Wins

Decisions settled by experiment, not taste

Weight tying and vocabulary size were decided by a 42-run factorial sweep: six cells (three vocabularies × tied or untied). Per cell, a learning-rate retune over a four-point grid on one seed (24 short runs), then the winning rate on three seeds (18 measurement runs); a separate companion ran one cell on each of the other cards as a hardware control. The noise floor was the widest per-cell spread, δ_seed 0.0076 bits per byte (spreads 0.0012 to 0.0076; δ_hardware 0.0026). Tying beat untied by 0.0090, 0.0121 and 0.0174 at 8k, 16k and 32k, every one above the floor and the margin growing with vocabulary, which is the interaction the factorial was designed to detect. 16k-tied and 32k-tied agreed to within 0.00003, a formal tie, and the smaller vocabulary won on cost by a rule I recorded before the sweep. A second, separate sweep priced epochs: costly, so single-epoch sizing stands. Each verdict became an architecture decision record before any training started.¹
v3 landed at 506M parameters on 4.8B tokens with 0.8575 bits per byte, the first Marilyn fed at roughly ten tokens per parameter instead of one and a half. Everything measured about her limits before this was untrustworthy; everything after is comparable.²
v5 sized from a study: composition needs depth and width, content precision needs tokens. That gave 32 layers at d_model 1,536, about 931M parameters on roughly 40B tokens, built as three variants so the comparison is between architectures.³

A promotion gate that is real and calibrated. Four Temporal workflows (gate, margin measurement, probe-set freeze, rollback). Its resolution was measured, not assumed: two frozen 24-document probe sets, one for the domain and one for retention, scored by a six-way paired bootstrap on a real checkpoint, giving δ_domain 0.0319 and δ_retention 0.0473 bits per byte. Calibration used real checkpoints from the v3 run, not perturbations: step 18,405 against step 4,000 (known better) promoted at 2.9× δ on the domain with retention improved; step 1,000 against step 18,405 (known worse) rejected at 7.5× and 6.7× δ on both clauses; and step 18,405 against step 10,000, whose difference of 0.0298 sits below δ, refused as unresolvable rather than promoted. The worker survived an eleven-minute outage with a run parked and completed it after the client died.⁴
v5 pre-training survived its own infrastructure. Heartbeat time-outs, a worker restart and a host reboot; recovered from step 126,000 on 2026-08-29 and resumed through the workflow, not by hand. ⁵
A sandbox for untrusted code, proven at the kernel level before any self-modifying experiment was allowed to run. ⁶
Contamination re-certification of the v5 corpus against every frozen probe set, with the four hits it found queued for me to rule on instead of silently passed.⁷

A calculator tool, taught from 4,000 demonstrations in about twenty minutes of micro-steps: well-formed tool calls appeared fast. On a held-out set of 48, the tool fired on 38 and the answer was correct on 20; the two numbers are form and competence, and the gap between them is the finding. Form is cheap; mapping problem family to recipe was half-learned.⁸
Execution reward beat imitation at a fifth of the retention cost. A client-side reinforcement pass with sandbox-graded rewards reached the same tool-family accuracy as three imitation passes while costing 2.3% on validation instead of 11.4%, a strictly better checkpoint by the gate’s own arithmetic.⁹
Comprehension is installable by supervised fine-tuning at a 133M branch, which turned “she is too small to understand” into a measured claim with a size attached.¹⁰
A policy installs; competence is capped. At rung 2 of the capability ladder the model reliably adopted the taught procedure while its accuracy stayed bounded by size, which is exactly the capacity-versus-misalignment split the charter wanted to measure.¹¹
Self-assessment demonstrated at rung 3: the model produced calibrated statements about its own answers on a reserved exam. ¹²

The shared trunk seats as a standalone model: intake admitted it 6 of 6 through the same workflows the nursery uses. ¹³
All six domain experts pre-trained on their own diets (code, math, science, web, Go, Python) and read-tuned to the trunk. The Go expert alone was 23,465 steps over 34 hours on 6.15B tokens. ¹⁴
The composition gate is built and compliant with the null-arm rule: a pass only counts against a wrong-expert arm and a shuffled read-out arm, so composition cannot be claimed by accident. It has not yet been run against a trained composition; when it is, the score against both null arms goes in this sentence.¹⁵

A weekly literature review that runs itself: nightly ingestion, a weekly timer, a verified-id gate, one reviewer per paper against a written brief. The first scan ranked 5,759 approved papers. ¹⁶
Mantis, at signal level. An untrained native-memory bolt-on showed a fact-specific trace of about one nat, at least four times the standard error across the 24-fact probe set, on each of two seeds, enough to price the trained version instead of guessing. See Reading that shaped it.¹⁷

Numbers on this page are quoted from the lab's own records. The records are private; each note gives the record's date.