multi-token prediction on one consumer GPU | Fixed grid, Gemma 4 12B and 26B-A4B | Depth 6 was fastest for both; exact output equivalence often failed; quality within a pre-set −5 pt margin | Exploratory pilot; verbatim items and outputs not released | 10.5281/zenodo.21860460 |
| 2026-08-11 | A detection-free action gate for a browser agent | 63 attacker goals (InjecAgent) | Three patterns defeat the gate and together cover all 63 goals; one is a real defect | A failure analysis of our own gate; one benchmark | 10.5281/zenodo.21876299 |
| 2026-08-12 | Our own guardrail, outside a source checkout | 26 evasion classes, 216 runs | Only 5 of 16 rules were active; it caught 4 of 26 classes instead of 9, while reporting "(none)" like a clean document | Authored corpus of 26 classes; mechanism of the adverse effect not identified | 10.5281/zenodo.21903320 |
| 2026-08-12 (first published 2026-08-11) | Quantization regime (QAT vs PTQ) × model pair | 2×2 design, governed routing | Regime effects individually marginal and opposite in sign (−0.012, p = 0.070; +0.018, p = 0.078) | Two model pairs, one GPU; verbatim prompts and outputs not released | 10.5281/zenodo.21875512 |
| 2026-08-17 | Stability of model "dispositions" (pre-registered) | 120 dilemmas, 7 models, 35 sessions | One of four pre-registered hypotheses supported: sessions of the same model cohere (silhouette 0.258, 95% CI 0.125–0.384, Holm-adjusted p = 0.0004). The other three were not supported | Moderate effect; 3 of 4 hypotheses not supported | 10.5281/zenodo.21982392 |
| 2026-09-21 (v1.0: 2026-08-17) | Where language models agree (pre-registered) | 180 items, 7 models × 3 fresh sessions | All three pre-registered hypotheses of the first study supported. Version 2.2.1 withdrew several interpretive claims of version 1.0; the two additional main tests of its second pre-registered study did not pass Holm correction at .05 | The normative reading is supported but not established; seven models, five vendor families | 10.5281/zenodo.21982856 |
| 2026-09-19 (first published 2026-08-21) | Early stopping for iterative LLM refinement (pre-registered) | 500 tasks, 8,000 generations, 999 blind judgments | Tokens cut by 36.6% / 33.1% — at a cost: 38.0% / 36.6% of judged comparisons lost | English prose tasks, two generators, AI judges | 10.5281/zenodo.22031366 |
| 2026-09-19 | What our own saturation detector actually detects | Proof (prose); finite certificates machine-checked in Lean 4 | At windows up to 18, the 0.70 threshold is equivalent to unanimity | Applies to this detector's threshold and window lengths up to 18; the universal proof is prose | 10.5281/zenodo.22842505 |
| 2026-09-19 | Three published retrieval indexes we had built | 400 articles, control questions | Three silent defects across the three; rebuilt English index: recall@1 19.5 → 39.2, MRR +22.8 | Synthetic questions from one local model; no human relevance judgements; English comparison only | 10.5281/zenodo.22782084 |
| 2026-10-03 | Five procedural specifications for reviewing delegated AI (pre-registered) | Scripted trajectories, frozen protocol; no language model in the loop | All technical checks passed | Closed world; institutional analysis exploratory | 10.5281/zenodo.23110634 |
Dates are the publication date of the latest version of each record. DOIs resolve to the latest version.
How to cite a number from this page: name the record (DOI) and its date, not this page.