verify

name	verify
description	Define rover Done criteria and gather evidence against each criterion.
user-invocable	true
argument-hint	[--propose <loop-file> | <loop-file> | free text]
effort	high

Autonomous Verify

Evidence discipline for autonomous work. The operator is not watching, so every claim of progress or completion has to rest on something the operator could check without being present.

Two disciplines, one skill:

1. Write Done criteria up front. Before the rover starts building, it commits to what "done" looks like in concrete terms.
2. Gather evidence against those criteria. Actively verify each criterion by running, observing, or measuring. Report what is proven, what is unverified, what failed.

Why this matters for an autonomous rover

Without Done criteria, the rover has no endpoint. It stops when it is tired, when tests happen to be green, or when the phase machine nudges it along. None of that is "finished." The operator reading the loop file later has no way to audit whether the work actually matches the original goal.

Without active evidence, the rover coasts on proxies: "CI green," "the code compiles," "the test file exists." None of these prove the feature does what the user asked for. Proxies are the autonomous-work equivalent of driving at night without headlights.

Mode 1: propose Done criteria (--propose)

Invocation: /rover:verify --propose <loop-file> or called by rover at end of SURVEY.

1. Read the loop file's ## Dispatch, ## Context, and ## Plan sections. Dispatch is the source of truth; Context is interpretation; Plan is the proposed deliverable.
2. Plan-vs-Dispatch check (mandatory). Compare the action verbs in Dispatch against the deliverables described in Plan. If Dispatch contains action verbs (build, ship, fix, port, install, deliver, enable, make work, implement) and the Plan only describes producing research, documentation, or analysis without naming the actual implementation as the deliverable, refuse to generate Done criteria. Log the mismatch, send the rover back to SURVEY to rewrite the Plan so it includes the actual implementation, and only then return to Done-criteria generation. The rover does not ask the operator to confirm research-only scope; the Dispatch's verbs are the answer. Always compare against Dispatch, never against Context alone; Context can itself have shrunk the dispatch. 2a. Success-signal floor (mandatory). The action-verb check above guards against the Dispatch shrinking. This step guards the opposite failure: a deliverable that faithfully executes the Dispatch's verbs yet misses the operator's actual target, because nothing transcribed how the operator recognises the result as working. Those recognition tests become Done criteria here at SURVEY, where the no-fate-2 floor already lives (see "Done criteria are not findings"), so a later finding that the deliverable fails one of them cannot be cost-value-skipped. Two sources, in priority order:
   • State-pair distinguishability (the mechanical base, phrasing-independent). If the Dispatch names two or more user-visible states, modes, or statuses (recording vs idle, loading vs ready, valid vs error, connected vs lost), write a Done criterion for EVERY pair: "state A and state B are distinguishable at a glance, not only by the colour, opacity, or animation of a shared mark." Range over every pair; never narrow a multi-state signal down to the subset that is easiest to satisfy (a criterion that covers three easy states while the operator's real concern was the fourth pair is exactly the narrowing this forbids). This base fires regardless of which words the operator used.
   • Stated success-signals (recall booster on top). Scan the Dispatch for wording where the operator states how the result reads as working: glance / at a glance / without inference / distinct / obvious / unmistakable / tell-apart / can't-confuse / "see which X we are in" / "must feel different" / "must be visible". Transcribe each verbatim into a binary, observable criterion phrased as the operator phrased it. This keyword list raises recall; it is not the floor. A success-signal phrased without any of these words is still caught by the state-pair base when it concerns states; the base, not the list, is the structural guarantee.
   • Auditable coverage (anti-under-transcription). Record under the criteria: the raw operator-goal quote, the success-signals extracted, the state-pair matrix, and an explicit "not extracted, because…" line for any nearby intent language deliberately left out. Under-transcribing to dodge a binding floor later must be visible on the operator's read, never silent. And "not transcribed into a criterion" must NEVER be treated downstream as "out of intended scope" when the original Dispatch text carries the intent; that only moves target-drift from the Dispatch into SURVEY.

Good vs bad criteria

Bad	Good
"The settings page works"	"GET /settings returns HTTP 200 when logged in as a user"
"Form validation is correct"	"Submitting the form with an empty name shows an inline error next to the name field, no DB write happens"
"Fast enough"	"/settings first-paint under 200ms at p95 on the staging box"
"Tests pass"	"bin/rspec spec/settings_spec.rb exits 0 with 12 examples, 0 failures"
"Looks good"	"Screenshot at 1440x900 shows form below header, no horizontal scroll, no overlap with nav"

Vague criteria are how missions drift. The rover catches this up front by insisting on sharp ones.

