Skip to content
Final StateRare for You, Common for the Corpus
VOL. I  ·  NODE 119▢  ATLAS

ONCE IN A LIFETIME

Rare for You, Common for the Corpus

Rare-for-you means a case with no local routine but many documented precedents in the corpus.

RARE FOR YOU, COMMON FOR THE CORPUS

It's a pattern-recogniser, not a fortune-teller

One local rare case connected to many similar documented cases in the wider corpus.The figure defines rare-for-you and common-for-the-corpus: no local routine, but enough documented precedents for pattern recognition.ONE CASE, TWO FREQUENCIESRARE FOR YOUFIRST TIMENO ROUTINECOMMON INTHE CORPUSTRACEABLE CASESCANDIDATES, NOT PROPHECY
  • Rare for you: no local routine
  • Common in an evidence base: many traceable, comparable cases
  • A retrieval-backed system can surface candidates for review

The useful system is not a fortune-teller and its model weights are not a case archive. It supports recognition when it can retrieve documented precedents with provenance, then show why they may be comparable.

THE GOVERNED EVIDENCE BASE

Local rarity becomes usable only when the evidence base is governed

Traceable local cases flowing into a governed evidence base with inclusion rules, outcomes, and provenance.The conceptual figure shows documented cases entering a governed case base and keeps the insurance analogy bounded; model weights are not treated as a queryable risk pool.GOVERNED EVIDENCE BASELOCAL CASESORG 01ORG 02ORG 03DOCUMENTED CASE BASEINCLUSION RULEOUTCOMESPROVENANCENOT MODEL WEIGHTS
conceptual analogy only: insurance pooling uses defined exposures and loss data; model training does not by itself create a valid reference class.

The insurance comparison in 001's corpus beat is an analogy about aggregation, not statistical equivalence. The advantage requires a governed evidence base with inclusion rules, outcomes, and provenance.

  • Insurers estimate rates from defined exposures and observed losses
  • A documented case base can aggregate precedents across organizations
  • An LLM's training mix is not automatically a valid or queryable reference class

THE SKEPTIC AND THE CHAMPION

An adversarial collaboration found two conditions for skilled intuition

Kahneman and Klein panels converging on a shared rule for trustworthy intuition.The collaboration resolves into two required conditions: a regular environment and feedback sufficient to learn from it. Missing either means confidence is not evidence of skill.A FAILURE TO DISAGREEKAHNEMANBIASRISKKLEINSKILLEDUSECOLLABORATEREGULARENVIRONMENTVALID FEEDBACKTO LEARNANDMISS EITHER:CONFIDENCE IS NOT SKILL
source-based: Kahneman and Klein, Conditions for Intuitive Expertise: A Failure to Disagree, American Psychologist, 2009.
  • A sufficiently regular environment with valid cues
  • Adequate opportunity to learn those cues through practice and feedback
  • Confidence alone is not a diagnostic of validity

Kahneman and Klein found substantial agreement while preserving real disagreements. Their 2009 conditions help test when intuition has earned trust and underpin 004's wicked-ground beat.

KIND GROUND, WICKED GROUND

Feedback is what turns experience into skill — or into a confident lie

Split diagram contrasting kind learning ground with wicked learning ground.The kind side closes a fast honest feedback loop; the wicked side frays into delayed or misleading feedback, including the typhoid-hand example of learning the wrong lesson.WHAT EXPERIENCE LEARNSKINDFASTHONESTFEEDBACKLEARNING LOOPWICKEDDELAYEDOR MISLEADINGTYPHOID HANDWRONG LESSONFEEDBACK TRAINS THE MODEL
concept-based: Robin Hogarth, Educating Intuition, 2001.
  • Kind: stable cues and timely, representative feedback can support learning
  • Wicked: selected, delayed, or misleading feedback can reinforce error
  • Rare, high-stakes decisions can be wicked when outcomes teach slowly or ambiguously

Robin Hogarth's distinction explains why experience counts only when the environment teaches. The typhoid physician example shows a correct diagnosis paired with misleading feedback about method; 004 applies that warning to consequential decisions.

Thirty years can be one wrong model, practised thirty times.

Seniority is not the same as feedback. In a wicked domain the sample is silent — 'it has never gone wrong for me' is not evidence, because a rare miss may leave no trace until it lands.

THE CORNER WHERE IT BREAKS

A missing class and OOD status are different boundaries

Separate evidence-class and OOD tests showing that the two boundaries can overlap without being equivalent.The figure removes the old no-class-to-OOD chain and asks independently whether a governed evidence class exists and whether a case lies outside a chosen reference distribution.TWO BOUNDARIES, TWO TESTSEVIDENCECLASSTRACEABLE?COMPARABLE?OUTCOMES?EPISTEMICOODSTATUSREFERENCEDISTRIBUTION?OUTSIDE IT?DISTRIBUTIONALMAY OVERLAPNEITHER IMPLIES THE OTHER
  • Evidence-base question: is there a traceable, comparable reference class?
  • OOD question: is the case outside a specified reference distribution?
  • They may overlap, but neither implies the other

The analogy is bounded: OOD status is distributional, while no reliable class is an epistemic limit on comparison and odds. Test each separately.

WHAT THE ENGINE HANDS YOU

A reference class and a base rate — not a verdict

Engine handing a reference-class dossier to an owner while withholding a verdict.The dossier contains comparable cases and base rates, but the commitment remains with the owner; the figure also flags false alarms at low base rates.HANDOFF, NOT VERDICTOWNERREVIEWSDECIDESENGINEFINDSDOSSIERCASESOUTCOMESPROVENANCEINCLUSIONRULEVALIDATE COMPARABILITYLOW BASE RATE? CHECK ALARMSCOMMITMENT STAYS WITH OWNER
  • Retrieve candidate cases with provenance and outcomes
  • Validate comparability before calculating a base rate
  • Use the outside view as evidence; keep authority and commitment explicit

The system can propose a candidate outside view, not an answer. Validate inclusion and comparability, then check low-prevalence false alarms before using a base rate.

