Heuristic Algebra Paper

**Author:** Henry Carstens **Date:** March 2026 **Pages:** 25 (including appendices) **Project:** heuristicchat — Version 1.0, Campaign H1 **Full paper:**.

heuristic-algebra research LLM reasoning-scaffolds formal-systems empirical-validation
source Updated 2026-04-09

Heuristic Algebra: A Formal System for Composing Reasoning Scaffolds for Large Language Models

Author: Henry Carstens Date: March 2026 Pages: 25 (including appendices) Project: heuristicchat — Version 1.0, Campaign H1 Full paper: PDF

Abstract

Presents heuristic-algebra (H), ten operators for constructing, combining, and selecting reasoning scaffolds injected into LLM system prompts. Validates the framework through a 26-condition Elo-rated scaffold tournament across four frontier providers. Every scaffold outperforms the baseline (+160 to +799 Elo). The top result — lucid-judgment, a three-generation harmonic composite — achieves 1862 avg Elo vs 1063 baseline. Same three sources combined via different operators produce a 231-point gap, demonstrating that the compression mechanism determines scaffold quality.

Three central results

  1. Scaffolding works. Every heuristic condition outperforms the unscaffolded baseline. Minimum gain: +160 Elo (Strategy). Maximum: +799 (Lucid Judgment).
  2. Algebraic composition works. Top 5 are all composites. Median composite: 1637 Elo. Median primitive: 1464.
  3. The operator matters. ⊕bio(DC, RR, SJ) → Sovereign Judgment, rank #8, 1631 Elo. ⊕ha(DC, RR, SJ) → Lucid Judgment, rank #1, 1862 Elo. 231-point gap, same sources.

Tournament design

  • 26 conditions (25 scaffolds + 1 baseline)
  • 4 providers: Grok-4, Claude Opus 4.6, GPT-5.4, Gemini 2.5 Pro
  • 5 strategic decision queries per condition per provider
  • 3-judge Elo panel (xAI, Anthropic, OpenAI), K=32, initial 1500
  • 185 total experiments
  • Kendall’s W = 0.560

Top 10 results

# Condition Avg Elo Derivation
1 Lucid Judgment 1862 ⊕ha(DC, RR, SJ)
2 Actionable Reasoning 1784 ⊕bio(CR, DC, AR)
3 Superior Judgment 1767 ⊕bio(6 sources)
4 Decisive Calibration 1734 ⊕bio(CR, MR)
5 Reflexive Reasoning 1725 ⊕bio(AR, MR, CR)
6 Forecasting 1647 Primitive
7 Master Reasoning 1637 ⊕bio(CT, F, SM)
8 Sovereign Judgment 1631 ⊕bio(DC, RR, SJ)
9 Calibrated Reasoning 1625 ⊕bio(CT, F)
10 Grounded Sovereignty 1623 ⊕inn(SV, CR)

Baseline (no scaffold): 1063 Elo.

Key findings from analysis (§6)

  • LJ on Claude (1983) is the single highest cell in the tournament. +740 delta over Claude’s baseline (1243).
  • Meta-reasoning sources + ⊕ha is the strongest derivation method tested. Ranking: ⊕ha(meta) > ⊕bio(meta) > ⊕inn(meta) > ⊕ha(domain).
  • Preserved tensions outperform resolved tensions. Sovereign Judgment resolved all contradictions. Lucid Judgment preserved two. The scaffold with productive tensions won by 231 points.
  • Generation predicts rank. 3rd-gen > 2nd-gen > 1st-gen > primitives.
  • LJ beats all three of its sources individually by 95–137 points (source avg: 1742).

Caveats (§6.8)

  • All rounds use the same 5 strategic decision queries
  • Elo values are relative to this pool
  • LJ shares deep lineage with its sources — meaningful test is whether combination exceeds sources
  • Kendall’s W = 0.560 is moderate, not high

Paper structure

Section Content
§1 Motivation — scaffolds make LLMs reliably thorough
§2 10 operators defined
§3 Tournament design and condition pool
§4 Provider-condition heatmap (26 × 4)
§5 Full results and structural observations
§6 LJ derivation, operator comparison, key findings
§7 Conclusion and reproduction instructions
App A Full algebra specification with per-operator falsification
App B Heuristics of Lucid Judgment (7 axioms, tensions, dependencies)
App C Five strategic decision queries

External resources

  • NotebookLM audio overview: link
  • NotebookLM text summary: link

See Also