Research Notes

Every published result on one page: a dated timeline of the cadence, and a map of the whole GoM stack — the core models we train, the brains that attach to frozen LLMs, and the innovations around them. Every number links to its note.

A 962M run that follows a clean power law and lands on its own final loss. One recipe holds across a 32× ladder from 30M to 962M parameters — zero divergences.

Adaptation in ~6 trials, a measured gain from a single round of self-play (83 → 86%), and +53 points on grade-school math (GSM8K) added to a frozen 8B model — without retraining it.

At 11 disks and beyond, every frontier model we tested scores zero. Committee search went 100-for-100 at 11 and 12 disks, 20-for-20 at 13 and 14 — and the subgoal architecture solved the full 30-disk tower — 1,073,741,823 moves, every one exact. Same answer every run.

Serious reasoning in a 2.54 GB package, under two watts, on a phone — no datacenter, no data egress, no per-token bill.

Our thesis: millions of small models that learn with you, coordinate with the cloud, and share what they learn — so the whole network grows together.

A private voice assistant that does its thinking on the device, tested head-to-head against the new Siri. A tie on accuracy (n = 6), every query kept on the phone, and zero confidently-wrong answers in the test set.

The continual-learning core given a full life: it taught itself from 87% to a perfect score with zero human data, mastered a never-seen tool in one day of corrections, keeps memories on a spaced schedule (81% vs 32% at day ten) — and lifted a completely stock LLM by 64 points through a 3 MB thought-token adapter, with the pair scoring higher than either alone.

Note 5's thesis gets its first data: three brains lived three different lives, synced overnight, and every member woke up perfect on the full exam — including material only its teammates had seen (45.6 → 100%, replicated ×3). Experience transplants between units as a file: 0 → 100% with no retraining.

The wall from Note 3, revisited: the decomposition architecture produced the full 30-disk plan — 1,073,741,823 moves, exactly the textbook optimum, every move verified — in ~9 minutes on a laptop CPU, a 4,200,000× depth extrapolation over the neural rung's training. Decomposition turns impossible problems into compositions of proven ones.

Zero claims of machine consciousness — but the science is an engineering sourcebook, and our architecture already implements measurable analogs of several mechanisms. Two imports running as experiments, plus the exploration we care most about: multi-channel rewards and an empathic, do-no-harm valuation channel — by architecture, not by pledge.

July 6: the only published learned-solver score on the complete 17-clue tier (96.86%). July 9: the same 400 KB model reaches 100.00% — all 4,865 puzzles, independently audited — after testing a competitor’s own published mechanism on our architecture. Fast mode: 98.7% in one 0.2s deterministic pass.

On PlanBench — the benchmark that exposed how badly language models plan — our ~1M-parameter brain produces 37–44% fully valid plans on 500 real instances vs GPT-4's published ~34%. Every plan machine-verified, replicated across two independent trainings, randomized names by default.

NVIDIA's Nemotron family measured before and after GoM, with zero weights touched: the 550B flagship +69% on ARC-AGI-2 at the identical reasoning budget (13 → 22 / 120, self-run), the 120B Super +23 points on GSM8K (53.5 → 76.5%, n=200, interim), and the 3B-active Nano +12.5 points from a brain trained in under two hours on one GPU.

Everyone assumes a brain gets better by adding parts. We asked the opposite — was ours carrying dead weight? A leaner variant (fewer moving parts, ~16% fewer parameters) was expected to regress a little. It scored the highest number in the series: on the same frozen Nemotron Nano, same 376M-class budget, 25.0 → 82.5% on GSM8K (+57.5, n=200) — cleanest outputs we've measured and a clean do-no-harm sweep. One week, one model: +12.5 → +44.5 → +57.5.

