Gtm decision

GTM Decision

An experiment ended. Your job: gather honest results, judge them against the rules that were locked before the experiment started, and record the human's final call.

Mode: with USECASEIFY_API_KEY set, report the outcome to the registry (connected mode — this closes the experiment's public page and unlocks your next enrichment). Without a key, everything below works fully offline.

Step 1 — Locate the experiment

Find gtm/exp-*/criteria.lock.json for the running experiment (if several, ask which). Read the locked rules, primary metric, minimum evidence, and duration from it — the locked file, not from memory, not from BRIEF.md edits made after the start.

Step 2 — Collect results (real data only)

Open the experiment's TASKS.md → "Results to report" — the exact quantitative fields this experiment committed to measure.

For each field: read it from the real data source the implementation set up (saved SQL query, export script, analytics view, database). If a value cannot be read from real data, set it to "unknown". Never estimate, extrapolate, or fill in a plausible number — a fabricated number poisons the verdict.

Then ask the founder: the qualitative questions from TASKS.md, whether anything unexpected or any implementation issue could distort the data, and a one-paragraph summary in their own words.

Step 3 — Confirm before judging

Show a table of every value and where it came from. Get an explicit yes. (Connected mode: this is also what gets published — aggregate numbers only, never customer-identifying data, never secrets.)

Step 4 — Judge against the locked rules

First run the data-quality checklist (deterministic — do this before any judgment):

• no_quantitative_data — no numeric results at all
• missing_primary_metric — fields needed for the primary metric are missing/unknown
• insufficient_sample — largest relevant count below the locked minimum sample
• ended_too_early — ran less than half the locked duration
• self_reported_only — only qualitative, self-reported evidence

Then decide — exactly one of:

• continue — evidence supports the assumption; the next investment is justified
• iterate — the underlying problem/demand looks real, but the approach, message, price, or implementation needs to change
• kill — evidence does not support continued investment in this assumption

Rules of judgment:

• Judge against the LOCKED rules first. If results are ambiguous between two verdicts, pick the one the locked rules support and say why.
• Confidence is low / medium / high — qualitative only. Any flag above caps it at medium; a severe flag (missing_primary_metric, insufficient_sample, ended_too_early) forces low. No exceptions, including for exciting results.
• Use ONLY numbers present in the collected results, quoted verbatim. Never compute figures that were not reported. Never say "market validated".
• List evidence for and against the verdict, and at least one limitation.
• Recommend exactly ONE next experiment (2–3 sentences).

Step 5 — Deliver and record the final call

1. Write gtm/exp-NNN/DECISION.md per docs/format.md: the verdict large and unsoftened, confidence, rationale against the locked rules, results table, evidence both ways, limitations + data-quality flags, what it means in plain business language, the one next experiment, and next actions for human and agent.
2. The founder decides, not you. Ask: accept the recommendation, or override it? An override requires a one-line reason, recorded in DECISION.md next to the AI recommendation — both stay on the record.
3. Connected mode: POST {base}/api/v1/report with the results, verdict, final call, and override reason if any — this closes the public page honestly (an abandoned experiment can also be closed as abandoned; that too is honest data). (If the server is unreachable, skip the report and note it — the local DECISION.md remains the record.)
4. Update experiment.json status to completed, suggest committing gtm/.
5. If CONTINUE or ITERATE: offer /gtm-next-experiment for the follow-up. If KILL: acknowledge plainly — a clean kill is a successful experiment; it just saved weeks of misdirected work.