What this model is, and what it has and has not demonstrated

Which artefact exactly, at which version, and what is in and out of its scope.

The systemic entity kernel, running at commit 97151460269d with source digest cf3386d9154e0d98. Both are served at /api/provenance together with the working-tree state. A clean-clone reproduction claim applies only when that endpoint reports a clean release tree.

The platform carries 161 measured capabilities, 35 that settle or report on settlement, and 94 that do neither and are named as not load-bearing. Only the first group is in scope for reliance. The classification is generated from the route tree and anything it cannot classify fails the build.

CHECK IT · /api/provenance · /api/capabilities · /map

What a user may conclude, and what they may not.

The kernel measures four buffers per entity and reports, in days, how long each lasts on its current trend. It answers "when does this constraint bind if nothing changes". It does not forecast that distress will occur: new financing or a policy decision resets the clock, and that is the normal case rather than the exception.

It is not a credit rating and must not be used as one. It is not investment advice, not a benchmark, and not a substitute for an issuer's own disclosure. A published figure is a measurement of public statistics at a stated date, and the date is part of the figure.

Why this method, and what the thresholds rest on.

Four primitives, each measured in days: LIQUIDITY, REFINANCING, DEPENDENCIES and RESPONSE. The unit is the point — a state, a banking system and a company answer the same four questions in the same unit, so they compose without a mapping layer that could hide an assumption.

The regime is the state the model reports, not a band of the score. Above 72 and below 48 the two agree by construction; between them the label comes from the hidden Markov state rather than the threshold, so a score of 55 can be reported as critical and that is the model speaking rather than an error. The thresholds are anchored to named historical crises and are published so they can be argued with.

CHECK IT · /api/prime/entity/{slug} · /evidence/{slug}

Which sources, at what lag, and how absence is handled.

Every input carries the authority that published it, the date it was observed, and an evidence tier saying how it came to exist: IDENTIFIED read from the authority, STRUCTURAL derived by an accounting identity, ESTIMATED scaled by a coefficient no natural experiment fixes, STYLIZED a stated assumption. The weakest tier in a chain sets the tier of everything downstream.

An input that is absent is reported as absent, never as zero. That rule exists because it was once broken: blank cells from the ECB became measured zeros and Czechia was published as having breached its liquidity floor. Three of the nine entries in the corrections register are the same fault in different places.

The figures can be re-performed against the authorities without this platform's code: `npx tsx scripts/reperform.mts <slug>` fetches Eurostat, the World Bank and the ECB directly and compares. Across eleven countries it reproduces 51 published figures with no disagreement.

CHECK IT · /evidence/{slug} · scripts/reperform.mts

What has it demonstrated, against what baseline.

None of the 4 instruments has demonstrated skill against its own base rate. That is the finding, and it is stated first because a pack that buries it is not documentation.

Skill is measured against each claim class's own climatological base rate, not against nothing. A forecast that a managed currency stays managed is correct most of the time and demonstrates nothing; that distinction is the reason the headline above reads as it does. An instrument with no settled outcomes is reported as unproven rather than as accurate.

The calibration curve — forecast probability against observed frequency — is published as an image at /api/prime/reliability and is regenerated from settled outcomes.

CHECK IT · /api/prime/instruments · /api/prime/reliability · /accountability

Where is it weak, stated by the model rather than found by the reviewer.

The four primitives are not equally covered. DEPENDENCIES carried no inputs at all for any entity until 2026-09-13 — a quarter of the kernel was decorative and nothing in the payload said so. It is now measured from the BIS exposure matrix for 35 sovereigns and 17 banking systems, bounded by which systems report to the BIS.

The sovereign lens reads no liquidity input for any euro-area member. The World Bank publishes reserve cover for each of them — Greece at 1.295 months for 2024 — and whether a currency-union member's national reserves are the right buffer is a modelling question that has not been settled. It is recorded here rather than closed quietly.

Concentration is measured and has no horizon: there is no rate at which a Herfindahl drains. The payload distinguishes "nothing was read" from "read, and not a duration", because it once did not and reported three IDENTIFIED figures as unmeasured.

CHECK IT · /api/prime/entity/{slug} · /corrections

What is checked after deployment, how often, and what happens on a failure.

A deterministic sentinel runs every thirty minutes and checks whether the platform's own stated invariants still hold. It uses no language model: each check is an assertion with a source. INDETERMINATE is never counted as a pass — an assertion whose data source has emptied can pass vacuously, so losing the evidence is reported as losing the evidence.

The verdict is public whatever it says, and /api/integrity returns 503 while it is failing. Every run is kept rather than a selection of them, so the transitions are computed from the record instead of written beside it.

One of the checks is that the code on disk still hashes to what the release was stamped with, so a file edited in place after a build is a public failure rather than an invisible one.

CHECK IT · /api/integrity · /api/integrity/history · /status

What has changed, what was wrong, and what stops it recurring.

14 corrections are on the public register, each naming the surface it was live on, what was claimed, what was actually true, the mechanism, and the regression test that now prevents it. The register is in the repository rather than the database, so it is versioned with the code that caused each entry and cannot be edited without a commit.

Every release is stamped with its commit and a source digest recomputed from the files on disk. A release whose files have moved since it was stamped reports DRIFTED rather than a version number.

CHECK IT · /corrections · /api/prime/corrections · /api/provenance

What a reviewer would look for and not find, said here rather than left to be discovered.

• No accuracy figure is quoted for the platform as a whole. Accuracy over a mixed population of claim classes is arithmetic without meaning; skill is reported per class against that class's own base rate, and where nothing has settled it is reported as unproven.
• No user or traffic figure is quoted anywhere in this pack or in the investor material. The analytics record is dominated by a single day of referral traffic whose signature is fabricated, and a number that cannot be defended is not published.
• No claim is made about regulatory classification. Whether a published sovereign score engages Regulation 1060/2009 is a legal question that has not been answered and is not answered here.