"AGI" has become a word companies say about themselves. That is exactly why DeepMind's framework matters: "Measuring Progress Toward AGI: A Cognitive Framework" (Google DeepMind, March 2026; arXiv:2605.28405) replaces the word with a list — ten cognitive abilities, each of which can be measured, compared, and argued about in public. DeepMind themselves flag five of the ten as the ones the field can barely measure at all: learning, metacognition, attention, executive function, and social cognition. Keep that list in mind as you read the chart below — our strongest green sits squarely on it. We think the right response to a published yardstick is to lie down on it. So this note is our self-measurement: what it does, not how — and where it doesn't, said plainly.
Where the evidence is strong
Reasoning. On depth-11 Tower of Hanoi — where every frontier model we tested scores zero across twenty attempts — our committee search went 100-for-100, and held 20-for-20 out to depth 14 with every completed solution exactly optimal. A second GoM matches the published BDH solver's accuracy (96.9% on the hardest 17-clue puzzles) at a fraction of its size. These are verified-search results: the answers are checked, not judged.
Learning. When the rules of a task silently change, our system recovers in a handful of trials; the frontier models we tested needed several times as many. One-shot recall of taught facts lands at 99–100% against a 17.5% baseline. And a single round of self-play measurably improved Sudoku with no retraining. Learning is one of the five abilities DeepMind flags as hardest for the field to measure — it is where our numbers are cleanest.
Problem solving. One recipe, no per-task tuning, across Sudoku, Hanoi, Sokoban, and Rubik's Cube — including domains where a committee solves what no individual member can. And at production scale: composing our selection layer around NVIDIA's frozen open 120B Nemotron Super on SWE-bench Verified — real GitHub bug-fixing — lifted the resolve rate from 42.3% to 49.0% on 149 paired instances, ten instances fixed and none broken, exact p = 0.002 against a pre-registered bar. This week we closed the scale-up's control arm: 52.5% single-shot on the 550-billion-parameter Nemotron Ultra flagship, all 440 instances graded, zero exclusions.
Where we are partway — including a result that went backwards
Metacognition. The system can decline to answer rather than confabulate — measured, not asserted — and when it fails, it can say why (out of budget vs. out of ideas). What it cannot yet do is emit a calibrated confidence number per answer. Partial credit, honestly labeled.
Memory. The most instructive row on the card. We diagnosed our episodic memory's retrieval as functionally inert, fixed it — retrieval concentration improved a thousandfold — and the downstream solve rate went down. The retrieval was fixed; the memories being retrieved were sludge from failed search branches. The lesson (only verified experience deserves to be remembered) now drives the next iteration. We publish that sequence because a scorecard that only reports wins is marketing.
Generation, attention, executive function, perception. Each has a real mechanism and a real number behind it — a +32.5-point jump on grade-school math when our bridge is attached to a frozen Qwen3.5-27B model; adaptive compute that spends 2 passes on easy queries and 200+ on hard ones; committee control with explicit budgets; a perception campaign where object-level decomposition decisively beats whole-grid search on abstract-reasoning grids. None of these is finished. All of them are underway with logged baselines to beat.
Where we claim nothing
Social cognition — modeling other minds, responding appropriately in social situations. We have measured nothing here, so we claim nothing. It is the weakest row on our card and will stay weakest for a while. We would rather show you an honest zero than a decorated one.
Two levers already in flight
Sleep learning. The memory lesson above pointed somewhere specific: only verified experience deserves to be remembered. That principle is now a running experiment — solved episodes written back into a frozen model's input space, no weights touched, against a pre-registered bar, with a generic-prefix control arm so the episodic claim has to beat a placebo. It reports on its own schedule, published either way. If it passes, the memory row above earns its first write-path receipt; if it fails, you will read that here too. Either way, the loop it belongs to — select what works by day, consolidate it by night — is how the amber rows are designed to turn green: with receipts, not adjectives.
The collective pathway. DeepMind's companion work names four routes beyond single-system ability; one of them is multi-agent collectives — which is precisely the bet behind our swarm thesis (Note 5) and the committee results threaded through this card. We hold the label to the same standard as everything else: nothing swarm-shaped has been measured as a faculty yet, so it stays roadmap — but it is also the honest route into our red row, because social cognition starts with agents modeling each other. The composition results in Note 15 are the first rung of that ladder.
Why publish a card with holes in it
Because the alternative is worse. When "AGI" is a self-awarded title, every claim inflates until none are believable. A published rubric plus logged evidence is the only version of this conversation that ages well. Our strongest rows sit exactly where DeepMind says the field struggles most to measure — learning above all — and our weakest rows are printed in the same font size. Claim what is measured. Scale what is roadmap. Never retrofit.
Every number in this note comes from a logged run; qualified partners can request the full evidence trail — raw logs, harnesses, and the pre-registered bars — under NDA.