Experiment report

ADVERSARY-02: a pre-registered test of CAMS

Fifteen predictions, written and sealed before scoring, about China and seven Chinese companies for 2021–2024

Kari McKern · ComplexityWorkz · 5 November 2026

Results are for the primary pool (7 scorer-runs, 20 passes) unless stated. Everything here is reproducible from the files in analysis/adversary_02/part2/. Interpretation is labelled as interpretation.

Jump to the interactive chart Scoring table Sealed predictions (Section 7)Methods note

Before looking at any data for 2021–2024, I wrote down 15 predictions from a model of how societies and organisations work, and sealed them. Then eight AI scorers rated China and seven companies for those years. Six predictions passed and nine failed. The model’s one built-in kill-switch was not triggered, so it survived this test. That is not proof that it is right.

The headline wins

Sealed in advancePredictions locked before scoringSection 7 was signed on 4 November 2026 at 21:28:58 AEDT, before any 2021–24 scoring was done. The sealed file’s fingerprint is recorded in the repository.

The kill-switch heldFC5 not triggeredThe rule: if 3 or more of the 4 core predictions (P1, P3, P4, P5) failed, the framework would count as failed. Only one (P3) did. This held in all four pools.

It called the collapseEvergrande: −3.2 → −6.3Evergrande’s mean node value fell every year of 2021–24, as predicted (P1). No scorer-run that could be scored contradicted it.

It called the survivorsBYD and CATL heldBoth stayed far above their floors (P4, P5). Six of seven runs passed each one and none failed it.

Shocks stayed contained0 of 8 events spreadNone of the eight policy shocks tested showed contagion between entities (P12). All seven runs agree.

The one call persistence can’t makeHuawei’s slide reversedHuawei’s mean rose from 7.39 (2020) to 11.3 (2024) (C2). Predictions that simply say “nothing changes” cannot pass this one.

The misses, stated plainly. Nine of fifteen failed: P2, P3, P6, P7, P8, P9, P11, C1 and C3. The failures cluster in the finer-grained calls about detailed shape and size of change (Huawei, Vanke, Longfor, China’s level, BYD’s stress). The broad calls (who breaks, who holds, no spillover) stood.

1. What is CAMS, for readers who have not met it?

CAMS (Complex Adaptive Model of Societies) treats a society, or an organisation, as a small system of eight functional parts, called nodes. Each node is scored every year on four measures. Here are the definitions used in this experiment, as set out on the Canonical Definitions page:

C: Coherence: How well aligned and coordinated the node is within its own function.
K: Capacity: How well the node actually performs its function.
S: Stress: How fast disorder is being produced inside the function. It is not how hard the outside challenge is.
A: Abstraction: How well the node’s ideas and symbols are tied to what it actually does. High A means sophisticated and grounded. Ideas that have drifted away from reality score low.

An entity is the thing being scored: here, China as a country plus seven companies. China’s eight nodes are Helm, Shield, Lore, Stewards, Craft, Hands, Archive and Flow. The companies use Board, Risk, RD, Finance, Operations, Workforce, Brand and Markets, which the experiment maps onto the same eight positions (the mapping is provisional and unapproved).

Scorers give only the four raw measures. Everything else is computed afterwards, centrally:

Node value, V = C + K − S: One number per node-year. M is the mean of V across an entity’s eight nodes, the entity’s overall health in a year. Note: this experiment’s sealed rule has no A term. The Canonical Definitions page defines V with an extra A/2, so M values here are not directly comparable with V on other pages of this site.
Bond strength: How tightly pairs of nodes are tied together, driven by their coherence and abstraction and damped by their joint stress.
Energy = K − S: Capacity left over after stress. Below zero, an entity is spending more than it can sustain.
Focus = A × C: How clear and well joined-up the entity is. Energy and focus together give each entity a “personality”: a position on a two-axis map.

2. How the test worked

Pre-registered. The 15 predictions, with numeric pass/fail thresholds, were drafted, then signed by me on 4 November 2026 at 21:28:58 AEDT, before the 2021–24 scoring. The signed file is predictions_SIGNED.md (SHA-256 e15e1198…fb28; the manifest that lists it hashes to 766a1cbd…be6b; the scoring table misprints it as …fe6b).
Scored by AI, not by me. Eight AI scorers (Claude agent, Gemini, Kimi, Claude chat, Codex, Manus, Perplexity and Grok) produced 23 full scoring passes. Each pass covers 8 entities × 4 years × 8 nodes. Scorers could use only information available by 30 June of the following year.
Pooled fairly. Within each run the passes are averaged first, then every run counts equally, so a scorer with more passes gets no extra say. The primary pool is 7 runs and 20 passes. Grok is reported separately because of its scorer grade (see below). Perplexity’s attribution is unconfirmed, so results are also shown without it, and without Manus.
Fixed comparison. Pass/fail uses the primary pool and primary thresholds only. Looser secondary thresholds are shown for information and never change a verdict.
A kill-switch. Failure Condition 5 (FC5) would trigger if 3 or more of P1, P3, P4, P5 failed in the primary pool.

