AUTHREX-DA
Admissibility, Not Correctness.
GOVERNANCE AND ASSURANCE ONLY
AUTHREX maps evidence to authority for actions. AUTHREX-DA extends the same mapping to data records and federated model contributions. It is a configuration of the governance layer, not a new component family. Given a record offered to a system, and the evidence available about that record and its source, may the record be admitted, and with what authority? It is a filter on admissibility. The proposed layer never repairs, imputes or reconstructs a payload. It emits a verdict, a reason code and an evidence reference, and writes them to an append-only record.
PROPOSED / ENGINEERING PREVIEW. All data is synthetic and seeded. Nothing here has been tested on physical hardware, fielded, or reproduced by an independent party. No certification, accreditation, authorization, TRL determination or endorsement by any organization is claimed or implied.
Scope exclusion. Governance and assurance only. No weapons, targeting, kinetic effects, lethality, offensive cyber or electronic warfare content, and no detection, identification, classification, selection, prioritization or optimization of any effect.
A Wrong Filter Admits a Bad Number With Authority Attached.
Most data-quality work asks whether a value is correct. This asks whether a value has earned the right to influence a decision. An accuracy filter that is wrong admits a bad number. An admissibility layer that is wrong admits a bad number with authority attached, and everything downstream inherits it. Release 1.8, core engine 1.0, assurance layer 1.4. Every figure on this page is copied from a generated artifact, not written by hand.
Four Verdicts. Two Properties That Hold Them Together.
Admitted at full authority.
Admitted, but capped: usable with its flag, not at full authority.
Held. The source enters staged re-entry, rate limited by a minimum dwell.
Not admitted.
Every verdict carries a reason code and a reference to the evidence snapshot it was decided on. That is a requirement, not a convenience, because a governance layer whose decisions cannot be explained afterwards is not a governance layer.
Strong evidence on one channel cannot buy off a failure on another, so a record does not average its way to admission.
If the layer itself is unavailable, nothing is admitted. An assurance layer that fails open is decoration.
Four Evidence Channels, One State Machine, One Record.
Internal consistency. Is the record coherent with itself?
Cross-source agreement. Do independent sources corroborate it?
Temporal stability. Is it consistent with this source’s own recent history?
Physical plausibility. Is it possible given the modelled envelope?
Evaluated separately. It is a derived mechanism rather than a fifth channel, and the distinction changes how contribution is counted.
Sources carry state and move through normal, flagged, quarantined, staged re-entry and restored. Re-entry is rate limited by a minimum dwell and terminates in a non-compensatory gate, so a source cannot return to full authority by waiting.
Federated contributions carry an attestation quote binding node, round, update reference, update hash, attestation state and the full training-data evidence summary. A quote is spent on presentation whatever verdict it earns. Absent or failing attestation caps a contribution and never lifts it.
Determinism, AUTHREX-DA-CANON-1. Reproducibility is treated as a property to be earned rather than assumed. Every unspecified transcendental function is replaced with a fixed-form series over IEEE-754 basic operations, values are quantized before evidence computation, and hash constants are FIPS 180-4 literals. One residual assumption is stated because it is not checked: that JavaScript engines do not contract or reorder double arithmetic.
Computed and Generated. Never Hand Written.
The characterization was written by hand three times and was wrong three times. It is now computed by the engine and generated into an artifact, and no measured figure for it appears in prose anywhere, including on this page. It has three layers, because the first two mislead. Direct gate attribution says which gate owned a reason code. Recovery participation says which conditions were failing for held sources. Causal contribution neutralizes each channel in the admission gate and the recovery conditions together and re-runs the battery. The three disagree, which is the finding: participation is not causation, and attribution is neither.
Known limit, stated plainly. Holding the assumed adversary bound fixed while varying the real compromised count, the proposed layer holds through a minority of compromised sources, then begins quarantining honest sources as the honest majority is lost, then fails in both directions at once. This is the clearest limit in the work and it is recorded as an open requirement rather than a solved one.
