Risk Register

The central risk in reasoning is epistemic overreach: stable artifacts and fluent claims can make a bounded verification result appear stronger than it is. Every accepted claim must preserve the route from exact evidence bytes, through a governed plan and tool lifecycle, to the applicable verification policy and complete run bundle.

Claim Trust Path

flowchart TD
    bytes["governed evidence bytes"] --> span["span, digest, and source identity"]
    span --> plan["content-addressed plan"]
    plan --> events["tool and claim events"]
    events --> policy["verification policy"]
    policy --> bundle["manifested run bundle"]
    bundle --> reader["reader or downstream decision"]

    bytes -. "false or stale" .-> reader
    span -. "wrong support" .-> reader
    events -. "unsupported inference" .-> reader
    policy -. "weaker than assumed" .-> reader
    bundle -. "partial or altered" .-> reader

Persistent Risks

Hazard	Severity	Detection signal	Required control	Residual owner
citation names a source but does not bind exact support	critical	claim marker, evidence ID, span, snippet, and digest fail to reconcile	require byte-addressable evidence and verify the full support chain	reason workflow owner
correctly linked evidence is false, stale, or irrelevant	critical	independent source review conflicts with the archived evidence or its freshness requirement	record authority and observation time; require domain review and corroboration	source owner
derived claim outruns its supports	critical	claim contains material not entailed by extractive evidence or declared derivation	enforce support links and refusal; evaluate entailment and counterexamples externally	model or reasoner owner
omitted counterevidence produces a one-sided trace	critical	alternate retrieval or expert review surfaces material evidence absent from the plan	define corpus scope and counterevidence search; retain insufficiency as a valid outcome	decision owner
confidence is treated as calibrated probability	high	observed correctness by confidence bucket diverges from the stated value	label confidence semantics; calibrate against representative held-out data	evaluation owner
plan topology or tool lifecycle cannot justify the trace	high	cycles, missing tool returns, invalid ordering, or unbound evidence appear	enforce content-addressed DAG and event lifecycle invariants	executor owner
runtime or tool behavior changes behind stable-looking configuration	high	runtime, preset, provider, tool, or seed fingerprint differs	bind exact runtime descriptor and tool context to the trace	integration owner
permissive verification is summarized as an unqualified pass	critical	policy identity is absent or warnings disappear from the consumer view	retain policy and complete findings; require strict policy for strict claims	API consumer
artifact path escapes the governed run root	critical	normalized path resolves outside the root or manifest contains an unsafe target	normalize and contain paths; verify digest before access	deployment operator
run bundle is incomplete or concurrently interleaved	critical	manifest, trace, plan, report, and provenance IDs do not form one generation	isolate writers, stage the bundle, verify all core artifacts, then publish	artifact owner
snapshot replay is described as live reproduction	high	replay succeeds after live corpus or provider has changed	label snapshot replay; treat live revalidation as a new run	reviewer
proxy metrics are presented as semantic quality	high	metric label is reported without its implementation definition or relevance judgments	publish metric definitions and add domain-judged evaluation	evaluation owner
process-local rate or resource guard is treated as deployment isolation	high	multiple workers exceed aggregate policy or remote tools exceed local accounting	enforce quotas, sandboxing, and egress at the hosting boundary	service operator

Claim Release Gate

flowchart LR
    claim["candidate claim"] --> support{"exact support valid?"}
    support -->|no| refuse["refuse release"]
    support -->|yes| policy{"required policy passes?"}
    policy -->|no| refuse
    policy -->|yes| bundle{"bundle complete and manifested?"}
    bundle -->|no| refuse
    bundle -->|yes| adequacy{"domain adequacy reviewed?"}
    adequacy -->|required and no| hold["hold for review"]
    adequacy -->|yes or not required| release["release with scope and evidence"]

Release the verification report and evidence scope with the claim. Never reduce a strict report to a Boolean that hides which checks applied, which findings were filtered, or whether the conclusion is a refusal.

Evidence Required By Change

• Evidence or claim changes require pass and tamper cases for IDs, spans, digests, citation markers, derivations, insufficiency, and unsupported claims.
• Planner, tool, or executor changes require topology, lifecycle, failure, deterministic identity, runtime drift, and replay scenarios.
• Verification-policy changes require strict, permissive, and audit comparisons with the complete findings retained.
• Artifact changes require missing-file, path-escape, checksum, concurrent-writer, incomplete-generation, schema, and manifest tests.
• Retrieval or evaluation changes require judged corpus evidence; structural trace metrics alone are insufficient to support a quality claim.

Verification narrows what can go wrong inside the declared model. Source authority, freshness, domain interpretation, provider governance, privacy, and the threshold for acting remain outside that model and must be explicit in the deployment decision.

See known limitations for unsupported claims and architecture risks for failure mechanisms.