Empirical validation

Empirical Validation

Purpose

Claims of efficacy are cheap to make and expensive to trust. An inherited service "has a 74% success rate"; a new rule "sharpens the model"; a vendor "cuts errors 40%." The default failure is to reason about whether these hold — plausibility-check them, argue both sides — and then build on them. Reasoning cannot distinguish a real effect from an overfit artifact; only measurement can. This skill makes the reflex: when a claim is load-bearing, find the cheapest experiment that could falsify it, run it against real data with confidence intervals, and let the result decide — before you invest in it or ship it.

It is the difference between "this metric looks predictive" and "over 52,000 samples it's 49.8%, CI excludes nothing"; between "the new rule should help" and "14-rule flagged 9/12, 15-rule 1/9 — it hurt." Both verdicts were unavailable to argument and decisive to data.

When to use this skill

• About to rely on an inherited/abandoned system's claimed advantage (revival, due diligence).
• About to ship a change to something whose value is empirically established (a proven prompt, snippet, model, config, few-shot set) — verify the change didn't erode it.
• A performance number appears with no committed, re-runnable artifact behind it (docstring win rate, vendor ROI, "we saw a lift").
• A decision worth real money/time/reputation rests on "X works."
• Someone asks "does this actually work?" and the honest answer is "nobody measured."

When NOT to use

• No efficacy claim is load-bearing — you're not betting on whether something works.
• The claim is already backed by a reproducible artifact you can inspect (read it instead).
• Pure correctness questions about your own output — that's adversarial-verify.
• The cost of the experiment exceeds the cost of being wrong (rare; usually the cheap experiment is far cheaper than the misplaced investment — check before assuming this).

The procedure

1. Name the one claim the decision lives or dies on, in falsifiable terms. Not "the bot is good" but "signal S predicts the 5-min direction >52% after costs."
2. Find the cheapest ground-truth for it. Free public data, a held-out slice, a historical log, a small controlled run. The best experiments cost cents and minutes (a free public dataset pull; 12 claude -p cells). Ask: what's the least I can gather that could prove this false?
3. Design to isolate and to falsify. Change exactly one variable (controlled A/B: same everything, differ only in the thing under test). Prefer a design where a null result is meaningful. Interleave arms so a partial run stays balanced.
4. Run it and compute uncertainty. Report rates with confidence intervals, not point estimates; state n; flag thin cells. A "0%" on n=12 is not the same as on n=3,000. Compute the CI on the independent unit (window, user, day), not the raw row count — correlated rows inflate n and manufacture false precision.
5. Adversarially check the result before trusting it: fat tails / worst cases (not just the average), sample-regime limits, whether your proxy equals the real settlement variable, whether a "positive" is an upper bound (naive fill/selection bias).
6. Act on the verdict, and publish the artifact. The experiment script + numbers are the receipt — reproducible, committed. Let a kill be a kill and a pass be a pass; update the plan, don't re-litigate the data.

Quality bar

• The load-bearing claim is stated in falsifiable terms before any experiment.
• The result rests on real data with n and confidence intervals, not on argument.
• The experiment isolates the variable under test (a controlled comparison, where applicable).
• The result's limits are stated (sample regime, proxy fidelity, upper-bound caveats).
• A reproducible artifact exists; the decision follows the verdict, including "kill it."

Common failure modes

• Plausibility theater: arguing a claim is reasonable instead of measuring it. If you can gather ground-truth for cents, reasoning about it is a choice to stay wrong.
• Confirmation-shaped experiments: a test that can only pass. Design so a null result would show; test the strongest way it could be false.
• Point-estimate confidence: "74%!" on n=8. Report CIs; small-n fractions (5/8, 2/3) are the canonical overfit trap.
• Uncontrolled A/B: changing five things and crediting one. Isolate the variable.
• Pseudo-replication / unnamed selection: CIs computed on correlated rows (12,000 trades from 250 windows) — effective n is the independent cluster, not the row count; and reporting the best-of-N slice as if it were the only hypothesis. Compute on the independent unit; name the search space.
• Ephemeral evidence: a verdict resting on data that can't be re-pulled (a live API with short retention) and was never snapshotted — commit the raw data, or label it "unreproducible after DATE."
• Ignoring the tail: a mean that hides ruinous worst cases (a rare catastrophic loss the average quietly absorbs until it doesn't).
• Proxy ≠ target: measuring the convenient thing (a proxy signal) and claiming the real thing (the outcome that actually pays) — name the gap.
• Not shipping the receipt: a verdict with no committed, re-runnable artifact is a claim, not evidence — exactly the thing this skill exists to stop.

Works with sibling skills

verification-discipline labels a claim's evidence level; this skill goes and generates the evidence when the label is "unmeasured." adversarial-verify attacks your finished work by reasoning; this attacks a claim by experiment (and step 5 runs adversarial-verify on the result). research-methodology gathers what others have measured; this measures it yourself when they haven't. product-thinking decides whether a claimed edge is worth building; this decides whether it's real first. plan-gate/deep-decomposition sequence the larger work once the premise survives.

Provenance and maintenance

Added 2026-07 after two independent wins in one session: a load-bearing efficacy claim that looked impressive on its face was falsified for $0 on real public data — the effect was overfit noise — and a snippet change believed to help was shown by a controlled A/B to have eroded the pack's one proven behavioral result. In both, argument had been inconclusive and a cheap experiment was decisive. Re-verify by the negative test: on the next "does X work?" that gets answered by discussion rather than a runnable measurement, this skill's trigger failed. See .claude/learnings/2026-07-08-prove-the-premise-before-reviving.md and …-adding-to-the-proven-snippet-needs-an-ab.md.