Ownership Boundary

Reason authority is the relationship between evidence and explicit claims. It owns how support is represented and verified, while leaving evidence retrieval, role scheduling, and final workflow acceptance to their respective layers.

flowchart TD
    change{"Which record changes?"}
    representation["source/chunk record"]
    retrieval["execution artifact/ranking"]
    claim["claim/support/finding"]
    workflow["role transition/convergence"]
    authority["flow verdict/replay acceptance"]

    change --> representation --> ingest["ingest"]
    change --> retrieval --> index["index"]
    change --> claim --> reason["reason"]
    change --> workflow --> agent["agent"]
    change --> authority --> runtime["runtime"]

Decision table

Change	Owner	Reason
change chunk normalization or byte mapping	ingest	changes evidence representation before reasoning
change ANN candidates, scoring, or replay bounds	index	changes retrieval execution rather than interpretation
add a check that derived claims have valid exact support	reason	changes grounding and verification semantics
add a planner node kind or content-identity input	reason	changes the reasoning record itself
repeat critique after an unsatisfactory verifier role	agent	changes role lifecycle and workflow control
reject the entire flow when reasoning is non-certifiable	runtime	changes end-to-end acceptance policy

Index-to-reason handoff

Reason consumes identified evidence and retrieval provenance. It may retain a pinned local corpus/index for the reference workflow, but it does not reinterpret the index package's score as support. Support is established separately through an evidence identity, exact byte interval, and snippet hash.

This distinction prevents a high-ranked passage from becoming an accepted claim merely because retrieval considered it similar.

Reason-to-agent handoff

Reason produces a self-contained plan, trace, claim set, verification report, and manifested run. Agent may schedule reasoning-related roles and preserve their outputs, but it must not rewrite claim status or hide findings inside workflow summaries. Conversely, reason does not decide whether another role should run or whether convergence has occurred.

Reason-to-runtime handoff

Runtime can apply flow policy to reasoning evidence. It may accept, reject, or mark a run non-certifiable, but the underlying verification findings remain unchanged. Arbitration over a complete flow is not a second reason verifier.

Three decisions that must remain separate

Decision	Owner	Required record	Does not decide
which material ranks for a query	index	request, artifact, backend, metric, score, rank and provenance	whether bytes support a claim
whether a claim is supported and passes checks	reason	claim type/status, exact supports, checks, findings, trace and manifest	whether another role runs or the whole flow is accepted
whether a complete flow is admissible	runtime	manifest, policy, lower-layer evidence, arbitration and replay envelope	what a reason finding originally observed

A high retrieval score cannot validate a claim. A validated claim cannot schedule another role. A runtime acceptance cannot rewrite a rejected claim. These are distinct authorities even when one application displays them in a single response.

Minimum reasoning handoff

Reason publishes enough evidence for agent and runtime to preserve, inspect and arbitrate its work:

Handoff field	Custody purpose
problem specification and content identity	identifies the question, constraints and expected output
plan DAG and node identities	exposes intended derivation and verification order
runtime/tool/retrieval descriptors and call records	identifies external execution and evidence acquisition
evidence records and exact content digests/spans	anchors supports to inspectable bytes
claims with type, status, confidence and support references	preserves the epistemic object being reviewed
all checks, findings and unavailable/failure outcomes	prevents a summary from hiding verification coverage
semantic trace, byte fingerprint and invariant checksum	binds ordered reasoning events and integrity evidence
run manifest and replay comparison	binds files into one run and records later divergence

Agent may embed these records in role inputs or trace entries. Runtime may apply flow policy. Neither may replace the packet with final prose and still claim to preserve reasoning custody.

Route a broken handoff

First false invariant	Owner
wrong source/chunk bytes or missing preparation mapping	ingest
wrong backend/rank/provenance for immutable evidence	index
invalid support span, changed evidence digest, wrong claim status or missing finding	reason
reasoning packet is correct but role order, merge, veto or trace projection is wrong	agent
packet is intact but tenant policy, certification or replay acceptance is wrong	runtime

The package that detects the break retains it visibly, but ownership follows the first false semantic record.

Ownership test

Locate the false invariant. Ranking and backend provenance point to index. Claim kind, support linkage, evidence bytes, check outcomes, and reasoning-run integrity point to reason. Role order points to agent. Final authority points to runtime.