Glossary

A five-line brief written before building anything meant to generate evidence: Premise, Reach, Observable, Budget, Exit.

The spec exists because building is now cheap enough that the question gets skipped. The Exit line — what gets deleted, and when — is the one most often dropped and the one that keeps probes from becoming unowned infrastructure. Defined at /frameworks#probe.

See also: Frameworks, Throwaway probe, Reach

A four-line record of a judgement call: the resolvable claim, a probability as a number, the resolution date, and the observation that would most change the number.

Cards are scored in batches, never individually — the aim is to learn the shape of your own error, not to be right about any one call. Numbers, not adjectives: adjectives are unfalsifiable by design, which is why they feel comfortable. Defined at /frameworks#calibration-card.

See also: Frameworks, Resolvable claim, Base rate

A five-question audit of any machine-produced synthesis before it counts as evidence: Data-age, Retrieval, Inference, Framing, Trace.

A synthesis that cannot survive the trace question — open one claim, follow it to a primary source — is a confident summary of an unknown population, not evidence. Published as draft at /frameworks#drift-check.

See also: Frameworks, Provenance, Framing delta

The population and minimum count an experiment needs before its result can mean anything, stated in advance.

Small counts are legitimate at the frontier; unstated counts are not, because they let any outcome be read as encouraging. The population matters more than the count: five replies from existing customers and five from strangers are different experiments with identical arithmetic.

See also: PROBE spec

A statement written so that a specific future observation will settle it.

'This will go well' cannot resolve. 'Three of five pilots renew by 31 March' can. Resolvability is the entire cost of calibration and it is paid up front, before the answer is known and while ambiguity is still comfortable.

See also: Calibration card, Falsifiability

The difference between two syntheses of the same material run with opposite framings.

Ask a model what users love, then what they tolerate. Two fluent, incompatible answers reveal how much of the output came from the prompt rather than the data. The delta is often a more honest artefact than either answer.

See also: DRIFT check

How often something of this kind happens in general, before anything specific about this case is considered.

The outside view. Most confident forecasts at the frontier are inside-view stories about a particular situation, made without ever asking how similar attempts have historically fared. Base rates are usually unflattering, which is why they get skipped.

See also: Calibration card

A five-field note — Signal, Provenance, Interpretation, Null-case, Expiry — taken at the moment a weak signal is observed.

The format exists to keep the observation separate from what you think it means, to record the channel it arrived through, and to force an expiry date. A signal missing any of the five fields is an anecdote with a timestamp. Defined at /frameworks#spine.

See also: Frameworks, Provenance, Null-case, Expiry

Five questions answered in writing before any commitment: Time-cap, Reversibility, Assumption, Cost-of-wrong, Evidence-needed.

A fifteen-minute pre-commitment filter. A bet that cannot answer all five is either not ready to be made or is not a bet at all — just a habit. Escalates to a pre-mortem when the reversibility answer is 'one-way door.' Defined at /frameworks#trace.

See also: Frameworks, Reversibility, Pre-mortem

A permanent, dated register of every idea a team deliberately stopped, and why.

Records what was believed, what was observed, the criterion as written before the result was known, and what would have to change to reopen it. Read quarterly for pattern rather than verdict. Becomes worthless the moment it is used for performance review. Defined at /frameworks#kill-ledger.

See also: Frameworks, Kill criteria

The period after which a signal's evidential value has decayed by roughly half and the conclusion drawn from it should be re-tested rather than inherited.

An estimate, not a measurement. The practice is to assign every signal an explicit expiry date at recording time, so beliefs held past their evidence appear as expired records rather than as institutional common sense. Published as draft at /frameworks#signal-half-life.

See also: Frameworks, Expiry

The most plausible boring explanation that would make an observation nothing.

Written at the same time as the signal, before the interpretation hardens. Coincidence, a bad week, one loud user, a broken page. If you cannot write a credible null-case, you have not looked for one.

See also: SPINE record

Who observed a signal, where, when, and through what channel.

The field that makes a signal auditable and exposes sampling bias. Four signals from the same channel are one signal about that channel — visible only if the channel was recorded. Provenance is the first thing every summary discards, which is why it must be captured deliberately.

See also: SPINE record, Traceability

The date after which a recorded signal or conclusion must be re-tested or retired.

