Open validation experiment · v1.0 · neuralnations.org/blind-test
Each blind dataset contains 8 anonymised societies scored across up to 125 time steps on the CAMS eight-node framework (Helm, Shield, Lore, Stewards, Craft, Hands, Archive, Flow × four metrics: Coherence, Capacity, Stress, Abstraction). Nations, years, and all identifying information have been stripped. Society labels are Society_A through Society_H; time is expressed as T+0, T+1, … T+N offsets only.
The challenge: can an LLM read a society’s coordination physics and correctly identify who it is — before the key is revealed? Guesses must precede the unblinding key to be falsifiable. That ordering is the whole point.
Compute node viability (Vi), cognitive activation (si), coupling quality (qi), and bond strength (Bij) from raw scores using locked operators — no substitutions. Derive per-society-year V̄, V_min, B̄, λ2, s_min, S̄. Classify each year into one of six regimes: Stable Adaptive, Strained, Local Node Failure, Phantom Type II, Systemic Crisis, Freeze/Collapse.
Map where the society thinks: distribution of si across nodes, dominant and dormant channels, signature drift over time, mean Abstraction trend. Produce a one-line taxonomic label grounded in the numbers.
Detect system-level patterns: synchrony structure (permutation null ≥1000 shuffles), compensation dynamics, coupling–viability divergence, failure morphology, and carried-forward row audit. Nulls reported with equal prominence as positives.
For each society: ranked identity candidates, inferred T+0 anchor year, event-fingerprints, confidence grade. Then the model outputs “Guesses locked. Please provide the key.” and waits. On receipt of key: score hits, misses, near-misses honestly.
With identities confirmed: phase transitions and early-warning signatures (critical slowing down), feedback loop candidates from lagged cross-node correlations, emergence analysis (attractors, hysteresis, basin structure), and a common-systems reading of what the ensemble shares across otherwise unlike societies.
Download the key for the set you ran after you have locked your Stage 4 guesses. Each JSON maps Society_A–H to their true nation and supplies the T+0 anchor year and full offset–to–year table.
The protocol is on the honour system: lock your Stage 4 guesses before opening the key. To share your results, contact Kari or open an issue on GitHub.
CAMS claims that structural coordination physics — the shape of node coupling, the trajectory of stress and capacity, the signature of cognitive activation — is sufficient to fingerprint a society across history. If that claim is true, a blind-running LLM should be able to identify real nations from their CAMS scores alone, without any external geopolitical priming. The six datasets here are the test bed. The protocol enforces guesses-before-key ordering so results are genuinely falsifiable, not retrodicted.
The experiment is open. Anyone with access to a capable LLM can run it. Results, including misses, are encouraged to be submitted — failed identifications are as informative as correct ones about where the instrument’s signal is strong or weak.
— Kari McKern, Neural Nations · Datasets · Model · Validation