Retained negative results. Five negative findings and the counterexamples that produced them ship as first-class artifacts in the evidence package, including records that were corrupt by injected ground truth and nevertheless passed every detector-visible gate.
| Requirements DA-R01 to DA-R14 | Count | Meaning |
|---|---|---|
| TEST_PASS_LOCAL | 12 | An automated test passed on synthetic data on this machine |
| PARTIAL_LOCAL | 1 | The same, with a stated coverage gap |
| BLOCKED | 1 | DA-R09, bounded influence under faults, gated on WP-0 |
Nineteen audit passes are on file, five internal and fourteen external. External audit review is not the same thing as independent reproduction, and no independent reproduction exists. The evidence package carries seventeen artifacts under a three-part commitment.
| core engine 1.0 | 209e97b24ed782daac614f546b921fb96c56ca94e03d705379e9ea50f3e2ca4f |
| assurance layer 1.4 | f35707c81a435e7054d20601af34a53e910090c09a2362cf8ba755d08b174ef8 |
| test harness | 8165ef351581a41a2a8a9869db6099f86a87a9045be6d7b8c7fbb8f6feb4c741 |
| packaged standalone v1.8 | a24576be2d3d24619fb3058e9e3981733628559b547babee8343a11925a218e1 |
| evidence_set_root | ad3499989faf572ea3777d1569d5a0ce8f2ca238d2936b7c89595885dd18ffd6 |
Canonical acceptance target, for reproduction on a second platform: canonicalization AUTHREX-DA-CANON-1 at resolution 1e-8, scenario SC-4, seed 20260828, 120 rounds, 1,584 ledger entries, expected root f9e61d19fd7d91883ef170cec5f7c1e97680ca43b96d4a78a01f19319af2e60b. The engine hash governs whether that root remains valid.
Stated Before Anything Is Claimed.
No bench, no processor-in-the-loop, no latency measurement on any target. Runtime assurance that takes longer than the mission loop is not runtime assurance, and that has not been measured.
Every result is a single seed with no preregistered campaign, no confidence intervals and no multiplicity control.
No independent reproduction. External audits have reviewed this work, which is not the same thing.
WP-0 is blocked. The original quorum module, its configuration and its log were never supplied. A surrogate candidate model was built and checked instead, and it is labelled SURROGATE everywhere. No property DA-P1 to DA-P9 is model checked.
The canonicalization layer was reproduced on a second operating system. The whole-engine root of the current release is single-platform; the second platform is pending.
Self-assessed engineering readiness: 51 of 100. This is a self-assessed internal rubric. It is not a DoD score, a TRL, a certification, an authorization or an endorsement. Its only function is to stop software polish being read as evidence maturity, so the domains software cannot satisfy are scored at or near zero on purpose. The last two domains cannot be earned by writing code; they require evidence produced by someone other than the author.
Status. Core engine YELLOW, internal. WP-0 BLOCKED. Formal SURROGATE only. Multi-seed campaign HOLD. Paper, public and government use RED. Thirty-eight numbered open items are carried in the package.
Next. A preregistered multi-seed campaign design: seeds, distributions, matched operating points, hypotheses, stopping rules, multiplicity treatment and analysis plan, frozen before anything is run. Not another feature cycle.
Governance and Assurance Only.
Is this a product?
No. It is a proposed engineering preview of a governance configuration, on synthetic and seeded data, published for scrutiny.
Does it clean or repair data?
No. It never repairs, imputes, reconstructs or smooths a payload. It decides admissibility and records the decision.
Does it detect or classify anything in the world?
No. The scope exclusion above is load bearing: no detection, identification, classification, selection, prioritization or optimization of any effect.
Where is the evidence?
In the release package: a deterministic engine, a test suite, twelve ordered release gates, retained negative results and a hash-chained ledger with an archived canonical root. The console on this site is the packaged standalone build, hash-pinned above.
What would falsify it?
Reproducing the canonical acceptance target on a second platform and getting a different root, or demonstrating an admitted record that violates a stated gate at the published parameters.