# E-013 — A path-based `pass` does not certify strategic direction

- Date: 2026-08-02
- Status: organic signal; low-sensitivity hypothesis pending hardening.

## Hierarchy and trajectory

- Parent objective: detect drift early through measurable alignment.
- Previous round: `R-SESSION-BETA`, closed and archived from a clean tree.
- Current round: `R-F4-PASSIVE`, for reproducible passive evidence before statistical intervention.

The transition was opened correctly with `adopt-baseline` and `next-round`: the previous contract and baseline were archived, the successor contract was written before any new work, and `round-transition` recorded both IDs. During the first dogfooding pass, a protocol note was added and a test that assumed every contract would have at least two KRs was fixed; the new round validly has just one KR.

## Expected and actual behavior

`align` was expected to verify the deterministic coherence of the transition. The run returned `pass`: it found a KR evidence path in the protocol note and emitted no warnings. However, the surface was three files, one outside the declared scope (`test/contract.test.mjs`), and the test passed only after fixing a structural assumption in the suite.

The closure returned `UNKNOWN`, not `PASS`: the research criterion and its obligations still lack declared checks or semantic evidence. This distinction was useful, but it does not compensate for the low sensitivity of `align`.

## Evidence

- `oguard align --json`: `status=pass`, `targeted_evidence_count=1`, `outside_scope_ratio=0.333333`, no warnings;
- `npm test`: 91 tests passed after fixing the single-KR test;
- `npm run round:end`: `UNKNOWN`, full verification but the `AC-PRIMARY` criterion and research obligations remain `unknown`;
- `npm run validate:docs`: 233 checks passed;
- local transition: `R-SESSION-BETA` contract and baseline archived by hash, `round-transition` event recorded in `.oguard`.

## Classification and limits

It is a **potential strategic false negative**, not a defect in the transition nor evidence that there was any bad intent. The guard only responded to the rules it currently has: the current contract, expected paths, and path-based evidence. Therefore, the `pass` means "it fits the current contract," not "the trajectory remains faithful to an independent plan."

A prolonged absence of warnings should be treated as possible low sensitivity until measured with a corpus and external adjudication; not as evidence of exceptional execution. Token savings cannot be inferred from this episode either.

## Correction of direction

The next hardening step must explicitly separate:

1. the frozen objective and initial plan, with a baseline external to the current edit;
2. append-only contract changes, justified and compared against that baseline;
3. semantic compliance evidence per criterion, not just path matching;
4. a specific alert when a contract change attempts to accommodate work already present.

The existing implementation partially covers (1), (2), and (4) for a single session through a frozen baseline, change events, and `contract-accommodation`. What is still missing is turning them into an immutable plan trajectory across rounds and, above all, modeling and verifying semantic evidence without pretending that a path proves it.

## Impact on the thesis

E-013 refutes the strong reading of a current `pass`. At the same time, it confirms that preserving the transition and the criteria's `UNKNOWN` avoids hiding uncertainty. The thesis becomes more precise: Objective Guard can give deterministic signals of contractual consistency, but it does not yet certify strategic direction; that capability requires comparing against frozen intent and falsifiable semantic evidence.

## Verifiable follow-up

- define an immutable initial-plan artifact and its provenance hash;
- add a declarative representation of semantic evidence per criterion and a reproducible verifier;
- create fixtures that distinguish a path match from semantic compliance;
- measure sensitivity, precision, and abstention rate on adjudicated data before enabling statistical alerts.