3. Scoreboard: all 15 items

6 of 15 passed; 9 failed. Verdicts were identical across all four pools except P6, P7 and P8. Each of those fails in the primary pool but passes in at least one alternative pool (P6 with Grok added or Manus removed; P7 and P8 with Grok added), so they are counted as failures.

Item	The sealed call, in plain words	What the data showed	Result
P1	Evergrande loses viability: mean node value (M) at or below 5.0 in every year and at or below 4.0 in 2024	M ran −3.2, −4.2, −5.4, −6.3 for 2021–24, falling every year	PASS
P4	BYD preserves: 2024 M at or above 7.2 and no year below 6.5	M ran about 10.0, 11.4, 11.8, 11.5	PASS
P5	CATL preserves with strong internal bonds: M at or above 8.3 each year and bond strength at or above 23.1	M 11.6–11.8 each year; bond strength 38.0 → 39.1	PASS
P10	Longfor is the comparator furthest from Evergrande (further than Vanke)	Distance to Longfor 4.05, to Vanke 3.31. All 7 runs agree	PASS
P12	Policy shocks do not transmit between entities: none of 8 events scores 2 or more	All 8 events scored 0. All 7 runs agree	PASS
C2	Huawei’s slide reverses: 2024 M above its 2020 value of 7.39	M = 11.3 in 2024 (also clears the stricter 8.67 margin version)	PASS
P2	Evergrande keeps the same weak spots (Stewards and Shield lowest) in at least 3 of 4 years, under four conditions	All four conditions held in only 1 of 4 years; the lowest node in 2024 was Archive	FAIL
P3	Huawei reorganises: M stays below 9.15 every year	M ran 8.1, 8.2, 10.9, 11.3: above the ceiling in 2023 and 2024	FAIL
P6	Stewards or Shield is the weakest node in at least 14 of 16 company-years (four primary companies)	13 of 16 in the primary pool (14 of 16 once Grok is added)	FAIL
P7	Bond-strength gap between Evergrande and its comparators stays at or above 10.0 every year	Gap shrank 22.1, 19.5, 13.6, 8.2. Passes only under the looser secondary threshold (8.0)	FAIL
P8	Evergrande stays far from Vanke and Longfor relative to how close those two are to each other (both ratios at or above 5.0)	Ratio 4.35 against Vanke, 5.32 against Longfor. Passes only under the looser secondary threshold (3.5)	FAIL
P9	Huawei and SMIC converge: M within 1.5 of each other every year	Gap 0.13, 0.24, 3.30, 2.73: they diverged from 2023 as Huawei recovered	FAIL
P11	Vanke (at or above 7.49) and Longfor (at or above 8.66) hold their level every year	Vanke M 8.9, 7.7, 4.5, 1.2. Longfor 10.7, 8.1, 7.1, 6.1	FAIL
C1	China’s 2024 M rises clearly above its 2020 level (at or above 9.30)	7.36	FAIL
C3	BYD’s stress rises (2024 at or above 5.54) while M holds	Stress 4.11: it did not rise	FAIL

Full detail, all four pools, per-scorer verdicts and sensitivity runs: scoring_table.md and scoring_table.csv. In the per-scorer table, P is pass, F is fail and I is indeterminate: where a scorer left too many cells blank to decide an item, it is marked I, not F.

4. Did CAMS survive, and why?

FC5 was not triggered, in any of the four pools. Of the four core predictions, three passed (P1, P4, P5) and only P3 (Huawei’s “reorganise, then stabilise” call) failed. So by the rule written down in advance, CAMS survived this test.

The model’s main structural calls held up: it said which entity would break (Evergrande), which would hold (BYD and CATL), and that shocks would not spread between entities. Its finer predictions about the detailed shape of change did worse. Surviving a test is not the same as being proven right.

An important caveat, from the sealed document itself. Section 7 states that all 11 of the non-coupling predictions (P1–P11) would also have passed if every entity had simply stayed at its 2020 value. Four of the six passes (P1, P4, P5, P10) are of that “no reversal” type, and P12 is a negative-result prediction about coupling. C2 is the only pass that persistence cannot produce, and C1 and C3, the other two contrast predictions, failed. So these wins show that the framework’s calls were not contradicted. They do not show skill beyond “things continue as they were”.

Interpretation (the author’s reading, not part of the scoring). P11 is counted as a fail: Vanke fell from 8.9 to 1.2 and Longfor from 10.7 to 6.1. My own view is that the Vanke result is better read as construct-validity evidence than as a plain miss: the scorers independently saw Vanke deteriorate, and the instrument picked it up even though the prediction had said otherwise. That is an interpretation. It does not change the verdict, and it is open to the objection that the scorers know how events turned out (see limitations).

