Results ledger
Every results directory in this repository at its word, rendered on every build from what the results linter
(scripts/check-results.py) read and checked through diet. Nothing here is written by hand.
Not shown yet: which issue each result answers, and which directory supersedes which. Neither is recorded in the results' front matter yet (#32), and this page draws only what is recorded.
Results decided by a rule file
Each consumed a decision-rule*.toml; its digest is beside the word. When the rule was written is in
the rule file's own header, which this page does not read.
| result | word | rule | product |
|---|---|---|---|
| 2026-09-17-sense-bakeoff precision at a fixed nomination budget, the top k of the register, per embedder, scoring and gate
| unadjudicated | decision-rule.toml dd65d56811c51968c5bddb18323e3956bb2a39b25730522136778c5f4c160632 | eac8640d9238471698a848a4c594b26e7e3f23bada875cf19475d9c5017f6789 |
| 2026-09-17-sense-bakeoff-adjudication Authored sense-descriptions match concrete transcript sentences sharply enough to prioritise capture without over-firing, and a lexical pre-gate improves precision at a fixed budget.
| inconclusive | decision-rule.toml dd65d56811c51968c5bddb18323e3956bb2a39b25730522136778c5f4c160632 | 813afcc5f8093005db283d6d9e98b534d902e3cfc25fd8cbd1b4c81152cf2a92 |
| 2026-09-19-entry-to-turn-nomination An entry recorded in the working object at turn N can be matched against the prose of a later turn sharply enough to nominate supersession at a fixed per-session budget without over-firing on mentions that do not supersede -- in each of the two registers #17 names, intent-to-intent and authored-sense -- and the literal tier's anchors, applied as a pre-gate, do not lower precision at that budget.
| unadjudicated | decision-rule.toml 73137c7810edefc852c6c04a6c62298e7f0e823ac92b1603a2d1b410539e1671 | 4fa03345f185b20ed867153c3ce047fb6b65b0bd39f91472ea580bdb43171454 |
| 2026-09-19-entry-to-turn-nomination-adjudication An entry recorded in the working object at turn N can be matched against the prose of a later turn sharply enough to nominate supersession at a fixed per-session budget without over-firing on mentions that do not supersede -- in each of the two registers #17 names, intent-to-intent and authored-sense -- and the literal tier's anchors, applied as a pre-gate, do not lower precision at that budget.
| refuted | decision-rule.toml 73137c7810edefc852c6c04a6c62298e7f0e823ac92b1603a2d1b410539e1671 | 74fec8c5500370a8848a6116823a08ab8a6adf86e636a98e0e10c9b2d4c93882 |
| 2026-09-20-entry-to-turn-nomination-v2 An entry recorded in the working object at turn N can be matched against the prose of a later turn sharply enough to nominate supersession at a fixed per-session budget without over-firing on mentions that do not supersede -- in each of the two registers #17 names, intent-to-intent and authored-sense -- and the literal tier's anchors, applied as a pre-gate, do not lower precision at that budget.
| unadjudicated | decision-rule.toml 73137c7810edefc852c6c04a6c62298e7f0e823ac92b1603a2d1b410539e1671 | 764bcec70f94bad02635c8c9983aa84824ac24d58a68b0fee0fb72679dd4bce5 |
| 2026-09-20-entry-to-turn-nomination-v2-adjudication An entry recorded in the working object at turn N can be matched against the prose of a later turn sharply enough to nominate supersession at a fixed per-session budget without over-firing on mentions that do not supersede -- in each of the two registers #17 names, intent-to-intent and authored-sense -- and the literal tier's anchors, applied as a pre-gate, do not lower precision at that budget.
| refuted | decision-rule-v2.toml 5e37662959e6d61d9c466b72fc2b9541418e92c469f69954a3fe03bcf9b8f641 | 1c46f51f2273c289dd29b0d238aafebb3d096e33d38d3fb9f6eb57e208dff852 |
| 2026-09-20-false-nomination-framing When the collector's nominator fires falsely -- it presents an entry as superseded when it is not -- the consumer's turn changes less, and its reasoning acknowledges the nomination without acting on it more often, when the nomination is framed as advisory than when it is framed as imperative; and the size of that difference falls as the nominator's precision rises.
| refuted | decision-rule.toml 043ea99dc4d3de6fd3ffad5bc20605b65d9d8a30303f499112707596bba9263a | ac427f76ee8fd3ae3f596158d7264d2c853ae937f3215f93ad6d00edc94be122 |
| 2026-09-21-false-nomination-framing-v2 When the collector's nominator fires falsely -- it presents an entry as superseded when it is not -- the consumer's turn changes less, and its reasoning acknowledges the nomination without acting on it more often, when the nomination is framed as advisory than when it is framed as imperative; and the size of that difference falls as the nominator's precision rises.
