Interpreting Reasoning Evidence

Review a reasoning result as a chain of custody from problem and plan to exact evidence, inference, verification, and manifested run. Final prose is a projection of that chain, not a substitute for it.

flowchart LR
    problem["ProblemSpec"] --> plan["content-addressed plan"]
    plan --> events["typed execution events"]
    evidence["retained evidence bytes"] --> support["exact SupportRef"]
    events --> claim["typed claim"]
    support --> claim
    claim --> findings["verification findings"]
    findings --> run["fingerprint + manifest"]

Follow One Claim

Review question	Evidence to inspect	What remains unproven
What was the system asked to establish?	ProblemSpec, constraints, expected output, stable identity	whether the problem formulation is complete
Why did this reasoning action occur?	plan node, dependencies, ordered trace event	whether the plan is sufficient for the domain
Which source content was available?	EvidenceRef, retained bytes, source identity and digest	source authority, freshness, and completeness
Which bytes support the statement?	SupportRef span, snippet digest, support kind	that the inference from those bytes is valid
What kind of statement is it?	observed, assumed, or derived claim kind and status	whether confidence is calibrated
Which checks evaluated it?	complete verification findings and policy disposition	defects outside the registered checks
Is the run internally complete?	plan, trace, verification report, fingerprint, metadata, manifest	real-world truth or generalization
What did replay establish?	frozen inputs, invariant checksum, trace comparison and diff	current external source or provider behavior

Bounded Reasoning Vocabulary

Claim	Required evidence	Bound on the claim
structurally valid trace	supported header, ordered events, complete lifecycle, linked calls and returns	says nothing about source truth
grounded claim	retained span, snippet digest, evidence identity, and support edge	does not establish source authority or entailment by itself
verified run	complete registered checks and explicit policy disposition	covers declared checks, not every possible defect
reproducible run	specification, preset, seed, runtime fingerprint, canonical files, and matching replay	applies to frozen recorded inputs and results
successful replay	invariant checksum and trace comparison pass over retained artifacts	is not a fresh call to tools or sources
evaluated behavior	named corpus, cases, constraints, expected refusals, definitions, and metrics	does not generalize beyond represented cases
confident claim	explicit confidence field linked to the claim and support	is not automatically a calibrated probability

Distinguish Linkage, Grounding, And Truth

A content digest proves identity. A support reference proves that a claim is linked to exact bytes under the recorded support kind. A verification report proves that registered checks reached their recorded findings. None alone proves that the source is correct, that contrary evidence is absent, or that a domain expert should act on the conclusion.

insufficient_evidence, unsupported capability, provenance drift, rejected claims, and verification failures are evidence-bearing outcomes. Do not remove them when projecting a run into a summary. The local extractive reasoner and BM25 path are inspectable references, not claims of general reasoning ability or state-of-the-art retrieval.

Preserve The Review Bundle

Retain the specification, plan, runtime descriptor, evidence bytes and source metadata, claims, support references, tool calls and results, typed trace, verification report, fingerprint, run metadata, and manifest. A citation label without the bytes, a hash without a retrievable object, or prose without claim identity breaks later review even if the output still looks plausible.

Continue with invariants for enforced reasoning laws, known limitations for epistemic and operational bounds, and the risk register for custody and interpretation risks.