# E-035 — A long multi-agent round exposed traceability gaps

## Context and hierarchy

- Date: 2026-08-05.
- Main objective: preserve the parent objective and evidence across long agentic development.
- Development context: a large effort progressed through definition, implementation, review, and correction with multiple agents and operational telemetry.
- Product question: whether Objective Guard preserved intent while explaining who changed what, why closure failed, and how each criterion advanced.

## Observed trajectory and evidence

Operational feedback reported that `task-flow`, `align`, and `round-end` kept the objective, scope, non-goals, and criteria visible across many rounds and agents. It also reported repeated contract-file rewrites at round transitions, manual correlation between execution and governance traces, noisy global checks with pre-existing failures, ambiguous closure outcomes, glob crossings during legitimate cross-layer fixes, and no cumulative criterion-to-commit view.

Available telemetry proves that the guard commands ran and that their decisions were recorded. It does not prove which agent action advanced a criterion, whether an infrastructure failure was introduced by the round, or whether a `pass` represented complete evidence.

## Expected and actual guard behavior

Expected: a stable parent objective governs inherited execution units; each unit attaches evidence; closure distinguishes newly introduced failures, baseline debt, unavailable infrastructure, missing evidence, and contradiction; the final artifact reconstructs requirement → change → proof → result → commit.

Actual: the objective remained visible, but round identity was coupled to a versioned contract rewrite and evidence had to be correlated manually with orchestration records. Existing checks and scope crossings could create noise without enough provenance to classify the outcome.

## Classification, correction, and limit

This is an organic protocol and traceability signal, not a confirmed detector false positive or false negative. The product contract now declares a future split between durable governance and local operational state, native inherited subrounds, explicit closure categories, cumulative criterion status, baseline fingerprints, precise preflight invalidation, and generic external evidence IDs.

Orca integration is deliberately deferred. The correction defines an agnostic protocol first; it does not claim that the current implementation supports these capabilities.

## Impact on the thesis and follow-up

The episode supports the thesis that explicit objectives and staged checks help multi-agent continuity. It weakens any stronger claim that the present round model already provides end-to-end causal traceability.

Follow-up is verifiable when two parallel subrounds inherit one parent objective, preserve conflicting evidence by criterion, distinguish baseline and infrastructure outcomes at close, and emit a reproducible requirement-to-commit matrix without rewriting the durable objective. A later adapter may prove Orca correlation separately.