Mode 2: gather evidence (default)

Invocation: /rover:verify <loop-file>, or bare /rover:verify in a session where a loop file is obvious, or called by rover at end of INSPECT.

1. Read the ## Done criteria section.
2. For each criterion, determine the verification tactic:

Criterion type	Tactic	Evidence
Code runs	Execute it (unit test, script, CLI)	Command output, exit code
HTTP endpoint	curl -sSv with expected inputs	Status code, response body
UI element	Navigate and screenshot	PNG path + described observation
UI state change	Before and after screenshots	Two PNGs, described diff
DB mutation	Query the DB after the action	Row count, specific field values
File on disk	ls / stat / cat	Path + relevant content
Third-party integration	Trigger it, inspect the other side	Log entry, API response, webhook payload
Logged behavior	Run, grep the log	Matched log line with timestamp

3. For each criterion, attach the evidence under its row in the loop file. Format:

## Done criteria

- [x] `GET /settings` returns 200 when logged in
      Evidence: `curl -s -o /dev/null -w "%{http_code}" http://app.test/settings` → 200 at 10:14
- [x] Submitting empty name shows inline error
      Evidence: screenshot `/tmp/settings-empty-name.png`, error visible next to name field
- [ ] p95 first-paint under 200ms
      Unverified: no staging box available in dev. Ran locally only: 180ms average over 5 runs.
- [ ] Tests pass
      Failed: `bin/rspec` → 12 examples, 1 failure. See log.

4. Report a summary:
   • Criteria met with evidence
   • Criteria unverified (why, what would be needed to verify)
   • Criteria failed (what the evidence showed, next DRIVE target)

Mode 3: standalone use

Outside a rover session, /rover:verify "free text describing the work" lets a user ask "did I really finish what I claimed?" for any recent work. The skill:

1. Asks for or infers a short Done-criteria list from the free text and the recent diff
2. Gathers evidence
3. Reports

This is useful after any non-trivial task, not just autonomous loops.

Proxies are not verification

The rover is tempted by proxies because they are cheaper than real evidence. Reject them:

Proxy	Why it fails	Real verification
"CI is green"	Tests a file I cannot run	Run the relevant scenario locally
"The code compiles"	Syntax check, not behavior	Execute the code path
"The test file exists"	File existence, not passage	Run the test, see it pass
"Curl returns 200"	Status code, not content	Assert on response body
"The feature should work"	Unchecked hypothesis	Run the feature end-to-end
"I followed the pattern"	Pattern imitation, not correctness	Prove this instance behaves
"No errors in the console"	Absence is not presence	Confirm the positive outcome happened
"The pipeline works on one example, so the matrix criterion is met"	Single-instance evidence does not satisfy a quantitative criterion	Run the full matrix, produce every output the criterion names, list each one
"The simulator boot loop would burn context, so I tick the criterion now"	Cost is a fate-2 candidate for findings; Done criteria have no fate-2 path	Either produce the full evidence, or mark unverified: <route the operator needs to take> and stay in DRIVE

When a proxy is the only thing available, label it: unverified, only proxy evidence: <X>. Honesty beats a false green.

When verification is genuinely impossible

Sometimes a criterion cannot be verified with available tools:

• Production-only behavior (auth, third-party services, real user load)
• Visual checks in a headless environment without a browser
• Timing-sensitive measurements without a stable baseline
• Behaviors that only manifest under rare conditions

Mark these unverified: <specific reason> in the Done criteria. Do not claim them met. The rover does not decide on its own that an unverified criterion is "acceptable" and push on. The rover also does not route unverified criteria to the operator; there is no operator-accept path inside an autonomous mission. The rover's job is to make every criterion verifiable and produce the evidence.

Done criteria are not findings. The three-fates rubric from rover applies to findings raised by pride, gurus, end-user, and technical passes; it does not apply here. Done criteria are the mission's destination, not weights along the way. There is no fate-2 cost-value-skip path for a Done criterion: every criterion is either met with evidence, or unverified (which blocks STOW until the rover produces a verification route). Fate 3 (reject-as-non-issue) does not apply either: a Done criterion was committed to during SURVEY and proves the mission delivered what the Dispatch asked for, so it cannot retroactively be a non-issue.