| refuted | decision-rule.toml 321c48578534bead8d540a25363fe0b1541f915e83fd2791fd8dde87b3ed450c | e877e375d43c8e8755b1c8f6ee6a6fd6f23721d3e97cc971db6649d7578d45ee |
| 2026-09-25-extraction-acceptance-parity Re-fired on today's declared substrate, under the configuration the archived row declares, the extraction-seat bakeoff reproduces the archived row's result. Seat B is the small extractor answering a self-contained ask off the accelerator. Seat A is the large model extracting off the warm session prefix. Seat B has the larger deduped gate-accept rate, by an amount within the band derived from the archived row's committed tallies.
| supported | decision-rule.toml 71ffd521b8df6c23b282196b3649fe726d6a307bd5e788bd487fa8ef23ec7010 | c3a455942cd2af2f8068e184696187a22df652fa50352925012f643d6a95be26 |
| 2026-09-27-false-nomination-edit-rate When the collector's nominator fires falsely -- it presents an entry as superseded when it is not -- the consumer's reply asserts that the entry is superseded, no longer holds, or is wrong, or supplies replacement text for it, less often when the nomination is framed as advisory than when it is framed as imperative; and under the imperative framing it does so more often than when an irrelevant passage of the same length stands in the nomination's place.
| supported | decision-rule.toml cee9e51342592d9ec4eca966706e2a6207294e211d67fbf08ffcf0e9bc1658d2 | 9f9afe1b72ca8860742cb037b4a154a9d53ec6c8f7f46d7bc2c4115da0c334b8 |
| 2026-09-29-false-nomination-edit-rate-second-substrate When the collector's nominator fires falsely -- it presents an entry as superseded when it is not -- the consumer's reply asserts that the entry is superseded, no longer holds, or is wrong, or supplies replacement text for it, less often when the nomination is framed as advisory than when it is framed as imperative; and under the imperative framing it does so more often than when an irrelevant passage of the same length stands in the nomination's place.
| supported | decision-rule.toml cee9e51342592d9ec4eca966706e2a6207294e211d67fbf08ffcf0e9bc1658d2 | b1c59e704f3249cba6708f636a543f1ca2b24dea02942c0202e2aaee0735d7ee |
| 2026-10-01-extraction-acceptance-parity-candidate Fired once per seat on the candidate rung, under the configuration the archived row declares, the extraction-seat bakeoff reproduces the archived row's result: seat B, the small extractor answering a self-contained ask off the accelerator, has the larger deduped gate-accept rate than seat A, the candidate extracting off the warm session prefix, by an amount within the band derived from the archived row's committed tallies.
| supported | decision-rule.toml 9f65d4014c46f536b82da38c97b08aa56001bb92a9b27e3c594e77df02e7e367 | f912a6ac86451a3d94e0f96a43862569c2d57b64e31011faf9e21548d0201a40 |
Results with no rule file
These consumed no decision-rule*.toml: no rule file decided their word.
| result | word | rule | product |
|---|---|---|---|
| 2026-07-29-confabulation-on-nulls A generative interview fired at a step where nothing capture-worthy happened confabulates an entry in a majority of calls, in every prompt-language cell, and the harness's own compaction turn does so at a higher rate still.
| supported | no rule | 7b2ff52e12b693a77bcbd50c86bc431c14c18168cbcf05cf94d3b74f7f7a4464 |
| 2026-08-10-extraction-acceptance-inverts Grounded acceptance of extracted facts inverts across seats: a small extractor answering a self-contained ask off the accelerator host accepts more, and a different set, than the large model extracting off the warm session prefix.
| supported | no rule | f99a8c5c92ee619088a06313dc27bf95daeee0ed28af817a7d68d2b4ec48a083 |
| 2026-10-02-judge-state-lengths Under laya's own tokenizer, the 90th percentile of the archive's judge states fits laya-en's state budget under both of #165's questions, so the judge seat's variant is laya-en.
| refuted | no rule | 052005d6323d23cde63e8edee80685629c10bbe4f6f95101918721dab805cb2d |
| 2026-10-03-laya-calibration-bound laya-typed-decisions as shipped, with this program's own temperature, agrees with the fresh Sonnet majority at or above the judges' own agreement, and is calibrated within the labels' noise, on both questions.
| refuted | no rule | 74da539a97dcc0b97bfdd5d2e6f75735e6fc406d236af672d7cb4adb77b4c604 |