RARE, THEN YOURS

The corpus lights the documented; the new stays yours

  • Ask for the comparable cases, inclusion rule, outcomes, and provenance
  • If the class validates, use its base rate; if not, mark uncertainty rather than inventing one
  • Next: which noticings are worth watching, in 004

Carry the governed evidence boundary back to the AI wins you can measure: retrieve precedents where a class validates, and mark uncertainty where it does not.

Read the transcript

01 · ONCE IN A LIFETIME

A business owner opens one letter under the lamp. A supplier's accounts are drifting toward trouble. Or a grant window has opened that fits her operation almost exactly. Or a clause buried in a contract will quietly cost her a year from now. Whichever it is, she has never seen it before. There is no routine for it, no drawer of past cases to reach into, so it lands hard on the little attention she has left at the end of a working day. It is, for her, a once-in-a-lifetime event. And here is the thing she cannot see from her desk. Cases like it may already be documented elsewhere, in other firms, other disputes, other filings. Rare for her. Common for the corpus.

02 · RARE FOR YOU, COMMON FOR THE CORPUS

Say plainly what the phrase means. A situation can be rare for one person and well documented elsewhere. But do not confuse model training with a case archive. A model's weights do not expose which cases it learned from, how they ended, or whether they are comparable. The useful system here is retrieval-backed. It searches a governed evidence base, surfaces traceable candidate precedents, and shows the features that may make them relevant. That is recognition support, not prophecy. The case is common for the corpus only when the corpus contains inspectable evidence, not merely because a model produces a familiar-sounding answer.

03 · THE GOVERNED EVIDENCE BASE

Insurance offers a useful but bounded analogy. An insurer estimates rates from defined exposures and observed losses across policyholders. A governed evidence base can likewise aggregate documented precedents that are locally rare. But model weights are not a case archive, and training is not insurance pooling. The useful system needs explicit inclusion rules, traceable provenance, recorded outcomes, and a maintained scope. Local rarity becomes informative only when that governed evidence supports a defensible comparison class.

04 · THE SKEPTIC AND THE CHAMPION

Daniel Kahneman and Gary Klein approached professional intuition from different research traditions. Their adversarial collaboration did not erase every disagreement, but it found substantial common ground. In 2009 they identified two conditions for skilled intuition. The environment must be sufficiently regular to contain valid cues. And the person must have adequate opportunity to learn those cues through prolonged practice and feedback. Where both conditions hold, intuitive judgments can reflect skill. Where either fails, subjective confidence is not a reliable diagnostic of validity.

05 · KIND GROUND, WICKED GROUND

Robin Hogarth called the contrast kind and wicked learning environments. Kind environments contain stable cues and provide timely, representative feedback. Wicked environments select, delay, or distort the feedback, allowing a mistaken method to feel successful. His typhoid example captures the trap: a physician could diagnose correctly by examining patients while unknowingly helping spread disease, so apparent diagnostic success did not validate the method. Rare, high-stakes decisions can be wicked when they recur too slowly, outcomes arrive late, or counterfactuals remain hidden. They are not wicked merely because they are rare.

06 · THIRTY YEARS, ONE WRONG MODEL

So here is the line to hold onto. Thirty years in a rare, high-stakes domain is not thirty years of learning. It can be one wrong model, practised thirty times, because the domain never sent back the feedback that would have corrected it. Seniority feels like expertise and is not the same thing. And the most dangerous evidence of all is the quiet kind: it has never gone wrong for me. With a rare event, the sample can be silent by nature. The absence of catastrophe is not proof it cannot happen. It only means the die has not landed on that face yet, while you were building your confidence on the rolls that missed it.

07 · Advertisement · Twain GTM Agents

A good outbound message should know why it is being sent. Twain GTM agents research accounts, qualify contacts, and turn live signals into relevant outreach. They work with your CRM, personas, and GTM stack, so the message is not just personalized. It is backed by a specific reason to start the conversation.

08 · THE CORNER WHERE IT BREAKS

Keep two boundaries separate. The evidence-base question asks whether traceable, comparable cases support a defensible reference class. The out-of-distribution question asks whether this case lies outside a specified reference distribution. They can overlap, but neither implies the other. An OOD case may have useful precedents in another governed source, while an apparently in-distribution case may support no stable odds for the decision. Treat no reliable class as an epistemic limit and OOD as a distributional relation, then test each with its own evidence.

09 · WHAT THE ENGINE HANDS YOU

What should the system hand you? Candidate cases with provenance and recorded outcomes. Then the human process tests comparability, states the inclusion rule, and calculates a base rate only if the class survives. That is the beginning of an outside view, not a verdict. It also needs a prevalence check. When the target event is rare, even a useful detector can produce an alert queue dominated by false positives. A confident flag therefore proves neither that the case is common in the evidence base nor that this instance belongs to the class.

10 · RARE, THEN YOURS

The phrase earns its value only with evidence attached. Ask to see the comparable cases, where they came from, why each belongs, and how they ended. If the class validates, use its base rate as an input. If the set is empty, hand-waved, or selected after seeing the answer, do not manufacture certainty. Mark the gap and use an accountable decision process suited to uncertainty. The corpus can light documented precedents. It cannot turn unsupported novelty into a known class. Carry that distinction back to the spine, where the next question is not merely what a system can flag, but which noticings deserve attention.

01 / 10 · ONCE IN A LIFETIME0:00 / 7:17