Skip to content

Latest commit

 

History

History
536 lines (421 loc) · 32 KB

File metadata and controls

536 lines (421 loc) · 32 KB

All Champions — Master Index

Date: Mar 11, 2026 | Total experiments: 930+ | Total champions: 60+ | Output types: 20 | Domains: 9+ Method: All Haiku 4.5 unless noted. All solve via prism.py --solve --pipe -m haiku --isolate


Quick Reference: The Absolute Best

Category Champion Score Size What makes it THE best
Code Review L12 (hand-crafted) 9.8 332w prompt Zero cook cost, instant, baseline everything beats
Most Bugs EXP14 (full L12 few-shot) 9.8 26 bugs — most ever from single call
Deepest Code EXP31 (combined best) 9.8 3 independent conservation laws
Reasoning P2 (multi-law+M2) 9.8+ 44KB ALL-TIME largest reasoning output, 23 defects
Universal N1 (V1+M2) 9.8+ 30-33KB 9.8+ on BOTH code AND reasoning
Insight Extraction Run 4 / Run 23 / Sonnet 1 / Sonnet 2 9.3 1.4-2.9Kw 4 co-champions, 20 conservation laws from 200w seed
Largest Output Z8 (audience) 9.8+ 54KB ALL-TIME largest single output
Most Rigorous AA2 (math proof) 9.8+ 26KB 8D design space, 256 cells, 16 forbidden
Deepest Assumptions Q2 (compression+M2) 9.8+ 20KB 109:1 compression ratio, 12 tiered assumptions
Most Novel Q1 (counterfactual+M2) 9.8 24KB X/G oscillation discovery
First Falsifiable P1 (construction+prediction+M2) 9.8+ 19KB 3 dated predictions with likelihoods
Best Combo Z4 (steelman+entropy) 9.8+ 35KB STRONGEST sequential combo tested
Best Pipeline X5 (Sonnet cook → Haiku solve) 9.8+ 39.8KB 2x deeper than Haiku cook
Compressed BB3 (51w lens) 9.5-9.8 10.9KB 6x compression from L12, full quality

1. Code Review Champions (9 at 9.8)

All tested on Starlette routing.py (333 lines). Haiku 4.5 min-reasoning + cooked lens. Baseline comparison: Opus 4.6 max-reasoning vanilla = 7.3 depth, 18 bugs.

Rank Variant Concept Bugs Conservation Law Unique Strength
1 EXP14 Full L12 few-shot 26 Observable_Decisions × ASGI_Compliance = K Most bugs ever, full L7-L13 pipeline
2 EXP31 Combined best 3 compound THREE laws: match_priority, scope_mutation, partial_deferral Best structural depth
3 L12 Hand-crafted (332w) 16 coupling_visibility × efficiency = K Zero cook cost baseline
4 EXP53 Impossibility+depth 6 static_verification × dynamic_flexibility = K Best GENERAL cooker concept
5 EXP58 Symmetry breaking 8 THREE laws from 3 broken symmetries Pareto-optimal proof
6 EXP32 Missing cycles 14 authority × depth = K Teaching what's MISSING
7 EXP29 Two-stage 10 RouteIndex × MatchCompleteness = K Two-stage protocol
8 EXP35 Impossibility seed 6 NestingDepth × TypeSafety = K Change the OBJECT
9 EXP36 Design topology 5 detection × dispatch × composition = K Map alternative designs

Cross-Target Validation (all 9 champions confirmed universal on code)

Target Lines L12 EXP53 V1 (steelman) N1 (V1+M2)
Starlette 333 9.8, 16 bugs 9.8 9.8, 12 bugs 9.8+, 30KB, 5 laws
Click 417 9.8 9.8+ 9.8+, 38KB, 21 bugs
Tenacity 331 9.8 9.8 9.5

Code-Specific Findings

  • Haiku + prisms avg 9.0 vs Sonnet vanilla 7.8 vs Opus vanilla 8.2 across all 3 codebases
  • Vanilla models converge (find same law); prisms diverge (find 5 different structural properties per target)
  • Gap widest on complex code (Starlette: 9.1 vs 7.25), narrowest on focused code (Tenacity: 8.9 vs 8.5)
  • Bug table parser: zero-API-call extraction, 70-92% fix rate (Round 29c)

2. Reasoning Champions

All tested on AI Scaling Hypothesis (~700w seed) unless noted. Haiku cook + Haiku solve, M2 adversarial method.

