Playable brain: deterministic model, real spiking network, per-sentence math, calibration (which caught an inverted trust score)

[slavazeph-coder]

Summary

Turns the arcade's brain from decoration into an instrument, and turns its numbers from asserted into calculated. Along the way the new rigor caught a significant scoring defect and two simulation bugs.

201 unit tests · 19 Playwright e2e · lint and build green.

The headline finding

Calibrating against a labelled corpus revealed trust was ranked backwards (Spearman −0.505): the engine scored outrage bait as more trustworthy than a sincere apology. trust counted trust vocabulary, not evidence — "share this because once it's gone" earned +13 for the bare connective, while "on Tuesday our update broke checkout for six hours, here is what we refunded" earned nothing.

Fixing it was evidence-driven, and one hypothesis was rejected by the data: removing "because" made things worse (−0.611), because it's sometimes a genuine evidence marker. Reverted; added specificity and stated-limitation signals instead.

dimension	before	after
trust	−0.505	+0.628
manipulation risk	0.387	0.629
viral pull	0.333	0.366
pair accuracy	65.8%	80.1%

What shipped

Deterministic brain — the sim ran on Math.random(), so the same text produced a different brain every time and no score was reproducible. Now a pure, seeded, parameterized model with interventions (lesion / cut / stimulate) and headless trials. New readouts: firing rates in Hz, PFC/AMY control ratio, hijack index, control-loop gain, E/I balance, spike correlation, settling time, STDP flux. It previously had no tests at all.

A real spiking network — despite the name, BrainSNN was a rate model. Added a proper LIF network in the Brunel (2000) formulation (Dale's law, delta synapses, axonal delays, Poisson drive) with standard measurements — CV of ISI, Fano factor, population spectrum. Validated against published regime behaviour: silent below the threshold drive, rate monotonic in drive, rate falling and irregularity rising with inhibition, near clock-regular firing when excitation dominates. Gamma-band power becomes measured rather than asserted.

Per-sentence math — leave-one-sentence-out attribution, with the same replicates read as a jackknife for genuine error bars replacing the asserted confidence. Runs inline (~27 ms).

Defend the Brain — the arcade's first machine-checked mission. All twelve existing labs print a mission nothing verifies; this one has a real win condition, fail state and budget. Thresholds were measured: doing nothing breaches for ~179 ticks, silencing threat or boosting judgment holds, and cutting the salience route helps but isn't enough — a real difficulty gradient.

Evidence layer — Krippendorff's alpha for inter-annotator agreement; AUC/Brier/ECE metrics; an external-corpus eval script; an annotation rubric; and verifiable run proofs where a score is checked by recomputation — inflating it requires finding a log that replays to it, not editing a number.

Bugs found by the new rigor

1. Zero-delay spike transmission — the LIF ring buffer wrote into the slot it was reading, so spikes reached later-indexed neurons instantly instead of after 1.5 ms. With correct delays the balanced network settles at 42 Hz instead of 69 Hz and the sweeps become cleanly monotonic.
2. Mission loss didn't latch — a run that failed mid-way but recovered scored as a win.
3. Divergent control ratio — a lesioned amygdala drove the denominator to ~0 (21.73 on screen). Capped.

Honesty notes

• The brain is driven by the lexical scores, so its dynamics add structure but no new information about the text. Every readout carries a claim boundary: simulated dynamics of a 7-region model, not a measurement of any human brain.
• Two limitations are pinned by tests rather than hidden: "verify within 24 hours" still reads as concrete detail (phishing over-scores on trust), and spike probability can't reach its own cap because activity is capped at 1.0.
• Brunel's asynchronous-irregular regime needs N ~ 10⁴–10⁵ for presynaptic independence, so the tests assert monotonic trends rather than claiming all four published regimes.
• A six-item corpus smoke run gives AUC 0.778 but ECE 0.290 — the scores rank reasonably but don't yet behave like percentages.

Not done, and why

Durable cross-user leaderboard storage: Supabase is wired for magic-link auth only — no tables, no persistence, no rate limiting. The verification logic that would make such a leaderboard trustworthy ships here; the storage is separate infrastructure. Public corpora are not vendored (licensing) — loader, rubric and metrics ship, data is fetched locally.

🤖 Generated with Claude Code

https://claude.ai/code/session_01725FqR5KopKJJXN3RTfLpD

[gemini-code-assist]

Code Review

This pull request introduces 'Mind-Hack Autopsy' (experiment 013), a new client-side interactive lab allowing users to simulate brain reactions to various content types locally. It updates the arcade layout, landing page, sharing dialogs, and E2E tests to support thirteen experiments, and adds a new evidence export feature for the attractor playground. The review feedback suggests several robustness and compatibility improvements, including defensive checks for clipboard API availability, safer destructuring of parameters in the evidence builder, and appending the temporary download link to the document body to ensure cross-browser compatibility.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.