Runner Fidelity Verdict v0¶

Status: internal derived helper contract. This page defines the assay.runner.fidelity_verdict.v0 vocabulary and claim gate derived from assay.runner.observation_health.v0. It does not wire the verdict into Runner archives, CLI output, Trust Basis, or any stable public report schema.

Purpose¶

Runner observation health already records the raw measurement-health signals:

• kernel_layer
• ringbuf_drops
• policy_layer
• sdk_layer
• cgroup_correlation
• network_protocol_coverage
• network_endpoint_claim_scope

Those fields are the source of truth. The fidelity verdict is a small derived read-model that answers one narrower question:

Which Runner measured-effect claims may downstream projection/report code
interpret from this observation health record?

It is a gate before interpretation, not a new capture layer.

Schema String¶

assay.runner.fidelity_verdict.v0

The current Rust helper lives in assay-runner-schema. The schema string is an internal helper/contract label at this stage; no JSON Schema sidecar or archive member is frozen yet.

Vocabulary Boundary¶

Do not confuse this verdict with the experiment-scoped observability calibration agreement vocabulary.

The calibration vocabulary uses match and drift because it compares a requested target with an observed artifact. Runner fidelity deliberately does not use those words: a health record can be clean without proving the workload content "matched" anything, and drift is a cross-state comparison, not per-run measurement quality.

The shared words clipped, failed, and not_applicable are allowed because they describe measurement states in both contexts. Consumers must still use the schema namespace to decide which semantics apply.

Verdicts¶

Claim Gate¶

The helper emits a claim_gate with explicit decisions for every verdict:

failed blocks claims authorized by this fidelity verdict. It does not erase separately validated SDK, trace, or external receipt artifacts; it means this Runner health record cannot safely authorize those claims.

The load-bearing rule is:

ringbuf_drops > 0 => clipped => bounded_negative_claims = blocked

Observed positive events can remain useful under clipped; missing events cannot prove absence. This is why clipped degrades positive measured claims but blocks bounded negative claims.

clean is capture-health clean, not protocol-universal. A record may be clean while still declaring network_protocol_coverage = absent and network_endpoint_claim_scope = not_applicable when no network protocol events were observed. A record with observed connect activity may instead declare network_protocol_coverage = connect_only and network_endpoint_claim_scope = diagnostic_only; downstream code must not stretch that into an exact QUIC/datagram peer-set claim. When Runner observes sendto or sendmsg destination sockaddr events it may upgrade network_protocol_coverage to datagram_peer_observed or connect_and_datagram_peer_observed; that is a stronger transport signal, but still not a request-level or exact peer-set binding while network_endpoint_claim_scope = diagnostic_only. If network events may have been dropped before any network event was emitted, coverage is unknown.

Composition With Projection claim_level¶

Projection helpers such as path projection carry claim_level values like:

• raw_observed
• projected_equivalent
• inconclusive

The fidelity verdict does not replace that vocabulary. It gates which projection claim levels may be interpreted as measured-effect claims.

Initial composition rule:

If a projection or report wants to make a bounded-negative claim, it must also consult claim_gate.bounded_negative_claims. If it wants to bind a claim to a specific run/cgroup/tool identity, it must also consult claim_gate.per_binding_claims.

This keeps claim_gate as a guardrail over existing claim levels instead of introducing a second independent claim hierarchy.

Composition With Coverage Descriptors¶

fidelity_verdict.v0 gates capture health. coverage_descriptor.v0 gates the per-dimension visibility ceiling. Consumers must apply both when interpreting measured side effects.

Example:

• A clean health record may allow bounded_negative_claims at the fidelity layer.
• A filesystem descriptor with completeness = open_syscall_only still blocks a bounded-negative file-activity claim because io_uring or mmap-backed writes can hide effects from the capture method.

So the combined decision is the stricter of:

fidelity claim gate
coverage descriptor claim-kind decision

This keeps positive observed effects useful while preventing a clean capture-health verdict from becoming a universal absence or completeness claim.

Derived Shape¶

Example clean verdict:

{
  "schema": "assay.runner.fidelity_verdict.v0",
  "source_schema": "assay.runner.observation_health.v0",
  "run_id": "run-001",
  "verdict": "clean",
  "claim_gate": {
    "reported_claims": "allowed",
    "measured_positive_claims": "allowed",
    "bounded_negative_claims": "allowed",
    "per_binding_claims": "allowed"
  },
  "reasons": [
    {
      "field": "observation_health",
      "observed": "clean",
      "rule": "complete_kernel_layer_zero_drops_clean_cgroup_correlation"
    }
  ],
  "non_claims": [
    "fidelity_no_observation_health_replacement",
    "fidelity_no_policy_correctness_verdict",
    "fidelity_no_runtime_safety_verdict",
    "fidelity_no_agent_quality_score",
    "fidelity_no_probabilistic_confidence_score"
  ]
}

Classification Rules¶

The v0 helper derives the verdict only from one ObservationHealth value:

1. Invalid observation-health records or cgroup_correlation = failed produce failed.
2. Non-Linux platforms or absent kernel layers produce not_applicable.
3. ringbuf_drops > 0 or kernel_layer = partial_ringbuf_drops produces clipped.
4. cgroup_correlation = partial produces correlation_partial.
5. kernel_layer = complete, ringbuf_drops = 0, and cgroup_correlation = clean produces clean.

The helper intentionally does not upgrade or reinterpret network_protocol_coverage or network_endpoint_claim_scope. Those fields remain an explicit honesty boundary on top of the capture-health verdict.

If more than one degradation is present, the helper preserves specific reasons and applies the stricter gate where needed. For example, partial cgroup correlation blocks per-binding claims even when the top-level verdict is driven by clipping.

Non-Claims¶

• The verdict does not replace observation_health.v0.
• The verdict does not prove policy correctness.
• The verdict does not prove runtime safety.
• The verdict does not score agent quality.
• The verdict does not provide probabilistic confidence.
• The verdict does not validate reported traces, spans, tool calls, or SDK events as true.
• The verdict does not convert connect-only network capture into an exact datagram/QUIC peer-binding claim.

Wiring Boundary¶

This slice adds only the contract and helper. It intentionally does not:

• add a Runner archive member;
• add CLI output;
• add Trust Basis claims;
• add capability-diff gating;
• add cross-runtime report wiring;
• add path/network projection Slice 2 or Slice 3.

Report wiring should wait for a concrete consumer or review surface that needs the derived verdict.