Does the Crisis Detector Work on Firms?

Part I · Design

The Test and the Frozen Protocol

Can a detector built for nations tell a failing firm from a healthy one?

The rule

The crisis-typology rule was sealed on 5 November 2026 after scoring 13 of 15 known national crises correctly (crisis_typology_rule_v1.md). It compares two clusters of nodes. If the economic nodes are clearly stronger than the governing and meaning-making nodes, the year reads as an epistemological crisis; the reverse reads as an economic one.

The firms

The control list was frozen before any scoring began and its SHA-256 hash recorded (P4_protocol_frozen.md, hash record). Controls had to share industry, scale and era with the case, and have no documented pathology of the same family in the window.

Corporate node mapping

The cases and primary controls were scored with repeated independent AI passes and averaged (ensemble-mean ENS files, with scorer spread in ENV files). The secondary controls come from a single-scorer run by Kimi (Moonshot AI), added as a sensitivity test once the primary results were in. Kimi scored in blocks of years rather than year by year; see Part V.

Part II · Results

The Rule Fires on Everyone

Firing rate does not separate the collapses from the survivors.

The national rule fires rarely: a third of society-years come out unclassified. On firms it fires in 44–100% of years, and the controls fire about as often as the cases. Averaged over each group, the three cases fire in 60% of their years and the three primary controls in 54%. In the WorldCom pair the control fires more often than the case.

Longest run = most consecutive years firing with the same classification. Source: P4_detector_summary_9firm.csv, P4_detector_results_9firm.csv.

Year by year

Each chart shows the gap for a case and its two controls. Above the upper dashed line the year reads epistemological; below the lower line, economic.

What might separate them

Threshold-only rule fires on controls as readily as cases. Separating features are persistence (Boeing: 7-year unbroken run) and magnitude trajectory — not raw firing rate.

That is the working finding from the analysis. It has not been tested: it is a hypothesis drawn after seeing the data, and Part V shows that it holds for two of the three pairs. A formal duration-and-magnitude score is listed under Part IV.

Part III · Interpretation

Structural Personalities

Reading each firm’s node hierarchy as a character.

Three patterns the portraits point to

Where the disorder is, not how often the rule fires. In all three cases the nodes that should hold meaning and governance (Lore, Archive, Shield) are both low-ranked and highly stressed. When a control has a weak node, the weakness is not paired with high stress.

Craft coupling. In every healthy firm, Craft is the top-ranked or the most tightly coupled node. Boeing keeps nominal Craft coupling, but the analysis reads its Craft as cut off from Lore and Archive.

Capital volatility. Stewards is the most volatile node in all three cases (Boeing 4.63, Theranos 6.57, WorldCom 5.64) and calmer in the controls quoted (Airbus 2.99, Quest 0.91, Lockheed 0.52).

Each of these could become a detector feature, but it would have to be frozen and tested on new firms first.

Part IV · Still to run

The Baselines the Protocol Requires

Three comparisons were written into the frozen protocol; a fourth was proposed after the results.

Item 4 was designed after seeing which firms collapsed, so on these nine firms it can only be exploratory. To count as validation it would need to be frozen and run on a new set of cases and controls.

Part V · Checks

What Was Verified, and What Doesn’t Hold Up

Verified

• Reproduction. All 86 firm-years were recomputed from the ENS files with V = C + K − S. Every gap and classification matches the results file.
• Stored Node Value. The ENS files’ Node Value column is C + K − S + 0.5A on every row, as the report says, so the detector correctly recomputed V instead of using it.
• Freeze record. The SHA-256 of P4_protocol_frozen.md matches the hash record (e5ca844c…74f8bcf), so the protocol has not changed. The file timestamps put the freeze (17:37, 10 Nov) before the first scored file (17:44).
• Sealed rule. crisis_typology_rule_v1.md is unchanged since it was committed.

Does not hold up, or needs a decision

1. Persistence does not cleanly separate cases from controls. Boeing’s 7-year run is the longest, but Airbus has a 5-year epistemological run in the same years. WorldCom’s longest run (4) equals Illumina’s and is shorter than Verizon’s (5).
2. Magnitude fails in the WorldCom pair. Boeing’s peak is above Airbus’s and Theranos’s above Quest’s, but AT&T’s epistemological peak (+2.90) is higher than WorldCom’s (+2.13). The working finding in Part II holds for two pairs out of three.
3. The protocol has no registered pass criterion. It requires every firing to be reported but never says what result would count as a pass or a fail. Under §0A (“where the protocol is silent, stop and ask”) this page does not declare P4 passed or failed. The plain observation stands: the threshold rule does not separate cases from controls on firing rate.
4. Secondary controls are weaker evidence. They are single-scorer runs (every ENV spread is zero), and Verizon covers only 2000–2004, so its 100% firing rate rests on five years.
5. Most secondary-control scores were carried forward, not scored each year. The secondary controls are Kimi’s run of 10 November 2026 (protocol note); the P4 files match it row for row. Kimi scored in blocks of years and tagged every score. In the P4 windows, 480 of the 832 scores for Lockheed Martin, Illumina and Verizon are tagged Carried_Forward: no year-specific evidence was examined and the previous score was held. Only 152 are New_Evidence. Kimi’s note says relational analysis must not treat carried-forward cells as independent observations. It shows in the gaps: Illumina reads +1.17 in each year 2013–2016, Lockheed Martin −1.17 in each year 2013–2015 and −1.50 in each year 2020–2022, and Verizon −2.00 twice and then −1.83 three times. Their firing runs are therefore partly built in by the scoring method, which matters for any persistence criterion and for the circular-shift null. It also weakens the portraits’ reading of Lockheed Martin’s “near-zero volatility” as a trait of the firm.
6. Scored windows are wider than the protocol’s. The protocol gives approximate windows (Boeing ~2018–2020, Theranos ~2003–2018, WorldCom ~1999–2002). The scored years are 2013–2022, 2008–2018 and 1996–2004. Nothing on record shows the scored years were fixed before scoring.
7. The baseline commit is missing. The protocol and hash record cite baseline commit 58fcaaf, which does not exist in this repository.
8. The sealed rule’s own note no longer fits. It says the rule “fires rarely and cleanly”. On firms it fires in 44–100% of years. The rule itself is unchanged; the note describes national data only.

Files

• P4_protocol_frozen.md (locked) · P4_protocol_hash.txt
• crisis_typology_rule_v1.md (sealed 5 Nov 2026)
• P4_detector_report_9firm.md · results · summary
• P4_detector_report.md (first run, six firms)
• P4_personality_portraits.md
• Kimi’s single-scorer run (source of the secondary controls): scores · per-score status · protocol note · build script
• Scores: ENS (ensemble mean) and ENV (scorer spread) for each firm in analysis/P4/

Run 10 November 2026 under the frozen P4 protocol and crisis_typology_rule_v1 (sealed 5 November 2026). Analyses by AI collaborators, presented as written; checks by Claude Code. CAMS: Complex Adaptive Model State.