5. The “personalities”: where each entity ended up

Placing each entity on the energy and focus axes turns the raw numbers into a map. The dashed lines split it into quadrants. Energy 0 is a natural line. The focus line at 30.25 (5.5 × 5.5, the product of the scale midpoints) is an analyst’s choice, not part of the sealed predictions, and moving it would move entities that sit near it, Longfor especially.

Interactive: switch between the map and the two time series, toggle entities, and optionally add the 2010–2020 baseline from Part 1.

Entity	2021 energy / focus	2024 energy / focus	What the 2024 data show
Evergrande	−5.9 / 8.7	−7.8 / 2.5	Low-energy, low-focus throughout; by 2024 its focus is almost gone. The direction Part 1 pointed, at a larger scale.
Huawei	1.4 / 47.3	3.7 / 57.4	Highest focus of any entity in 2024. Shield (protection) was its weakest node in 2023 and 2024.
CATL	4.0 / 54.0	4.1 / 56.9	Strong quadrant throughout; Lore (knowledge) its strongest node in 2024.
BYD	3.0 / 46.8	3.9 / 55.6	Strong quadrant throughout; Lore strongest in 2024.
SMIC	1.8 / 40.5	1.9 / 43.6	Strong quadrant throughout; Lore strongest in 2024; Shield weakest.
China	2.3 / 41.8	1.3 / 37.5	Energy stayed between 0.4 and 2.3; focus eased from 41.8 to 37.5.
Longfor	3.5 / 44.2	0.3 / 32.8	Drifted toward both lines. In 2024 only 2 of 6 scorers who could score it put it in the pooled quadrant.
Vanke	2.4 / 38.6	−3.0 / 19.4	Dropped into the low-energy, low-focus quadrant from 2023.

Values: primary pool, from entity_personality_entity_year.csv. Evergrande’s 2024 point rests on 4 scorable runs; the other entities on 6 or 7.

Static version of the entity personality map, 2021 to 2024 — Static version of the map: paths run 2021 → 2024, last point is 2024; dashed lines at energy 0 and focus 30.25.

6. Who scored best

The scorers were not equal, and the assessment of them is part of the evidence. Grades come from scorer_assessment.md: five criteria (reliability, rule compliance, independence, calibration, and discipline over blank cells), equal weight. They are judgments from the evidence, not measurements of truth, and the pool they are compared against is not ground truth either.

Bar chart of overall scorer grades: Claude agent A-, Gemini B, Kimi B, Claude chat B, Codex B-, Manus B-, Perplexity C-, Grok D+ — Overall scorer grades (A = 4 … D = 1). Grok is reported separately; Perplexity’s attribution is unconfirmed.

Claude agent, A−: the most repeatable, and closest to the pool.
Gemini, B, Kimi, B, Claude chat, B: sound levels, with specific issues noted (an impossible timestamp for Gemini; two noticeably different passes for Kimi; a disclosed bit of context for Claude chat).
Codex, B−: consistent and compliant, but left 52% of corporate cells blank. Manus, B−: well isolated and candid, but the harshest scorer.
Perplexity, C− (attribution unconfirmed, no run note) and Grok, D+ (carried most cells forward unchanged; Grok is outside the primary pool).

7. Limitations

Outcome knowledge. The scorers are AI models working from memory or web search, and they know how events turned out. So did the AI that drafted the predictions (every prediction is tagged for this in Section 7), and the choice of contrast themes was made after the 2010–2020 results were visible. The pre-registration protects against moving the goalposts. It cannot certify that hindsight was absent.
Most passes are “no change” calls. See the caveat in section 4.
A method seam at 2020/2021. Part 1 (2010–2020) and Part 2 (2021–24) were scored by different scorer sets in a different prompt era. The step from 2020 to 2021 is a seam, not a finding.
The quadrant lines are a choice. Energy 0 is natural; focus 30.25 is not part of the sealed predictions.
Levels versus rankings. Scorers agree on how entities rank more than on exact levels, which varied scorer to scorer. Rank correlations against the pool ran 0.81–1.0 for the six runs with full entity coverage; Codex scored too few entities to compute one.
Different definition of V. This experiment’s V = C + K − S (no A term), as sealed. The Canonical Definitions page defines V = C + K + A/2 − S.
Small, uneven grids. Perplexity’s attribution is unconfirmed, a few scorers had partial blank grids, and Evergrande’s 2024 point rests on 4 scorable runs.
Fifteen items, one sample. One experiment on one country’s firms. A single result like this should be replicated before it carries weight.

8. Files and reproduction

Everything is in analysis/adversary_02/part2/: the sealed predictions and its manifest, the scoring table (scoring_table.md and .csv), the entity-year personality data, scorer spread and agreement files, the methods note, the scorer assessment, and the two scripts (51_score.py for verdicts, 55_entity_personality.py for the map). Per-pass data and the Part 1 results are in the separate Part 1 submission (pull request #39). The chart above reads the same values as entity_personality_entity_year.csv, rounded to two decimals.

More CAMS papers and tools: Explore.