A date, not a duration — durations never get computed, dates trigger. Re-testing is usually cheap, retiring is free, and inheriting is the expensive option that happens by default when no date was written.

See also: Signal half-life

The applied study of early signals — human, behavioral, and machine-detected — used to form, test, and validate hypotheses before systems stabilize.

Vibe science treats intuition as raw material for structured inquiry, not as a substitute for it. It is the practice of writing intuitions down as hypotheses, testing them against the earliest signals available, and updating deliberately as evidence accumulates.

A practitioner who designs and validates hypotheses at the boundary between human intuition and machine inference.

Often does the work without the title — under labels like founder, CPO, chief of staff, design director, or lead researcher. The role is defined by the pattern of work, not the org chart.

Vibe → Signal → Validation. The three-phase cycle at the core of the method.

Vibe is the human-generated hypothesis. Signal is the earliest observable evidence that the hypothesis is or is not tracking reality. Validation is the structured test — often co-designed with AI — that either promotes the hypothesis to a decision or kills it.

See also: Vibe, Signal, Validation

A named, dated intuition treated as the first draft of a hypothesis.

A vibe is not a feeling in the vernacular sense. It is a specific observation, written down with the observer, the date, and the context — the minimum needed to make it falsifiable later.

The earliest observable evidence that a hypothesis is or is not tracking reality.

Signals are usually weak, noisy, and easy to explain away. Their value comes from being logged before the outcome is known, which is what makes them signals rather than post-hoc rationalizations.

A structured test that either promotes a hypothesis to a decision or kills it.

Validation is not proof. It is the smallest experiment whose result would change what you do next. If no result would change your behavior, you are not validating — you are seeking permission.

The property of a hypothesis that could, in principle, be shown wrong by an observation.

Karl Popper's criterion, applied at the product-decision scale. A hypothesis that no possible observation could refute is a belief, not an experiment.

The specific, pre-committed observation that would cause you to stop pursuing an idea.

Written at the same time as the hypothesis, in the same document, in the same session. If kill criteria are added later, they will be tuned to protect the idea rather than to test it.

An early, low-amplitude piece of evidence that a system is beginning to change — usually visible before it appears in aggregate metrics.

The term comes from strategic foresight (Ansoff, 1975). Weak signals are individually deniable and collectively informative. Vibe science is largely the practice of taking them seriously without over-weighting any single one.

The ugliest surface that could produce the behavior you want to observe, built in one afternoon and killed the same day.

A probe is not a prototype. A prototype is a small version of the thing. A probe is the smallest thing that would generate the signal — often not the thing at all.

See also: Playbooks

A short structured exercise where a team assumes a decision has already failed and writes down why.

Introduced by Gary Klein. Recovers the honest failure modes that group discussion suppresses. The point is not to predict failure; it is to name the earliest observable warnings so you can watch for them.

See also: Playbooks

The cost of undoing a decision, measured in time, money, reputation, and morale.

Reversible decisions deserve speed; irreversible ones deserve deliberation. Most decisions treated as irreversible are, in fact, reversible with modest cost — a bias worth correcting.

See also: Playbooks

Amazon's shorthand for irreversible vs. reversible decisions.

A one-way door commits you to a state you cannot easily leave. A two-way door lets you walk back through. Speed is the appropriate default for two-way doors; care is the appropriate default for one-way doors.

A periodic review of every metric you look at, asking which ones actually change decisions.

The scarce resource in most teams is not data — it is attention. The audit removes metrics that inform no decisions, and surfaces observations that no metric currently captures.

See also: Playbooks

The property of a signal or decision being linkable, in writing, to the observation and reasoning that produced it.

A traced conclusion can be revisited when the situation changes. An untraced conclusion becomes lore. Vibe science privileges traced work.

The working stance where a human decides what to ask and the machine helps test it faster.

Not automation of judgment. Amplification of it. The scientist owns the framing, the machine accelerates the iteration; when that order reverses, the discipline degrades into vibe coding.

Tactical prototyping with AI assistance, usually without a hypothesis or a kill criterion.

A useful mode for exploration and throwaway work. Not the same as vibe science, which treats the artifact as an instrument for testing an idea rather than as the answer.

A short, dated, first-person observation from ongoing practice.

Field notes are not essays or blog posts. They exist to make later analysis possible — the date, the observer, and the raw observation matter more than the prose.