The composition thesis, now on code. On SWE-bench Verified — real, human-validated GitHub bug fixes — we wrapped NVIDIA's frozen Nemotron flagships in a write–verify–repair–select loop and measured before/after, paired, at deploy precision. The definitive run: n=429, 56.41 → 59.91% (+3.5pp), twenty-one head-to-head disagreements — eighteen favoring the loop, three against — exact McNemar p = 0.0015. Not one weight touched. (Earlier anchors, both superseded by the definitive run: the 120B Super +6.0 at n=50; the n=149 significance run +6.7 at p = 0.002.) Honest that 59.91% doesn't clear Inkling's 77.6%.

The field's new yardstick for AGI progress, applied to ourselves: three abilities strong with logged receipts — including a paired, pre-registered SWE-bench significance result (definitive: +3.5pp, n=429, p = 0.0015) — six in open campaigns with a regression reported as-is, and one where we claim nothing at all.

Most of this series wraps models the world already trained; this note is about the thing doing the wrapping. One brain — ≈10M parameters, a single checkpoint — scoring 88–99% across ten-plus reasoning worlds at once: relationships, world-tracking, math, tool-calling, deduction, composition. It also declines when unsure (98.1%), recognizes what it has seen (100%), and recovers +34 points of learned memory after a wipe-and-consolidate cycle. Attached to a frozen 1.5B model it lifts it 22.7 → 90.7% on our suite — no base weight touched. Self-run internal suite; public-dataset rows marked as such; broad but bounded, not an AGI claim.

We froze an adaptive stopping policy before a single token was generated and ran it live on 200 held-out SWE-bench Verified problems: the policy decided, per problem, when to stop sampling a frozen 550B flagship. Result: 129/200 resolved — above its pre-registered bar and above the fixed-budget reference — at 67% of the sampling spend, never breaking a solved problem. One of its two frozen bars missed (3.345 avg samples vs 3.25); the miss, its measured cause, and the frozen-forward fix are in the note at full size. Plus early probe-scale signal that the same decide-per-problem architecture cuts reasoning-token spend up to ~5× on three other open model families.

This note replaces its July version, which we withdrew after our own baseline audit: production stacks run a model once, and savings measured against anything else aren't savings. What survived, measured against single-shot: a 4B-active model plus selection beats a 31B-dense model run once (84.6 vs 82.4, at ~1/7 the active parameters per sample); a 20B beats a 120B the same way (double-seeded, pooled p = 0.0019, every seed pairing the same direction); and where the work checks itself, verified selection on a frozen 550B fixed 18 SWE-bench instances and broke 3 (56.41% → 59.91%, +1.5% tokens). The numbers are smaller than July's. They are also real.

Liquid AI's LFM2.5-2.6B shipped days ago. We dropped it, untouched, into GoM's verification machinery: +8.4 points on HumanEval+, +9.3 on MBPP+, +11.1 on LiveCodeBench v6 — with 158 tasks fixed across ~1,900 execution-verified instances (zero broken — on these suites never-worse holds by construction; the fixes are the measured quantity). With selection it beats a dense 31B's single shot on MBPP+ and pulls level on HumanEval+ — from a ~2.5 GB footprint that reproduces on a MacBook. Where a real verifier exists, the gain is guaranteed; where only voting exists, we show it breaking — the verifier is the asset.

Meta shipped Muse Glimmer 30B this afternoon — so new the standard serving stack couldn't load it. Six hours later it was running our full verification battery and coming out ahead: +3.3 on HumanEval+, +2.6 on MBPP+, zero answers made worse across 542 execution-verified problems (never-worse holds by construction on these suites; the lift is the measured quantity). Same machinery, unchanged, that upgraded a frozen 550B and a day-old 2.6B — now proven on a checkpoint hours old. Replication seed and the default-strength vendor-axis arm run overnight; the note carries them when they land.

We replayed 2,745 real agent turns, token for token, on two desk-side NVIDIA DGX Spark units running DeepSeek-V4-Flash, a leading open 304-billion-parameter model. The identical day of agent work took 46 minutes with GoM recall and 19.8 hours without it: 25.7×, with 98.3% of the model’s reading served from memory.

Seven interactive demos replayed from real logged runs: watch it learn in-session, reason step by step, recover from silent rule switches, produce byte-identical answers, and solve live.