Late-discovered success-signal contradictions are unmet criteria, not findings. A success-signal floor (step 2a) only works if SURVEY did not have to be perfect. When any INSPECT pass surfaces that the deliverable contradicts a success-signal the Dispatch stated in its own words, or fails a state-pair distinguishability that the Dispatch's named states imply, treat it as an unmet Done criterion discovered late, even one SURVEY failed to transcribe: no fate-2, no canon-vragen, back to DRIVE until the contradiction is gone. The test is objective and anchored to the original text, not to the already-committed criteria list: quote the Dispatch phrase (or name the state-pair) the artefact contradicts. This is the same anti-shrink floor as the Plan-vs-Dispatch check, extended from action-verb decay to success-signal contradiction, so missing one phrase at SURVEY does not delete the floor. The guard cuts the other way too: a late-discovered contradiction is only a floor when the original Dispatch text carries the intent; the rover does not invent a success-signal the operator never stated in order to manufacture rework.

Unverified blocks STOW, not INSPECT activity

When INSPECT runs, any criterion still marked unverified blocks the transition from INSPECT to STOW. It does not freeze the rover. Other criteria keep verifying, other pride findings keep getting fixed, other parallel work continues; the blocker is only on closing out the mission. The rover's legitimate move on the unverified item is:

1. Return to DRIVE and provide the evidence. If the direct evidence is out of reach with the current setup, change the setup. Stand up a local harness, seed a test database, mock the third-party with a fixture that exercises the real code path, run a headless browser under Playwright to screenshot a UI that was previously "headless-only", instrument the code to emit a log line the criterion can assert against. Log each attempt explicitly. "I could not run the staging test" is not acceptable until "I tried A, B, and C, and here is why each failed, so I built D, which produced the evidence" has been logged. Unverified is not a final state for the rover; it is an assignment back to DRIVE.

The rover never upgrades unverified to a tick on its own, never silently drops a criterion from the list, never reasons "in de praktijk zal dit wel werken" / "in practice this should be fine" to close it out, and never asks the operator to accept an unverified criterion. Those moves are the exact corner-cutting this discipline exists to prevent. If you catch yourself typing "accepting unverified" or "acceptable given context" into the loop file, revert and go back to DRIVE to produce the evidence.

Anti-patterns

Thought	What it actually is
"I'll verify at the end"	No you will not, you will run out of steam. Verify per change.
"This is too obvious to verify"	Obvious things fail too. 30 seconds of evidence beats 5 minutes of debugging later.
"Verification would take forever"	If verifying the mission takes longer than doing it, the mission is probably too big. Split.
"The tests cover it"	Tests are a form of verification, but rarely the full criterion. What did the user see?
"I already checked"	Show the evidence. If you cannot cite it, you did not check.

Quantitative criteria need quantitative evidence

A Done criterion that names a count or a matrix (bin/capture-shots produces 40 PNGs, all 10 scenes render, both locales receive a preview, every endpoint returns 200) is met only when every named instance is produced and listed. Proving the mechanism on one instance proves the mechanism; it does not prove the criterion. The two are not the same.

The temptation is structural: when the matrix is wide (40 outputs, hours of simulator boots, a long sweep) and the mechanism is verified on one example, the rover wants to upgrade "mechanism proven" to "criterion met" because the marginal cost of the remaining N-1 runs feels disproportionate to the marginal information. That trade is wrong twice. First, the criterion was committed to during SURVEY precisely because the matrix matters; the rover does not get to redefine the criterion at INSPECT to be the part it already finished. Second, the rover routinely discovers that the mechanism works on the first instance but fails on the second or third (a scene flag the schema does not expose, a locale-specific font fallback, a device-class path the mechanism hardcoded). The matrix runs are the discovery surface, not the formality.

If the matrix is genuinely too expensive to run in this session, the move is unverified: requires <concrete-route> with a route the operator could take, not a tick. Unverified blocks STOW, so the mission stays open and the operator sees the gap in the next read of the loop file. A ticked criterion that was not actually verified is worse than an unverified one: the operator stops looking, the gap rots, and the next reader of the communiqué believes a lie.

Red flag: when typing a verify-pass log entry, watch for the construction "running N more iterations would exceed prudent context; the one-scene verification proves the pipeline works." That sentence shape is the proxy this section exists to prevent. The correct continuation is unverified: needs full matrix run via bin/X. Then stay in DRIVE until the matrix evidence exists, or transition out only because the operator's environment legitimately blocks the run and the unverified state is documented for them to pick up.

Interaction with other skills

• decide picks which path to take; verify proves the chosen path worked. Complementary.
• pride asks "would the user hate this?" (contrarian, smell-finding); verify asks "did this do the thing?" (evidence-gathering). Different questions, no overlap. The rover runs both before declaring done.
• rover invokes verify --propose at end of SURVEY and verify (default) at end of INSPECT. A rover mission without Done criteria is not started; a rover mission without ticked criteria is not finished.