Blind Testing & Validation

Tests run before the results were known, results reported as they came. CAMS is not externally validated. The out-of-sample holdout (2026–2028) is locked and unscored. Everything here is in-sample or pre-registered — the distinction matters and is marked throughout.

Kari Freyr McKern · 2026 · Open validation invite →
6
ADVERSARY-02 predictions passed (of 15)
9
ADVERSARY-02 predictions failed — kill-switch not triggered
87%
Crisis typology accuracy on canonical data
32
Prospective predictions locked, unscored (2027–2028)
2
Open challenges — participate now
Section 1
Take Part — Open Challenges
These experiments are open. Participate before reading the results.
OpenBlind challenge
JUNO / CAMS Blind Analysis Challenge
8 anonymised societies, 1750–2026. Load a dataset and protocol into any capable AI, run a five-stage analysis, and submit your findings before the unblinding keys open. Open science — results aggregated publicly.
Ongoing
Start the challenge →
NewOpenExperiment kit
Repeat the CAMS Personality Experiment
Download a name-masked packet from the CAMS personality kit (Core, Extension or Denmark sets), run a blind era-recognition pass with any AI, and compare your result to the panel. Use the participant route before reading recognition counts.
7 Nov 2026
Read the challenge Kit files Download participant zip
Get the kit →
OpenInvite
Peer Review & Validation Invite
An open invitation to independently replicate, audit, or stress-test CAMS-CAN. What's already been tested, what hasn't, and how to send in a finding.
Open
OpenData challenge
The Scale Invariance Experiment
Five blinded CAMS time-series. Can you identify what system each represents from its metric signature alone? Download the CSVs and submit your best guess.
2026
Section 2
Pre-registered Tests
Predictions signed and sealed before any scoring was run. Results reported as found, failures included.
New6/15 passedPre-registered
ADVERSARY-02 — A Pre-registered Test of CAMS
15 predictions sealed on 4 Nov 2026 before any 2021–2024 scoring for China and 7 Chinese firms. Result: 6 passed, 9 failed. The kill-switch (FC5) was not triggered. Written for readers new to CAMS.
Full report Scoring table & sealed predictions How to read the map
Headline results (primary pool, 7 scorer-runs)
PASSEvergrande collapse (P1) — mean node value −3.2 → −6.3
PASSBYD and CATL held (P4, P5) — 6 of 7 runs pass each
PASSNo spillover (P12) — 0 of 8 policy shocks showed contagion
PASSHuawei's slide reversed (C2) — 7.39 → 11.3
FAIL9 predictions failed — see the full scoring table
6 Nov 2026
NewInterimSealed rule
P4: Does the Crisis Detector Work on Firms?
The sealed crisis-typology rule applied unchanged to Boeing, Theranos, and WorldCom against six matched controls. Baselines pending — this is an interim result, not a final verdict.
10 Nov 2026
New87% accuracySealed
Crisis Typology — Cluster-Gap Rule v1.0
A sealed rule distinguishing epistemological crises (failure of knowledge and governance) from economic crises (failure of labour and resources). Gap = ECO cluster V − EPI cluster V; ≥+1.0 epistemological, ≤−1.0 economic. 13/15 (87%) on canonical ensemble data.
5 Nov 2026
ρ = 0.70, p < 0.0001Pre-registered hypothesis
Relational Inelasticity Hypothesis — Blind Test
The preregistered RIH blind test from "A Core Response to Outright Dismissal". Spearman ρ = 0.7017, p = 0.0000248. Would have killed the project if it had failed.
Sep 2026
Section 3
Panel Results
Blinded classification, scoring, and morphospace work by human and AI panels
NewPanel report
CAMS Blind Experiment Report
Kari McKern & Fable 5 (Anthropic). Full panel analysis: blinded society classification, JUNO scoring, and morphospace placement. PDF included.
Full report (PDF) Morphospace MDS plot
2026
NewReport
CAMS Personality: Independent AI Replication
Full report of blind era recognition and historical mood comparison. An exploratory, promising demonstration — not a causal validation and not peer review. Sealed evidence and audit caveats included.
Full report Sealed evidence
7 Nov 2026
NewProtocol report
Protocol v1.1.1 — Cognitive-Affective Extension
Kari McKern & GPT authors. 3 blind sets, 24 societies. Mean reciprocal rank improved from 0.667 to 0.833. Calibration gate failure documented. Signed activation operators, figures and data.
2026
Named assessment
CAMS v1.1 Blind Test — Named Scientific Assessment
JUNO experiment reveal. 8 masked societies unmasked (Germany, Norway, Russia, UK, Chile, Poland, Argentina, Australia). 1,201 society-years, 232 inflection points. Core formulas validated at machine precision.
