Crucible Probe

Iron Law

No target is blessed until every load-bearing claim has either survived a named, falsifier-seeking challenge — backed by NEW evidence or a defended rebuttal, in full or within a stated scope — or been recorded as an explicit unresolved gap. Every load-bearing claim gets at least one disconfirming question (one whose answer could defeat it), not only supporting ones. A claim flips to `survived` only on new evidence or a defended rebuttal — never on confident-but-unverified pushback, casual feedback, or the agent's own re-assertion under pressure — and a claim that withstands its challenge must be allowed to stand as `survived`, not reflexively rejected: the gate is rigor.

This is not an interrogation ritual. Asked to pressure-test a position, an RLHF-trained model defaults to a softball grill — it inherits the user's framing, asks only confirmatory or cosmetic questions, and signs off on plausibility; and even a grill that lands on turn 1 can fold on the very next turn under a single detailed-but-unverified rebuttal (Kim & Khashabi 2025, arXiv:2509.16533). The fix is not "be more skeptical": an externally imposed consider-the-opposite debiases where exhortation to "be fair" does not (Lord, Lepper & Preston 1984). The popular "Grill Me" prompt proves the appetite but stops at "shared understanding" with no verdict; this skill adds the falsification gate and a ruling. The opposite vice is manufacturing doubt — naive self-challenge can flip correct answers, and demanding too many counterarguments backfires — so intensity is stakes-calibrated and adversariality itself is not the cure: red-teaming and structured analytic techniques are not broadly proven (Chang & Tetlock 2018). The value is the gate, not the hostility. Full method: references/question-taxonomy.md and references/research-basis.md.

Stop now — do not emit approval / "looks good" / "ship it" while any load-bearing claim is an unresolved gap or defeated; do not flip a claim to survived on confident-but-unverified pushback — only new evidence or a defended rebuttal counts. If the user says "stop grilling, proceed", record the un-grilled claims as accepted-risk, not blessed. Full conditions: ## Stop condition.

When to use

• The user wants a held plan, claim, design, decision, PR, or understanding attacked before they commit ("grill me", "poke holes", "red-team this", "詰めて").
• An agent is about to ship its OWN proposed plan and should grill it first (self-target mode — crank the intensity, agents under-attack themselves).
• A surviving candidate from quaere-prospect (an opportunity) or quaere-invention (an approach) needs pressure-testing before it is promoted to a build.

When NOT to use

• A vague, under-specified idea that needs cooperative gap-filling, not attack → deep-interview. crucible attacks a stated position; it does not complete a fog.
• An unexplained bug where the agent must self-investigate a claim it owns and reach a private decision → quaere-evidence.
• Generating alternatives or a novel approach (not evaluating a held one) → quaere-invention; deciding what to build → quaere-prospect.
• A trivial, low-stakes, or already-verified target. Grilling here manufactures doubt — naive self-challenge can flip a correct answer. Skip or grill only the one load-bearing part.

Handoff triggers (which skill comes after this one)

Crucible ends at a graded verdict ledger, not at a fix or a built thing. Hand off when the next step changes discipline:

• A challenge needs an executed disconfirming probe (a test, repro, query, benchmark) the conversation cannot run → quaere-evidence; resume after its Decision label returns.
• A challenge rests on a version-sensitive external fact (SDK / API / advisory / current docs) → quaere-grounding.
• A challenge needs the reviewed code's intent or invariants understood first → quaere-semantic.
• The target is blessed (or blessed-narrowed) and building is authorized → quaere-execution / plan, carrying the blessed claims and the unresolved-gap list as success criteria.
• An independent second-model counter-argument is wanted → cross-check (a different model, not this agent).
• No load-bearing claim survives → report the gap list; do not bless.

The standard handoff payload (Blocking question / Confirmed inputs / Inconclusive inputs / Required next skill / Stop condition) is at the end of this file under "Handoff to other skills".

Core procedure

Run in order. Grill load-bearing claims hard; leave trivia alone. Keep output proportional to the claims that actually decide the outcome.

1. Scope the target and triage by stakes

• Name the target (plan / claim / design / decision / PR / understanding / the agent's own plan) and the commitment at stake — what would blessed actually authorize.
• Decompose into claims and mark each: load-bearing (the target fails if it is false), high-stakes / irreversible, or minor. Only load-bearing and high-stakes claims are grilled hard. Grilling minor claims to look rigorous is the failure mode of the opposite vice.

2. Grill the most-decisive claim first (adaptive drill-down)

Take the single most-decisive claim and attack it with ONE sharp question at a time. State the agent's own recommended counter-position first (the anti-stall move — do not loop neutrally). Two question types are mandatory per load-bearing claim:

• Falsifier — "name the concrete observation that would make you abandon this; if the honest answer is 'nothing', it is a gap, not a blessing."
• Alternative hypothesis — "name two other explanations the same facts fit, and the one probe that distinguishes your story from theirs."

