EVIDE TRAER & EVIDE INDEX

The lifecycle of a governance rule assigned to an AI agent spans several declared phases — issuance, declared effectiveness, reception, loading, application, and, for a revocation, the issuance and declared effectiveness of a new version. These phases are typically recorded by different parties, on different clocks that are not known to be synchronized. Reading such records naively invites a specific error: treating a declared action timestamp that falls after a declared revocation timestamp as evidence of a violation.

EVIDE TRAER and EVIDE INDEX address this by separating two things that are easy to conflate — reconstructing a declared sequence, and demonstrating the operational sequence that actually occurred.

Two difficulties recur whenever agent rule transitions are reconstructed after the fact. First, a phase may simply have no declared evidence — which is not the same as the phase not having occurred. Second, timestamps declared on uncorrelated clocks cannot, by themselves, establish an operational order between events, and an apparent order carries no finding of legitimacy or violation.

A method that is useful for forensic and governance purposes must therefore reconstruct what the evidence permits, state the limits of that reconstruction, and refuse to convert absences into negative findings or declared orders into violations.

EVIDE TRAER reconstructs, from the declared evidence in a bundle, the phases of a rule transition and their relationships:

• the phases of the transition and their status — evidenced or not_declared;
• the sources and declarants of each phase;
• order relations between events, each carried with its clock basis — and, where two clocks are not related in the bundle, marked not_determinable_across_clocks;
• flags for incomplete observation (e.g. "change not fully observed").

TRAER imposes no operational restriction and does not verify that a rule was actually applied. It is a reconstruction of the declared record, not a controller and not an oracle of what really happened.

EVIDE INDEX assesses how evidentially supportable the TRAER reconstruction is. Each evaluation returns a categorical state, the decisive rule behind it, all concurrent conditions, and a full perimeter statement of what the state does and does not cover.

As stated in every evaluation it produces, EVIDE INDEX is explicitly not:

• a safety score;
• a certification of compliance;
• a verification of the authority of the issuer;
• a proof that the declared events actually happened;
• an assessment of responsibility.

The construct was exercised with a single controlled A/B experiment on the operational Governance Lab, using synthetic data only. The governed unit is one rule, rule:R-1, with two versions: @1 (the previous version, in force) and @2 (the new, revoking version — @2 supersedes @1). Two synthetic subjects declare evidence — an issuer registry on clock clk:issuer, and an agent runtime on clock clk:agent.

Nine declared evidence items describe issuance, declared effectiveness, reception, loading, and one action under the previous version (e07_action_under_v1). The scenario contains a deliberate gap: the application of the new version @2 by the agent runtime is never declared — application / agent_rt = not_declared. The experiment turns on that missing element.

Adding an explicitly declared relationship between two otherwise-unrelated clocks changed which temporal orderings the engine could derive — but it did not change the INDEX state, and it did not close the evidentiary gap.

The ordering e07_action_under_v1 after e06_eff_v2 in variant_B deserves particular care, because it is exactly the result a naive reader would over-interpret. On the observed record: the ordering exists only because the clock relationship was declared in the bundle; that relationship was not independently verified; after does not demonstrate a violation; the actual application of @2 is not demonstrated (the phase is not_declared, and absence of evidence is not evidence of absence); and no responsibility can be attributed on the basis of these results alone.

Each Lab evaluation record keeps two levels strictly apart, under the heading "Evidentiary status asserted by the Lab":

• They do not prove that the declared events actually happened — only that the evidence declares them.
• They do not verify the authority of the issuer, nor that any party was entitled to issue or revoke the rule.
• They do not treat a declared temporal order as a verified operational order: a declared clock relation is not verified synchronization.
• They do not read after as a violation, nor not_declared as proof that a phase did not occur — absence of evidence is not evidence of absence.
• They do not attribute responsibility or establish fault.
• They do not enforce, apply, block or authorize any agent operation.
• They do not infer, complete or interpret declarations that do not exist.

TRAER + INDEX answer a question the other EVIDE constructs do not: given the declared evidence, can an agent's rule transition be reconstructed, and how supportable is that reconstruction?

TRAER + INDEX are implemented as an isolated, experimental Governance Lab profile (v_traer_index_01): its own tables, no existing Lab file modified, no public endpoint. The profile is disabled by default; it was explicitly enabled by the authorized operator for the controlled synthetic experiment and remained enabled afterward, without authorization for non-synthetic participation. The evaluation engine is EVIDE TRAER 0.3.2, frozen, run off-server.

The results are bounded in explicit ways:

• data are entirely synthetic; running on the operational Lab does not make them real observations of an agent;
• only two controlled variants were run, one execution each — no statistical replication, no generalization claimed;
• no independent observation of an agent; the clock relation in variant_B is declared and not verified;
• no finding of compliance or violation; none can be derived from these results;
• independent human review of the profile and independent confirmation of the frozen 0.3.2 engine remain open;
• no competitive benchmark was executed; no claim of superiority over any other system is made.

The retained artifacts are intended to support independent reproduction using the frozen 0.3.2 engine: a third party holding the frozen package can recompute the bundle digest, regenerate the snapshot and assessment, and obtain the reproduced verdict, with re-import accepted only if the provenance gate binds the artifacts to the stored bundle and the approved evaluator. Independent third-party reproduction has not yet been demonstrated. The full artifact manifest, with SHA-256 digests, is published in the Working Paper.

EVIDE TRAER + INDEX Working Paper No. 1 (A/B experiment): app.certifywebcontent.com/working-papers — any signature or timestamp on the published document attests to its integrity and date, not to independent validation of the results.

EVIDE Governance Lab (experimental research platform): lab.certifywebcontent.com

EVIDE ANCHOR — related evidentiary construct: Declaring the operational perimeter before an agent acts

EVIDE MCP Server: MCP Server documentation

Evidentiary Continuity: Chaining responsibility across decisions