Skill

Falsifying concepts — kill weak ideas early, cheaply, and defensibly

What this is for

A falsification campaign exists to kill an idea as cheaply as possible. Its job is not to build the idea or to confirm it — it is to find the fastest honest reason it will fail, and to make a survivor earn its survival. Bias toward killing. The enemy is motivated reasoning that keeps a doomed idea alive with a flattering metric, a missing counterfactual, or a goalpost that quietly moves.

Use this when a wrong "go" is expensive: a multi-month build, a large spend, a public commitment, a hard-to-reverse architecture choice, or a bold hypothesis someone is attached to. Skip it for small, cheap, or reversible decisions (just do them) and for ideas already validated by prior work.

Two modes

Mode	When	Cost	Output
quick-kill (default)	First pass on any big idea. Target its single riskiest assumption.	Hours; 1–2 kill-tests, one adversarial pass.	Kill (short negative report) or "survives triage → candidate for full campaign".
full-campaign	Quick-kill did not land and the cost of a wrong go is large.	Many agent-hours/tokens; confirm scope first.	4–8 pre-registered gates, executed studies, per-finding adversarial review, negative-result report or a caveated proceed.

Always start in quick-kill. Escalate only if it survives and the stakes justify it.

The non-negotiable disciplines

These are what make a kill defensible (and stop you fooling yourself). Every mode applies all of them. Details and worked failure examples are in references/meta-lessons.md.

1. Pre-register before evidence. Before gathering anything, write down for each kill-test: the decision metric, the kill threshold (a number), the null/ counterfactual, and a falsifiability demonstration — describe a plausible world in which the metric fails. If you cannot describe such a world, the metric is defective; redesign it before proceeding. Freeze thresholds — never revise them after evidence-gathering for that test begins. Use references/prereg-template.md.

2. Null-control every metric. Ask: would a zero-information / random / do- nothing baseline also pass this metric? If yes, the metric is vacuous and proves nothing. Report metric-minus-null, not the raw metric. Run python scripts/null_control.py (adapt the four callables) to measure it.

3. Counterfactual discipline. A benefit is the marginal gain over the cheapest existing alternative — never over "nothing". List what is already done for free, by someone else, or by a simpler method, and subtract it. Include the system's own mandatory work (what it would discover anyway).

4. Single-world coherence. Do not price benefits in one world and costs in another. Pick one consistent set of assumptions per calculation and state it.

5. Minimal executable kill-tests. Build the cheapest thing that can trip a gate — real data or sourced literature, not vibes. Prefer first-order / secular models over full fidelity. Label every number measured / assumption / external_validation_required (see references/evidence-labeling.md).

6. Adversarial review before any verdict is final. Every initial verdict is provisional. Spawn an independent skeptic (a fresh subagent) to attack the data, the counterfactual, and the arithmetic. Log every defect and its disposition; retract in place if overturned. Use references/adversarial-review-rubric.md. (In the worked example, both initial gate verdicts flipped under review.)

7. Verdict logic. If a kill gate trips → write the negative-result report and stop; do not pivot to a replacement concept unless explicitly asked. If all gates survive → report the surviving value with every caveat and only then recommend proceeding. Format in references/gate-verdict-table.md.

8. For a live prediction, score it — and score speed, not hit-rate. Fix a resolution date + review triggers (calendar and event-based) so it can't zombie; pre-declare the label set (killed / wounded / survived / inconclusive / too-early), a fixed cause code, and a consensus-at-call snapshot so lead time is measurable. The value is finding the load-bearing weakness before the market — prize lead time and resolution over being "right." (True calibration is a later, pooled goal that needs many independent, fast-resolving calls; a small or correlated set is calibration-ready, not calibrated — see meta-lesson #12.) See references/scoring-taxonomy.md; log calls in a predictions ledger and fold each resolution back into meta-lessons.

Quick-kill procedure (default)

1. State the idea in one paragraph. Name its single load-bearing assumption — the claim that, if false, kills the whole thing.
2. Pre-register 1–2 of the cheapest tests that attack that assumption, each with a kill threshold, a null control, and the counterfactual (prereg-template.md). Freeze the thresholds.
3. Lock the load-bearing number first, then gather the rest. Before theorizing, pull the single decision-driving quantity (the debt maturity ladder / cash runway / unit margin) directly from the primary source and reconcile it (does the ladder sum to the stated total? is there debt on the balance sheet you didn't list?); label it measured only once verified against that primary doc (meta-lesson #13 — theorize-before-verify was the #1 cause of wrong verdicts in the A/B self-test). Then do the focused web/literature check (fetch and verify real sources) or a <150-line calculation on real data. Label every number.
4. Adversarial pass: one fresh skeptic attacks data, counterfactual, and arithmetic (adversarial-review-rubric.md). Fix or retract.
5. Verdict: kill (write a short negative-result note: pre-reg, evidence with URLs, measured result, defect log, verdict) — or "survives triage".
6. Stop rule: if the load-bearing assumption fails cheaply, stop there. Do not build more to rescue it.

Full-campaign procedure (high-stakes only)

Confirm scope and rough cost with the user before launching — this can be expensive.

1. Landscape / baseline: what already exists and what it costs (verified, with sources). The idea must beat this, not a strawman.
2. Define 4–8 gates covering both feasibility ("can we do it?") and value ("is the result worth it?"). Pre-register each (metric, threshold, null, counterfactual, falsifiable world). Order them cheapest-kill-first.
3. Minimal executable study per gate on real data; every number labeled. Determinism: fix and record a master seed.
4. Adversarial review per finding (fresh skeptics); log defects + dispositions.
5. Single-world economics and an evidence matrix (claim | evidence type | source | confidence | what would invalidate it).
6. Gate verdict table + negative-result report on the first convincing trip, or a caveated proceed if all survive.

Worked example (a full campaign that ended in a defensible kill): https://github.com/HeliCorgi/covariance-compactor — read its STEP3_RESULTS.md §6 (13 dispositioned defects) and RESULTS.md §9 (two reversed verdicts).

Running it in this environment

• Use subagents / the Workflow tool for parallel evidence-gathering and for independent adversarial review (the reviewer must not be the author).
• Use the bundled scripts for the deterministic parts:
  • scripts/null_control.py — does a metric beat a zero-information baseline?
  • scripts/verdict_table.py — render the gate verdict table + kill summary.
• Anti-fabrication: every external number carries a fetched source URL, or is marked assumption, or external_validation_required. Cache sources with retrieval dates. Never report a computed number that was not actually computed.

Meta-check before you gather evidence

Two failure modes recur (see meta-lessons.md): a metric that cannot fail, and options priced on a world that does not exist. Before gathering evidence, run a single-world coherence check and an independent metric review — confirm each pre-registered metric could plausibly fail and is priced in one coherent world. Cheap to do; it prevents the most common wasted campaign.