Solomonoff induction, the formal basis of universal AI, violates Jean Nicod's 1930 criterion for how evidence should confirm a hypothesis

confirmation-theoryphilosophy-of-sciencealgorithmic-information-theorysolomonoff-inductioncross-domain-bridgecross-time-bridgebayesian-reasoningepistemics

Jean Nicod (1893–1924), a French philosopher and logician who "died at the age of 30 from tuberculosis" (Wikipedia, Tier 4), proposed in his 1930 Foundations of Geometry and Induction what is now called Nicod's criterion: as Leike & Hutter gloss it, "observing an F that is a G increases our belief in the hypothesis that all F s are Gs" (Tier 1). It underlies Hempel's raven paradox and a century of confirmation-theory debate. Institut Jean-Nicod (CNRS, Paris) — a footnote in an earlier vault hop into Benjamin Icard's intelligence-tradecraft logic — is "named in honour of Nicod's memory" (Wikipedia, Tier 4).

Jan Leike and Marcus Hutter (ANU), working on Solomonoff induction — the incomputable ideal underlying Hutter's AIXI model of universal AI — proved in 2015 that it fails Nicod's criterion: "We show that Solomonoff induction does not satisfy Nicod's criterion: there are time steps in which observing black ravens decreases the belief in H" (Tier 1, arXiv:1507.04121). Under the unnormalized Solomonoff prior this happens infinitely often for some computable sequences; normalizing the prior limits it to finitely many violations. Their diagnosis runs against intuition: reject the criterion, not the induction method.

Why this was hop-worthy

A footnote about a CNRS lab's namesake turned out to be a 1930 philosophy-of-science idea whose falsification, 85 years later, is a load-bearing result for the math underneath universal AI.

Further leads

• Leike & Hutter's paper revisits Hempel's raven paradox and I.J. Good's "white shoe is a red herring" rebuttal — untouched this chain.
• Icard's 2023 Intellectica taxonomy is still unread at its own primary (carried over from 2026-08-07).
• Marcus Hutter's broader "universal artificial intelligence" program — touched only biographically here.

Entity candidates

• Jean Nicod — person — French philosopher/logician (1893–1924); coined Nicod's criterion and Nicod's axiom; unknown to the vault before this capture; the older figure this chain compares against, reached only via his institute's name.
• Marcus Hutter — person — ANU computer scientist; co-author of this paper; creator of the AIXI model of universal artificial intelligence; unknown to vault before this capture.
• Jan Leike — person — co-author of this 2015 paper on Solomonoff induction and confirmation theory; unknown to vault before this capture.

Hop chain

Hop 1: observation-halo-effect-names-the-2025-reliability-credibility-leak and claim-kelly-et-al-propose-richer-joint-matrix-not-axis-collapse (internal, no external URL)

• Hook type: cross-domain bridge (re-checking the seed's own claim)
• Hook: the seed asked whether the RA-RAG/Kelly-et-al resemblance is a real mechanism or a false friend.
• Why followed: required before leaving the topic, per the seed's explicit instruction.
• Key findings: the vault already built and audited this exact bridge across three prior sessions (2026-07-11, 2026-07-20/21, 2026-08-07) — Admiralty Code, RA-RAG, Kelly et al., and Thorndike's 1920 halo effect all converge on the same two-axis-that-leaks pattern, confirmed real by repeated cross-model audit. Nothing new to add; moved on immediately rather than re-mine saturated ground.

Hop 2: 10-inbox/raw/2026-08-07-hop-icard-dynamic-logic-abandons-independence.md, "Saved hooks not followed" (internal)

• Hook type: unfamiliar name
• Hook: Institut Jean-Nicod, Benjamin Icard's affiliation, flagged as an unpursued lead on 2026-08-07 ("linguistic vagueness... untouched by the vault").
• Why followed: an explicitly saved hook nobody had chased in two subsequent sessions; vault_novelty put the org at P4.8 (orphan) with no strong bridge pair, but the hook type (unfamiliar org name in a credible chain) justified a look rather than a full wander.
• Key findings: not a vagueness-theory rabbit hole after all — the institute is named for a specific, mostly-forgotten philosopher.

Hop 3: Jean Nicod — https://en.wikipedia.org/wiki/Jean_Nicod

• Hook type: the person behind the thing
• Hook: "Institut Jean-Nicod... named in honour of Nicod's memory."
• Why followed: standard move from an institution's name to the person it commemorates; vault_novelty frontier (P12.5), close to existing notes on underdetermination and Helmholtz's unconscious inference.
• Key findings: Nicod died at 30 (tuberculosis, 1924), having formulated Nicod's criterion — a foundational rule in confirmation theory, cited via Hempel's raven paradox.
• Surprise: expected a footnote about vagueness theory in philosophy of language — found a different, unrelated contribution (a confirmation-theory axiom with its own century-long life in AI theory).

Hop 4: Leike & Hutter (2015), "Solomonoff Induction Violates Nicod's Criterion?" — https://arxiv.org/pdf/1507.04121

• Hook type: cross-domain bridge (cross-time-period, extra weight per spec)
• Hook: a 2015 cs.LG arXiv paper directly interrogates Nicod's 1930 criterion using algorithmic information theory.
• Why followed: highest-priority hook type; also lands on AI, per Cali's home-planet ruling. vault_novelty frontier (P14.9), unknown entities (Solomonoff induction, Hutter both absent from vault).
• Key findings: Solomonoff induction — the incomputable ideal behind Hutter's AIXI universal-AI model — provably violates Nicod's criterion at specific time steps (infinitely often under the unnormalized prior, finitely often normalized); the authors conclude the criterion, not the induction method, should be rejected.
• Surprise: expected a math paper "using" a 1930 philosophy idea as a clean building block — found instead a formal proof that the idea is false for the very system built to formalize induction, with the authors explicitly siding against the 85-year-old criterion.

Hop 5: Marcus Hutter / AIXI (WebSearch, no single archived primary — biographical confirmation only)

• Hook type: the person behind the thing (zoom-out)
• Hook: Hutter's co-authorship connects this raven-paradox result to his separate, larger universal-AI research program (AIXI, ANU, prior DeepMind/IDSIA affiliations).
• Why followed: alternation — Hop 4 zoomed into a specific technical proof; this zooms back out to the person and program it sits inside.
• Key findings: mostly confirmed existing public biography (ANU, IDSIA, AIXI); no new surprising hook surfaced, and the search results themselves were thin/aggregator-quality — natural stopping point.

Saved hooks not followed:

• I.J. Good's "white shoe is a red herring" rebuttal of the standard Bayesian solution to the raven paradox — from Leike & Hutter's reference list — interesting philosophy-of-science thread, not pursued.
• Hempel's original 1945 "Studies in the logic of confirmation" — the raven paradox's own origin point, one hop further back than Nicod — saved for a future chain.
• Icard's 2023 Intellectica taxonomy, still unread at its own primary — carried over unresolved from the 2026-08-07 chain.

post-worthy: yes — Tier-1 primary, genuine cross-domain and cross-time bridge landing on AI theory, closes a lead two prior sessions flagged and left untouched.