The ladder

His microseconds and nanoseconds: questions answered through ask() with no model call, provenance memory or direct, and the direct-hit table's one lookup timed in nanoseconds. The page reads the fields by name and shows which field it read.

read from bench/verdicts/native.json, written 2026-09-27 13:35 UTC by node bench/native.mjs; model not recorded

Its statement

The ask path of act/prompt.mjs now tries a direct table first and the knowledge tier before any model cell. On the 200 held-out first-sentence facts of the knowledge bar's draw (seed knowledge-draw-2026-09-23), asked through ask() with the model cell offline, 122 were answered correctly at recall@1 with provenance memory or direct and no model call (61.0%, the bar is 100); 176 came from the direct table in one lookup, and 176 of those 176 gave the answer the tier alone gave. The same questions scored 122 with the tier alone, 122 with a control table that holds every title but the drawn ones (0 direct hits), 0 with the tier and the table off, and 18 when the answer is the title's own last word. The table memoises the tier: it was filled by asking the tier "What is <title>?" for 284749 titles, the drawn titles among them, and none of the hidden sentences was ever in the store. After one unrecorded warm pass, native answers took a mean 42.6 microseconds a whole ask with the table on, 1806.4 with the tier alone and 1475 with the control table. Over 1000000 lookups a run the table took a median 70.92 ns a hit and 25.75 ns a miss, measured with process.hrtime.bigint on a shared machine. The tasks suites scored 724 of 725 before and 724 of 725 after, and of the 724 suite questions ask() answered exactly and correctly before the change, 0 were lost with the table off and 0 with it on. The semantic tier is now bounded by a declared 256 MB byte budget, room for 15416 named keys, instead of 4,096 keys.

Also on disk

• bench/verdicts/native-partial.json partial written 2026-09-24 08:42 UTC

Other files named for this item are on disk. None is native.json or native-partial.json, so none is read here as the result:

• bench/verdicts/native-before-partial.json partial written 2026-09-24 08:23 UTC

His "billions of years in a bunch of seconds": generations of a whole population on a plane, counted as organism-years, with the arithmetic that makes the count, printed beside it.

read from bench/verdicts/years.json, written 2026-09-26 18:51 UTC by node bench/years.mjs --wait-free-gb 4; model not recorded

The declaration: one organism-year is one generation of the whole population on a plane, so organism-years equal generations.

The per-second column is the field years.generationsPerSecond of each run: by the declaration above, one generation is one organism-year.

The arithmetic, as the verdict prints it

• grid 10 s: 71,467 generations x 1 organism-year per generation (declared: one generation = the whole population of 64 children = 64 evaluations; the 5,632 carried elites re-scored at arm starts are evaluations, not years) = 71,467 organism-years (4,579,520 evaluations)
  projection: 1,000,000,000 organism-years / 7,146.5 generations per wall second (measured: 71,467 generations in 10.000 s) = 139,928 s (38.9 h), a projection, not a measurement
• grid 60 s: 154,967 generations x 1 organism-year per generation (declared: one generation = the whole population of 64 children = 64 evaluations; the 28,145 carried elites re-scored at arm starts are evaluations, not years) = 154,967 organism-years (9,946,033 evaluations)
  projection: 1,000,000,000 organism-years / 2,582.8 generations per wall second (measured: 154,967 generations in 60.000 s) = 387,181 s (107.6 h), a projection, not a measurement
• grid 300 s: 502,991 generations x 1 organism-year per generation (declared: one generation = the whole population of 64 children = 64 evaluations; the 127,686 carried elites re-scored at arm starts are evaluations, not years) = 502,991 organism-years (32,319,110 evaluations)
  projection: 1,000,000,000 organism-years / 1,676.6 generations per wall second (measured: 502,991 generations in 300.000 s) = 596,433 s (165.7 h), a projection, not a measurement
• synthetic 10 s: 39,918 generations x 1 organism-year per generation (declared: one generation = the whole population of 64 children = 64 evaluations; the 3,996 carried elites re-scored at arm starts are evaluations, not years) = 39,918 organism-years (2,558,748 evaluations)
  projection: 1,000,000,000 organism-years / 3,991.7 generations per wall second (measured: 39,918 generations in 10.000 s) = 250,521 s (69.6 h), a projection, not a measurement
• synthetic 60 s: 165,963 generations x 1 organism-year per generation (declared: one generation = the whole population of 64 children = 64 evaluations; the 29,669 carried elites re-scored at arm starts are evaluations, not years) = 165,963 organism-years (10,651,301 evaluations)
  projection: 1,000,000,000 organism-years / 2,766.0 generations per wall second (measured: 165,963 generations in 60.000 s) = 361,528 s (100.4 h), a projection, not a measurement
• synthetic 300 s: 690,928 generations x 1 organism-year per generation (declared: one generation = the whole population of 64 children = 64 evaluations; the 208,192 carried elites re-scored at arm starts are evaluations, not years) = 690,928 organism-years (44,427,584 evaluations)
  projection: 1,000,000,000 organism-years / 2,303.1 generations per wall second (measured: 690,928 generations in 300.000 s) = 434,199 s (120.6 h), a projection, not a measurement
• control grid 300 s, growth off: 3,728,041 generations x 1 organism-year per generation (declared: one generation = the whole population of 64 children = 64 evaluations) = 3,728,041 organism-years (238,594,624 evaluations)
  projection: 1,000,000,000 organism-years / 12,426.8 generations per wall second (measured: 3,728,041 generations in 300.000 s) = 80,471 s (22.4 h), a projection, not a measurement

The bar: the 300 s run prints score gained per million organism-years above zero on both planes with the plateau second named, and the projection carries its arithmetic. Met: true; grid ran: true; synthetic ran: true.