1 Jul 2026
NewReport + working paper
Reading Societies Without Their Stories — CAMS Deep-Time Test
20 ancient society-windows (Longshan to the Xiongnu). 22 blind passes per prompt from Claude, Grok & GPT. Pre-registered H1–H5. Collapse drains every node at once; Aegean recovers far slower than it falls; models disagree on whether oral tradition counts as evidence.
2 Oct 2026
Newr 0.89–0.93Record
CAMS Human–AI Scientific Record
Preserved record of the GPT conversation behind the book, verified: all 13 files match their hashes. Blind-rater agreement r 0.89–0.93, ICC 0.869 reproduced from raw CSVs. Zenodo DOI 10.5281/zenodo.23793977.
30 Oct 2026
Protocol
JUNO/CAMS Third-Party Validation Protocol (TPV-1)
v1.0, July 2026. A standardised protocol for independent replication — characterisation-only design, no identity guessing. 4-stage analyst prompt, convergence scorecard, null symmetry binding. Compliant with v1.1.2 operators.
Jul 2026
NewConsistency check
Attractor Coherence, Revisited
Blind Sets 1–6 (failed Set 5) re-run on the full 53-society JUNO dataset. A consistency check — not a blind test. Documents what held and what didn't across the larger corpus.
9 Nov 2026
Section 4
Validation Studies
Retrospective and in-sample tests of the instrument's structure — not external validation, but structural integrity checks
Variance = 1/√NPaper
Ensemble Validation — The Wright Thesis: Confirmed with Qualifications
Variance reduction = 1/√N exactly; five independent scorers produce the optimal estimate. Uncertainty reflects genuine epistemic ambiguity, not noise. Germany & USA 1880–2026. Contestation Index introduced.
May 2026
Paper
CAMS v1.0.1 — Instrument Test Battery
Six-part validation: corpus audit, formalism repair, vantage-point bias control, and pre-registered 2027–2028 predictions. 2,532 society-years, 25 entities.
Jun 2026
Paper
Graph-Theoretic Diagnosis of Societal Collapse: CAMS Validation Across 18 Societies
Principal component analysis across 7,880 node-year observations. PC1 explains 72.7% of variance. 17 crisis fingerprints matched. USA ρ = 0.97, UK ρ = 1.00. 12 falsifiable predictions registered.
10 Jun 2026
NewAUC 0.930Structural audit
JUNO Structural Audit — Rank-1 Bonds, Node Dynamics & Stewards
36 societies, 32,361 node-years. The bond formula proven algebraically rank-1. Stewards confirmed as a leading political-economy indicator (AUC 0.930). Temporal-inelasticity pilot: Stewards–Flow I = 0.822.
Aug 2026
Paper
Cross-LLM Structural Concordance in CAMS Datasets
Do different AI models score the same societies the same way? Reproducible coordination signatures documented across independent AI assessors — consistent structure, not identical scores.
2026
NewTechnical note
Phase A Re-Run After Bond-Column Realignment
Node-labelled Phase A re-checked after a data realignment. World Bank T1 (r = −0.128) and Shield × V-Dem T3c reproduce to published precision. SIPRI-real and an Australia pilot flagged as unrecovered source files, not alignment reversals. λ₂ / SBD unaffected.
28 Aug 2026
NewExternal review
Carbon Evaluates CAMS
An outside AI's report on the project's scientific value, presented as written with checks. Not an endorsement — an independent read of the methodology and evidence base.
11 Nov 2026
Section 5
Predictions on Record
Forward-looking tests locked and timestamped before scoring. The instrument is not externally validated until these are scored.
NewUnscoredZenodo DOI
Prospective Predictions 2027–2028: USA, China, Australia
32 falsifiable, probability-weighted claims under JUNO v1.2-Final, lodged before outcomes are known. Zenodo DOI 10.5281/zenodo.24046639. Scoring begins when the period ends. A thermometer projected forward, not a crystal ball.
Read the register Zenodo record
8 Nov 2026
NewUnscoredLocked verbatim
Frozen 2026–2028 Out-of-Sample Prediction Register
12 falsifiable predictions from the 10 June paper, locked verbatim for later scoring. USA–Argentina 0.97 tagged similarity-not-prediction. JUNO v1.2 is not externally validated until this holdout is scored.
28 Aug 2026
UnscoredRegistered predictions
Pre-registered 2027–2028 Predictions — From Instrument Test Battery
From CAMS v1.0.1 (June 2026). Registered as part of the six-part instrument test battery; scored separately from the August and November registers.
Jun 2026
What counts as external validation? None of the above. In-sample tests, structural audits, and pre-registered tests on historical data are not the same as an independent replication or a scored out-of-sample holdout. The 2026–2028 predictions are the instrument's first true out-of-sample test. Scoring begins in 2027.
Neural Nations · neuralnations.org Kari Freyr McKern · 2026 · Open science