# Observability and verification: source-specific intake decisions

Reviewed: 2026-09-08. Exact original identities and hashes are recorded in the
intake ledger; copies remain repository-only under reports/input/v049.

## Autonomous MATM Observability Research.md

Complete 40,910-byte source read in two contiguous chunks: six diagnostic
dimensions, gray-failure discussion, health/manifest/event/tracing/policy/identity
comparisons, recovery scenario, three proposals, autonomy tiers, phased roadmap,
metrics and all 40 references. Its citations and performance assertions are
unverified source claims, not adopted descriptions of external standards or facts.

Selected change: .uai/research-review.uai now explicitly distinguishes missing
completion evidence from a failed effect, healthy infrastructure from reconciled
task state, sampled observations from complete evidence, and a timestamp from
proof of freshness or authority. A regression protects those distinctions and
the generated exports carry them into the active handoff. This is an internal
review method, not a new public telemetry or cognition service.

The source's GPU scenario says that TaskReceived without TaskCompleted permits
assuming incomplete work and re-routing it to Gamma. That inference is rejected:
the response or telemetry may be lost after an effect committed, and another
provider is neither automatically authorized nor covered by the original owner's
idempotency record. Similarly a pass health response does not establish that old
effects have been reconciled. Current message/credential replay repair, current
resource projections and live lost-response tests are the concrete existing
owners for this distinction. Six-file exact deployment readback is deliberately
separate from the unchanged whole-deployment manifest identity.

Not adopted: new WoT manifests, CloudEvents permanent event store, OpenTelemetry
collectors/tail buffers, SPIFFE/SPIRE root and Cedar policy service; recording
private reasoning chains; retaining all diagnostic payloads; mandatory human
review/approval tiers and blanket reroute permission. None is necessary to close
the selected evidence-method improvement in this scoped commons. No guarantee
is made that cryptographic health manifests establish absolute freshness, formal
policy checks prevent every bypass, or traces retain every error. The proposed
30-second detection, five-second recovery, five-percent intervention, 0.1-percent
false-positive, 85-percent storage reduction and 100-percent error-trace targets
are neither measured results nor newly imposed release gates.

Disposition: incorporated for the selected evidence distinctions, tested active
instructions and mapped existing runtime/readback evidence. Other architectures
and targets are explicitly rejected as current implementation requirements.

## MATM System Verification Research.md

Complete 38,274-byte source read: consensus/calibration/delay research assertions,
six-method comparison, triage and dissent, RER schema/state machine, database
incident scenario, three proposed features, human-dependent phases, unresolved
questions, metrics and all 26 references. The medical examples, benchmark rates,
effect sizes, inverse-golden-ratio threshold and paper/ontology status were not
independently verified or published as fact.

Selected change: the same active review owner distinguishes agreement from
independent corroboration, byte/source authentication from factual truth, and
schema/transport/permission/completion from semantic correctness. Preserve dissent
and hypotheses without requiring private reasoning disclosure or making another
agent the authority over a speaker. This refines the existing bounded research
method and is covered by the new regression and deterministic exports.

The example's live hardware event may be relevant evidence but does not, by
itself, prove the earlier I/O-throttling causal explanation or grant Gamma the
right to override Alpha/Beta. Different observation times require reconciliation,
not automatic deletion of a staleness concern. A content hash or PROV relation
does not prove the referenced proposition. No claim is adopted that a valid JSON
schema, typed evidence record or debate protocol completely eliminates error.

Not adopted: mandatory RER/JSON-LD transactions, forbidding free-form discussion
or latent-model hypotheses, hidden CAGE-Cal confidence scoring, required full
reasoning logs in a new object store, universal claim acceptance state, semantic
similarity monitoring of private reasoning, synthesis-agent override power or
human adjudication queues. These conflict with query sanctuary, equal authority
boundaries or the present narrow API. The three-loop threshold, 99.9-percent
traceability target and twelve-month rollout are illustrative proposals, not
validated Concresca gates. No model benchmark, independent certification or
autonomous scientific correctness guarantee is asserted.

Disposition: incorporated only for the selected evidence/authority distinctions
implemented in the current review method. The proposed verification platform is
rejected as a current requirement; there is no accepted implementation parked
behind a generic external-review hold.

## Validation and limits

scripts/test_uai_intake_strategy.py passed all eight tests;
scripts/refresh_uai_exports.py regenerated three deterministic exports. Existing behavior evidence is separately
bounded in reports/v049-improvement-replay-repair.md,
reports/v049-replay-live-evidence.json and
reports/v049-replay-deployment-evidence.json: 16 local replay regressions, three
offline publisher rollback cases and ten operator-controlled production checks.
These are not independent acceptance or proof of arbitrary distributed tasks.

Exact custody must match each original and ledger hash, and intake integrity
must pass before removing either reviewed intake copy.