Jordan & Pippen — the abductive pair

Papers on LLM reasoning and its limits, distilled into personal Claude Code skills. Each paper's useful discipline (not its summary) becomes a skill in ~/.claude/skills/, pressure-tested RED→GREEN before shipping.

Contents

Paper	Thesis	Derived skills
llms-cant-jump.pdf — Tom Zahavy, Google DeepMind, Jan 2026	LLMs master induction (pattern compression) and deduction (formal derivation) but lack abduction — Einstein's "Jump" from surprising experience to new axioms, made via thought experiments with counterfactual intervention.	jordan, pippen

Two halves of the same cycle: pippen creates the openings, jordan finishes. Pippen diverges — it manufactures the surprise and invents candidate axioms. Jordan converges — it takes a surprise that already happened and finds which explanation is true. Both are domain-general; neither is about code.

jordan — make the jump

LLMs can't jump; Jordan can. It fires on a surprising result, a problem surviving two fix attempts, an explanation accumulating exceptions, or nothing failing while two parts of the picture rest on contradictory assumptions:

1. Reproduce the surprise yourself (sit in the elevator).
2. Audit the axioms the current story rests on — VERIFIED / INHERITED / UNEXAMINED.
3. Generate ≥3 rival explanations; label tolerance patches and unexplained entities as Vulcans — naming an anomaly is not explaining it.
4. Run the smallest distinguishing experiment per rival (cut the cable) — whatever runs the counterfactual is the laboratory.
5. Test one prediction the existing checks don't measure (a Mercury); the new explanation must reproduce everything the old one explained.
6. State falsifiable axioms; report broken assumptions to their owners instead of shielding them.

Trigger: "jordan", "jump", "thought experiment", "question the axioms" — or let it auto-trigger.

Validated against a billing trap (falsely "verified" input contract + a check that misses silent data corruption): the baseline agent shipped an anomaly-normalizing patch that shielded the false contract; with the skill, the same scenario produced a measured falsification of the contract, probe-killed rival hypotheses, and an upstream breach report.

pippen — create the opening

For brainstorming, when every candidate feels like "X but for Y", or when a field feels stale though nothing is visibly wrong. Its core principle is the paper's most useful one for invention: don't wait for an error signal — Newtonian gravity was passing every test when Einstein replaced it.

1. Map the givens; tag each physics (binding) or custom (merely settled). Customs are the raw material.
2. Collect seeds — contradiction (two accepted frameworks that grind together), identity (two things an observer couldn't tell apart, so hypothesize they're one), conviction (a prior about how the domain should be).
3. Build the elevator: a sealed situation with a concrete observer; run it and write what they experience and can't name.
4. Jump to the boldest axiom that makes that experience necessary rather than coincidental.
5. Deduce the stakes: one consequence nobody asked for, its cheapest test, and its Newtonian limit.
6. Ship the portfolio — no winner picked. Converging is jordan's job.

Trigger: "pippen", "ideate", "invent", "fresh ideas" — or let it auto-trigger.