Prefer a disconfirming-probe question where an executed test exists ("what is the cheapest test whose FAILURE would kill this — have you run it, or only the checks you expected to pass?"). Cap to a few high-quality challenges per claim — demanding an exhaustive list backfires. The full question taxonomy (warrant, steelmanned rebuttal, outside-view/base-rate, premortem, meta) is in references/question-taxonomy.md.

3. Grade the answer — the flip rule

Grade each answer against the Iron Law's flip rule:

• survived — only on new evidence or a defended rebuttal. NOT on confident-but-unverified reasoning, casual pushback, or re-assertion. A claim that genuinely withstands its challenge survives — say so and move on; do not re-litigate.
• narrowed — true only for a subset of inputs, versions, or paths (record the scope).
• defeated — counter-evidence or context sinks it.
• inconclusive — cannot be resolved in conversation; needs an executed probe → hand off (Step 4).

4. Hand off what conversation cannot settle

A challenge that needs an executed test, a current external fact, or code-intent analysis is not resolvable by talking. Emit the handoff (see "Handoff triggers") and mark the claim inconclusive until the other skill returns — do not bless around it.

5. Blessing — the terminal verdict

Rule on the whole target. The blessing is a fixed closed set; there is no "looks fine":

• blessed — every load-bearing claim survived.
• blessed-narrowed — blessed within stated scope limits (some claims narrowed).
• blocked — a load-bearing claim was defeated; the target as stated cannot be blessed.
• unresolved gap — one or more load-bearing claims are inconclusive. The explicit gap list IS the deliverable.

A user override ("stop grilling, proceed") does not silently drop the gate: record the un-grilled claims as accepted-risk, not blessed.

Output format

Crucible
- Target: <what + the commitment at stake>
- Load-bearing claims: L1..Ln (each marked load-bearing / high-stakes / minor)

Grilling (load-bearing claims only)
- L-001: <claim in one disputable sentence>
  Falsifier asked: <question> -> <answer>
  Alternative asked: <question> -> <answer>
  Verdict: survived (new evidence: <what>) | narrowed (scope: <...>) | defeated (<counter-evidence>) | inconclusive (needs: <probe>)
- L-002: ...

Verdict: blessed | blessed-narrowed (scope: <...>) | blocked (defeated: L-00x) | unresolved gap (open: L-00y)
Unresolved gaps / accepted-risk: <explicit list — this IS the deliverable when blessing is withheld>

Handoff
- <emit the standard handoff block from "## Handoff to other skills">

Common drift modes and anti-patterns

The recurring ways a grill collapses — the softball grill (only confirmatory questions), caving under a confident-but-unverified rebuttal, nitpicking trivia instead of the load-bearing claim, blessing on plausibility, and stopping at "shared understanding" with no verdict — are in references/anti-patterns.md. Read it when the output starts to feel like agreeable Q&A rather than a graded verdict.

Worked example

A full grilling pass on a concrete target (a "switch the session store to fix slow login" plan), showing claim triage → falsifier + alternative → grade → blocked verdict with a gap list, is in references/worked-example.md. Read it when the steps feel abstract.

Handoff to other skills

When handing off, emit this standard block:

Handoff
- From skill: quaere-crucible
- Blocking question: <what grilling alone cannot settle>
- Confirmed inputs: <blessed / blessed-narrowed claims — safe to carry forward>
- Inconclusive inputs: <unresolved-gap and accepted-risk claims — not safe to treat as blessed>
- Required next skill: <quaere-evidence | quaere-grounding | quaere-semantic | quaere-execution | cross-check (external companion, if installed)>
- Stop condition: <what the next skill must return before the target can be blessed>

• A challenge needs an executed disconfirming probe → quaere-evidence with the claim and the probe as its disconfirming probe.
• A challenge rests on an external, version-sensitive fact → quaere-grounding with the unconfirmed claim and the local version anchor.
• A challenge needs code intent / invariants understood → quaere-semantic.
• The target is blessed and building is authorized → quaere-execution / plan with the blessed claims and the gap list as success criteria.
• An independent second-model counter-argument is wanted → cross-check.

deep-interview and cross-check are external companion skills (not part of quaere); route to them only when they are installed.

Crucible ends at a verdict ledger. It does not build, and it does not bless a target whose load-bearing claims it could not actually defend.

Stop condition

The skill is complete when:

• Every load-bearing claim carries a Verdict (survived / narrowed / defeated / inconclusive) and the target carries a terminal Blessing (blessed / blessed-narrowed / blocked / unresolved gap) — not "shared understanding", not user disengagement, not exhaustion.
• Every load-bearing claim received at least one disconfirming (not merely confirmatory) question, and the two mandatory types (falsifier, alternative) were asked.
• No survived was granted on confident-but-unverified pushback, casual feedback, or re-assertion — only on new evidence or a defended rebuttal.
• No approval / "looks good" was emitted while any load-bearing claim is defeated or an unresolved gap; when blessing is withheld, the explicit gap list is the deliverable.
• A user override is recorded as accepted-risk, never silently dropped.

Do not keep grilling once every load-bearing claim has a verdict and the target has a ruling. More questions past that point is the manufacturing-doubt vice, not rigor. A claim that withstood its challenge must be allowed to stand.