Research Fuzzer

The rules (hard — these ARE the skill)

1. No action without a bet. Before every action, append one bet line: what you will do, what you predict will happen, how confident, and which sub-question it serves. The prediction is written BEFORE you see the result — that ordering is the entire defense against fooling yourself. A result you would have "expected either way" teaches nothing; make the prediction sharp enough to be wrong.

2. Every settle picks exactly one verdict.

• confirmed — the prediction held. You learned little. A streak of confirmed is a leave signal, not an achievement: no new coverage here.
• surprised — the prediction broke. You MUST write the update: which belief changed, in one sentence. A surprise without an update line is not settled.
• anomaly — you cannot explain it. It goes on the unexplained-results ledger, may never be silently dropped, and blocks the gate (rule 5) until resolved or explicitly carried as a limitation.

3. Every settle asks: what did this open? Each result exposes leads you now know you could chase (a segment you noticed, a condition you held fixed, a control you lack). Append them to doors — even when you do not take them. This queue is your only honest measure of "how much is left": never "40% of the world" (unknowable), always "9 leads seen and never tried" (exact).

4. Three unresolved anomalies force a reframe. When unexplained results cluster, stop asking "which belief is wrong" and ask: "what axis is missing from my frame for these to make sense?" Publish frame v(N+1) with the new sub-question. A reframe is not an embarrassment; it is usually the most valuable single discovery of the run — record it as one.

5. The gate: no claim without passing it. Before stating any conclusion:

• it must account for every settled bet, or the exceptions are listed;
• the unexplained-results ledger is empty, or each open item is named in limitations;
• you have bet and settled a kill shot — the one action you believe most likely to break the claim, aimed somewhere you have not yet tested;
• the claim carries its scope: "within frame vN and the doors entered" — never an unscoped "the answer is X".

The panel

After each batch, produce the panel — run python scripts/tally.py fuzz-notebook.jsonl if Python is available, otherwise compute it by hand from the notebook (every reading is a count):

PANEL — frame v2 · 23 bets settled
  explored    : price? 6 · UX? 4 · need-evaporated? 8
  leads       : 9 untried · 5 followed
                oldest untried: "interview RETAINED users" (seen 12 settles ago)
  novelty     : last 10 settles → 8 confirmed, 1 surprised, 1 anomaly
                nothing new here lately — consider a far lead
  unexplained : 1 open (b17: retention spike in oldest cohort)
  spread      : 8 of your last 10 bets probed the same area (need-evaporated?)
  gate        : BLOCKED (1 unexplained open; no kill-shot bet on record)

The statistics window adapts to your pace: with only a handful of settles (slow, expensive experiments), the panel reports plain counts and withholds trend advice — a "last 10" reading means nothing when you have run 3.

Read it, then decide. The panel informs; it never commands. It only makes your own footprints visible: where you walked, what you saw and skipped, whether the last stretch taught you anything. State in one sentence what the readings imply before placing the next bet — a reading that is never read is not feedback.

What this is not

• The panel steers; it never certifies. Passing the gate means "not yet proven wrong within what you explored" — never "correct".
• Confirmed streaks feel like progress. They are the fuzzer's strongest signal to move: this region yields no new coverage.
• The queue only contains doors you noticed. Doors you never saw are not on it — which is why conclusions carry scope, and why a fat queue is a sign of good peripheral vision, not of failure.
• Do not tune the instrument mid-run: no reclassifying anomalies as confirmed, no editing old lines, no writing bets after seeing results. Editing the instrument to please the reading is the exact failure this skill exists to prevent.