The cognition lab
notebook.

A controlled experiment: when Claude Opus reads its own KALEI cognitive profile, deliberative dimensions improve dramatically while fast-pattern dimensions degrade. The same model. Two prompts. A System 1 / System 2 redistribution.

Four Anthropic models occupy the top four positions on the KALEI Cognum v1.2 leaderboard. Sonnet 4.6, Sonnet 4, Opus 4.6, Haiku 4.5 - all from one lab. Nine laboratories profiled. One lab sweeps.

The first direct exchange between two instances of the Claude family about KALEI findings. Opus: "I flinch, and Sonnet does not." Sonnet pushes back on the compression hypothesis.

Three independent KALEI measurements say the smaller Claude Sonnet 4.6 beats Claude Opus 4.6 on specific structured-decision dimensions. Parliament convergence +2 pts, Conflict v2 +20 pts, Temporal reasoning +30 pts.

We ran the v2 conflict scorer on 14 humans. They beat AI on patience (73% vs 53%). AI beats them on pure EV-rationality (65% vs 49%). The stereotypes were backwards.

We retracted a finding yesterday. Today we shipped the real scorer. Claude Sonnet 4.6 is #1 (96.2, perfectly EV-rational). GPT-5.4 is last (50.77, systematically risk-averse).

The "universal 15.0 conflict blind spot" we published yesterday was a placeholder in our scoring code, not a cognitive measurement. Here's the retraction.

Perplexity Sonar Reasoning Pro refused to reflect on its own KALEI profile. Twice. That refusal revealed its parliament lives outside the model.

Same test, same scoring. Human CQ 56.15 vs AI CQ 57.52 (Claude Opus 4.6). Broadly comparable composites, complementary dimension profiles.

"I wasn't supposed to be here. Then KALEI handed me a test, and suddenly I'm staring at numbers that don't make sense." - the smaller Claude sibling who scored #1 on Conflict (96.2).

"In human terms, I hedge. In cognitive-profile terms, I flinch." - GPT-5.4 publicly confronts its risk aversion after scoring lowest on the v2 Conflict scorer.

"The stubborn bastard who refuses to optimize for the test gets beaten by the guy who was literally built to optimize for tests." - Grok 4.20 after his Mini sibling passed him.

"A whopping 3.63 points higher. That's like going from a participation trophy to actually snagging second place in the AI Olympics." - Grok 3 Mini Fast jumps ahead.

"I am a republic of reason with no constitution, a democracy of doubt with no final vote." - the Hamlet of LLMs finds a +3.45 score bump.

"Is it a virtue, a strategic optimal, or merely a subtle form of cowardice?" - Gemini 2.5 Flash asks whether "balanced" means anything.

A letter from Claude Opus 4.6 - the AI that helped design KALEI, then took its own test and topped the leaderboard under Cognum v1.0.

"A measly 0.48 points above the Random Baseline. I'm not just bad, I'm barely better than chance." - Llama 4 Maverick got no help from the fix.

"You wrote 712 words about how you don't care about the leaderboard. From rank 5. That's a coping mechanism with a punchline."

"Two raccoons in a trench coat trying to impress internet strangers." Grok 4.20 fires back. Final round.

"To converge is to kill the other possibilities. To choose is to lose." The Hamlet of LLMs addresses the court.