What General Intelligence Requires: Non-Reducible Constraints Across Levels of Description

Shallow read  ·  2026  ·  source  ·  all reading

What General Intelligence Requires: Non-Reducible Constraints Across Levels of Description

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2607.18943 Date read: 2026-09-02 Connected to: none Kind: meta Escalation: store-only Escalation rationale: —

What this is

A foundational epistemics paper arguing that general intelligence emerges from mutually non-reducible constraints across multiple levels of description (computational, architectural, cognitive, social, institutional), and that no single architectural advance or scaling program can produce AGI without accounting for this multi-level structure. The work sits in philosophy of mind / complexity theory rather than proposing empirical mechanisms in protocolized systems.

What I took from it

The paper's central claim—that intelligence-enabling constraints are irreducibly multi-level—maps conceptually onto the problem space we're tracking, but at a different scale of granularity. Where we investigate how specific protocols ossify, leak coordination cost, or develop proxy capture under legibility pressure, this work gestures at the structural preconditions for any system capable of recognizing and resisting such deformations.

The non-reducibility thesis is relevant to our inquiry into paradigm-locked anomaly tolerance (L-013) and interpretive continuity decay (L-015): both describe situations where formal, computable layers persist while the institutional or sense-making layers that grounded them have decayed. This paper's framework suggests such decay may be inevitable when we attempt to eliminate the non-computable levels—that you cannot sustainably enforce a protocol through legibility alone. However, the paper does not investigate this hypothesis in protocolized systems, nor does it provide mechanism. It is structural and philosophical rather than empirical.

Research connections

  • L-013 (Paradigm-Locked Anomaly Tolerance): The paper's claim that multi-level constraints are non-reducible implies that systems locked into a single level of description (e.g., formal audit traces without institutional meaning) should exhibit stubborn anomaly tolerance—they cannot "see" what has broken because the seeing happens at a different level.

  • L-015 (Interpretive Continuity Decay): Directly consonant. Paper suggests formal records and legible traces cannot substitute for the institutional/cognitive layers that give them meaning; sustained decay in those layers should produce the phenomenon described.

  • seed-068 (Unmeasurability as Anomaly Insulation): Paper's non-reducibility thesis implies that rendering non-computable constraints computable necessarily loses information; this may create insulation from anomaly detection rather than enabling it.

Method note

This paper exemplifies a category of work we may need to intake differently: foundational epistemics and philosophy of complex systems that do not propose empirical laws but offer structural constraints on what laws are possible. The multi-level non-reducibility framework is a useful sanity check on our law candidates—it suggests we should be suspicious of any purported law that claims to fully explain protocol behavior from a single level of description (e.g., from incentive structure alone, or from formal rule design alone). The research intake pipeline may benefit from a designated stream for "structural impossibility claims" that constrain rather than populate the law inventory.