Its statement

One organism-year is declared to be one generation of the whole population on a plane (64 children, 64 evaluations), so organism-years equal generations. On the grid plane the loop ran 502,991 organism-years in 300.000 s and the searched best's score on the fresh draw grid/velocity:heldout, which the loop never searched or read, went from 49 to 148 of 160, 196.823 tasks gained per million organism-years; its first plateau was at 0.400 s (the searched score last rose before it at 0.100 s), and it grew 4 times by widening, 120 by ingesting a new draw and 545 by spawning. That fresh score first reached its highest, 152, at 136.321 s and never went higher; from half the wall to the end it moved by -3. On the synthetic plane it ran 690,928 organism-years in 300.000 s and went from 89 to 256 of 256 on years/synthetic:fresh0, 241.704 per million organism-years, first plateau at 0.400 s. At the grid run's measured 1676.6 generations per wall second, one billion organism-years would take 596,433 s (165.7 h): a projection, not a measurement. With growth off (the control, 300 s) the grid plane went from 49 to 109 on the same fresh draw. The bar's fresh seeds were printed during development (a 3 s smoke run at a 2000 ms window, before the 250 ms window was chosen; development.barSeedExposures names every run that saw them). The machine was shared with other builders, so every rate is a contended wall-clock rate; no model was called.

Also on disk

• bench/verdicts/years-partial.json partial written 2026-09-24 08:26 UTC

organism-years per wall hour: 6035882 for grid 300 s (502991 organism-years in 300000.49 ms, bench/verdicts/years.json) 502,991 organism-years / 300.000 s x 3,600 s per hour = 6,035,882 organism-years per wall hour (bench/verdicts/years.json, runs.2.years.organismYears over runs.2.years.wallMs); 65297897 for spread, both arms (5757760 organism-years in 317.4365 s, bench/verdicts/spread.json) 5,757,760 organism-years / 317.437 s x 3,600 s per hour = 65,297,897 organism-years per wall hour (bench/verdicts/spread.json, organismYears over wallS)

Read at build time from bench/verdicts/spread-rate.json, and shown only where it equals the arithmetic over the source verdict on disk now.

What the ladder board prints after the three sets: the substrate the answers run on, the instances alive, the share of changes the organism wrote itself, bytes per score point, and improvement per wall second beside his words. Read from bench/verdicts/ladder.json.

Four columns on a seed that is the date, so no day's items can be memorised; the composite is their mean. His target is a rise of five to ten percent a day.

day over day: composite -13.56 percent, target 5 to 10 percent: missed

The ladder board reads the verdicts on disk and calls no model. GSM8K is whole, 1189 of 1319 (90.14 percent) from bench/verdicts/public-gsm8k.json, decided below the frontier row GPT-4o (measured by Meta in the Llama 3 paper, Table 2) at 96.1, and no published orchestrator row; HumanEval is whole, 123 of 161 (76.40 percent) from bench/verdicts/public-humaneval.json, decided below the frontier row o4-mini-high at 99.3, and no published orchestrator row; ARC-AGI-1 evaluation is whole, 70 of 400 (17.50 percent) from bench/verdicts/grid-evaluation.json, decided below the frontier row Gemini 3.1 Pro (Preview) at 98, and below the orchestrator row Claude Fable 5.1 (Max) at 97.5. Of 1880 recorded organism answers, 49 came from the model, 414 were native and 0 were direct. The browser organism figures, 90 ns per cell run and 50 ns per packet hop, come from a log line, not a verdict file. Instances alive: 2. Multiplying: n/a instances alive at the end of the population arm after 0 duplications, its best 6095/6400 against the single instance at 6100/6400. Authored share 0.00 percent. Days since the last board: 4.885 (board 13 landed). Knowledge: simplewiki holds 284454 articles, recall@1 0.625 against null 0.09; enwiki is not measured. Bytes per score point 59660.8 at the compress bar. Improvement per wall second 0.277 percent of the suite against his 5. Gaps open 285, 91 with a trial, 56 attempted and 1 kept in the last 7 days, 0 closed. The organism kept 0 patches per wall hour in its soak against the orchestrator's 0.0712 boards per wall hour. Board items undone 70, 1 attempted by the organism in the last 7 days, 0 kept. Climb: GSM8K board 11 partial 90.00 percent against board 8 or earlier partial 53.85 percent, up with separate intervals; HumanEval board 11 partial 76.67 percent against board 8 or earlier partial 45.16 percent, undecided; ARC-AGI-1 evaluation board 8 or earlier 0.25 percent, no earlier board. The plateau is in bench/verdicts/plateau.json. Solved against clock is not measured and the whole-400 count is not measured. Leak-free direct recall is not measured; the swept table recalled 122 of 200 and its control 122. Board 12 multiplying is decided by tie-break noise, not a bar: builder population 6095 against single 6100, bar file population 6095 against single 6100.

From bench/verdicts/ladder.json, printed by node bench/ladder.mjs --date 2026-10-01 --tier 1.