GoM · Recorded — real logged runs, not live

The rule changed and nobody said a word.

A hidden rule pays off — then, mid-run, it silently flips. No announcement, no new prompt. The only signal either system gets is "right" or "wrong" after each guess. GoM's online learner detects the flip from that feedback alone and re-locks in about six trials — the frontier takes ~30. A frozen frontier model can't update its weights from in-session feedback — so it keeps playing the old rule, and the cost (regret) climbs.

The task — 45 shared trials. Each trial shows two numbers, a and b; the right answer is (a + b) mod M — and M is hidden. For the first 15 trials M = 10; at trial 15 it silently becomes 7. After each guess the only feedback is "right" or "wrong." Every item is chosen so the two rules disagree — (a+b) mod 10 ≠ (a+b) mod 7 on every trial — so the feedback genuinely identifies the rule, and both systems face the identical 45-trial stream. GoM keeps a running guess about the hidden rule and updates it after every right/wrong (a Bayesian belief with a change-point prior, for the technical reader); the frontier model receives the full trial history in its prompt. “Recovered” = 3 consecutive correct answers after the switch.

Trial 0

GoM regret 0

Frontier regret 0

GoM — online learner (updates per trial) Silent switch — rule flips, no warning

What you're looking at: cumulative regret (running count of wrong answers) vs trial, for one shared schedule both systems faced. Lower is better; a flat line means the system is getting it right. At the dashed violet line the hidden rule silently changes. GoM's curve bends flat again within a few trials (it re-learned); the frozen model's curve keeps climbing (it's still answering the old rule). Every point is a real logged trial — adapt_data.json (shipped with this page).

Frozen frontier model · after the switch

— trials

Still playing the old rule. A frozen model can't update its weights from win/loss feedback — its only channel is the prompt, and here that wasn't enough to recover in-window.

→

GoM · after the switch

— trials to recover

Detected the flip from feedback alone. The belief over the hidden rule shifts after a short run of errors, its best-guess rule flips, and it re-converges — genuine in-session adaptation.

Trials to recover after the silent switch

real logged run

Honest framing. A 2026 reversal-learning study (arXiv:2604.04182) finds frontier models — GPT-5.2, Gemini-3, DeepSeek — perseverate for ~20 trials after a silent reversal vs ~3 for humans, underweighting losses. In the published study, DeepSeek was the slowest to adapt; in our own logged runs GPT-5.5 (30 trials) and DeepSeek (28) were comparably slow — we chart and report both. The load-bearing claim is not "slow learning" — it is structural: a frozen model cannot update its weights from in-session win/loss feedback. Its only adaptation channel is the in-context prompt; any genuine, persistent weight-level adaptation is a categorical capability GoM has and a frozen model does not.

In production, rules change without an announcement — pricing, policy, intent. The system that re-locks in six trials is the one you can leave running.

The rule changed and nobody said a word.

Trials to recover after the silent switch

See the rest of the deck.