Hoo verify outcome

Verify an outcome (recompute first, read the claim second)

Takes an item slug or a path to a record in outcomes/. The job is to recompute the number from raw data and only then look at what the record claimed, so the recomputation cannot be steered by the answer.

Read-only over the brain except for the run report and the one signal in step 5. Nothing here changes an item's status, edits an outcome record, or touches the claimed numbers.

1. Identity, URL, windows - and nothing else

PYTHONPATH="$CLAUDE_PLUGIN_ROOT/lib" python3 -c into core.outcomes.parse_outcome(<path>), and use ONLY these three from what comes back:

• item - which item this is
• url - the page to pull
• windows - the before and after date ranges the claim was measured over

Do not read claimed_before, claimed_after, or claimed_delta yet, and do not print them. This is the whole exercise. A number you have already seen is a number you will unconsciously reproduce: the pull gets framed around it, a mismatch gets re-run "to check", and the verification becomes a confirmation. Steps 2 and 3 happen with the claim unread.

No windows in the record: derive them from applied_at (the 28 days after, against the same 28 days before) and say in the report that the windows were derived, not read. No applied_at either - stop and report the record as unverifiable, naming the missing key. Do not invent a window.

An item slug rather than a path: resolve it to outcomes/<item-id>.md in the active brain (registry.get_active()). No such record - say so and stop; there is nothing to verify.

2. Pull the raw windows independently

Through the configured capability slots (ADR-0009), not from anything the brain already stored:

• search-data (GSC for most sites): clicks, impressions, CTR, and average position for url over each window.
• analytics (GA4 for most sites), when configured: sessions for url over the same windows.

Never substitute a stored signal line for a fresh pull. A signal is the brain's own prior reading, so checking a claim against it verifies nothing.

Connector unreachable, or the window falls outside its retention: the claim is unverified. Say that, name which connector and why, and stop. An unverified claim is not a confirmed claim, and a verification run that cannot pull data has found nothing - not agreement.

3. Compute the delta from the raw data

After minus before, per metric, from the numbers pulled in step 2. Write them down in the run report before step 4. GSC lags 2-3 days: if the after window ends inside that lag, say the window is still filling and treat the result as provisional.

4. Now read the claim, and compare

Read claimed_delta (or claimed_before / claimed_after when the record states those instead) and call core.outcomes.compare(claimed, recomputed, tolerance=0.05).

Report what it returns as it returns it:

• agrees: True - the claim holds within tolerance. Say which metrics were compared. It confirms the arithmetic, nothing more.
• divergences - list each one: the metric, the claimed number, the recomputed number, and how far apart they are.
• unverifiable - metrics only one side has, or values that are not numbers. These are neither agreement nor divergence; name them as unchecked.
• agrees: False with an empty compared - nothing was actually compared. Report it as unverified, never as disagreement.

5. Report it, including the parts that look bad

Write runs/<UTCdate>-verify-outcome/REPORT.md: the item, the URL, the windows used (and whether they were read or derived), the raw pulls, the recomputed delta, the claim, and the comparison.

A divergence is a finding, not an error. Append it as a signal (core.contracts.append_signal) in the same shape the daily uses: outcome <item-id>: claimed clicks +120, recomputed +38 over the same window. Do not re-run the pull hoping for a better number, do not widen the tolerance to make it pass, do not quietly drop the metric that disagreed, and do not edit the outcome record to match. The reflector reads these signals; a divergence that never reached the brain teaches it nothing, and a loop that hides its own misses is worth less than no loop.

What agreement does and does not mean

Agreement means the recorded number matches what the raw data says over the same window. It does not mean the change caused the move. Attribution follows the same rule everywhere else in this plugin (canonical in hoo-daily's anomaly section, step 2.7): the claim, the comparison actually run, and what would falsify it. Verification and attribution are two different questions, and this skill answers only the first. See plugin/docs/reproducing-results.md for how someone outside the project runs the same check by hand.