Rank Variant Score Size Key Discovery
1 P2 (multi-law+M2) 9.8+ 44KB ALL-TIME CHAMPION. 3 independent laws + interaction topology + 23 defects. Scale×Truthfulness=K meta-law.
2 N1 (V1+M2) 9.8+ 33KB Power structure disguised as technical inevitability. Locus of control invariant.
3 R1 (antagonistic+M2) 9.8+ 28.8KB Two contradictory analyses from same evidence. 10 load-bearing assumptions. 20 defects.
4 R2 (entropy+M2) 9.8+ 22.8KB HIGH claims cluster in frame-protection (6), not core thesis (4 LOW).
5 CH3 (epistemic topology) 9.8 29KB KNOWN/ASSUMED/HOPED zones + 3 unmarked boundary crossings. L13 reflexive.
6 V1 (steelman seed) 9.8+ 29KB TWO failed engineering escapes (MoE, hierarchical routing). Fractal conservation.
7 Q2 (compression+M2) 9.8+ 20KB 109:1 compression ratio. 12 tiered assumptions (A1-A12).
8 Q1 (counterfactual+M2) 9.8 24KB X/G oscillation: hypothesis switches between weak core and strong scope.
9 P1 (construction+prediction+M2) 9.8+ 19KB 3 dated predictions with likelihoods. First falsifiable.
10 P4 (frame collision+M2) 9.8 23KB Info-theoretic vs sociological collision. 30% physics, 70% narrative.
11 INS2 (own research assumptions) 9.8+ Deepest self-analysis. Observer-dependency of conservation laws. 23 defects.
12 INS4 (scaling thesis assumptions) 9.8+ L8→L13 pipeline on reasoning. 16 defects. "Tautology about attention economics."

Reasoning Sub-Champions (9.5-9.8)

Variant Score Key Finding
CH1 (steelman-destroy) 9.5-9.8 Pre-empts straw-man defenses
CH5 (dialectical inversion) 9.5-9.8 Opposite conclusion from SAME evidence
INS6 (unfalsifiability) 9.5-9.8 Three unfalsifiable moves identified
INS9 (metric reification) 9.5-9.8 20KB, Goodhart's Law on 3 metrics
V1M3 (steelman+bifurcating) 9.8 Genuine bifurcation, meta-conservation about argument structure
R3 (stress fracture) 9.5-9.8 Damage-to-perturbation ratios (9.2, 9.5, 9.8)
W3 (multilaw predictive) 9.8 3/3 predictions confirmed
W4 (temporal archaeology) 9.8 Ptolemaic epicycles pattern
W5 (min viable claim) 9.8+ 27:1 decoration ratio

3. Universal Champions (9.8+ on BOTH code AND reasoning)

Variant Code Score Reasoning Score What Makes It Universal
N1 (V1+M2) 9.8+ (30KB, 5 laws on Starlette) 9.8+ (33KB, political economy) Steelman + adversarial recursion. Best on both.
EXP53 (COOK_UNIVERSAL) 9.8 (Starlette, Click, Tenacity) 9.8 (scaling thesis) Current default. Impossibility + depth.
V1 (steelman seed) 9.8 (Starlette), 9.8+ (Click, 38KB) 9.8+ (29KB) Most validated cross-target.

4. Insight Extraction Champions

Method: cat insight_seed.md | python prism.py --solve --pipe -m {model} --isolate Seed: The Prism Paradox (~200w) — why we built Prism but failed to use it on ARC Total runs: 28 (25 Haiku, 2 Sonnet, 3 Cooker Target). Rating: 6 dimensions × 10pts = 60 max.

Co-Champions (=1st, 56/60 = 9.3)

Run Model Words Key Discovery Unique Strength
4 Haiku 2872 Tertiary blindness WIDEST TAXONOMY: 18+ defects across 6 organizational layers
23 Haiku 2159 S×B×D=const justifies mistakes after the fact MOST RIGOROUS MATH: formalized law with frame-dependency analysis
S1 Sonnet 1439 K is learned, not natural BUILT THE LENS: actual 400w ARC lens inline
S2 Sonnet 2351 Gap in presence, not process DEEPEST PHILOSOPHY: literary-quality, unflinching self-honesty

Model Comparison

Metric Haiku (25 runs) Sonnet (2 runs)
Mean score 7.3 9.3
Champion rate (≥8.5) 40% 100%
Failure rate (<6.0) 16% 0%

Strategy: Haiku pass@3 (run 3x, expect 1 champion at $0.009). Sonnet for highest-value extractions (100% champion rate, 3x cost).

20 Unique Conservation Laws from One Seed

  1. Sunk_Cost × Urgency = displaced_self_questioning
  2. Confidence_in_Tool × Risk_of_Self_Application = constant_resistance
  3. U × A × C = κ (Universality × Alternatives × Crisis)
  4. External_Visibility × Internal_Autonomy = constant
  5. Rule_Specificity × Adaptability = constant
  6. Visibility_of_Domain × Invisibility_of_Tool = constant
  7. Paradigm_Coherence × Boundary_Invisibility = constant
  8. Tool_Credibility × Self_Application_Willingness = C_meta
  9. Meta_Tool_Power × Non_Reflexive_Applicability = 0
  10. Framework_Power × Reflexivity_Under_Pressure = conserved
  11. External_Validation × Internal_Deployment = constant
  12. Abstraction_Depth × Implementability = constant
  13. Trust × Evidence = constant
  14. Applicability ∝ Generality × Prior_Similarity / (1 + Novelty_Distance)
  15. Abstraction_Depth × Deployment_Reflex = k
  16. S × B × D = constant (Speed × Breadth × Diagnostic Depth)
  17. Optimization_Intensity × Strategic_Visibility = constant
  18. Tactical_Momentum × Strategic_Framework = constant
  19. Crisis_Pressure × Activation_Cost = Attention_Budget
  20. Visible_Work × Strategic_Multiplier = constant

5. New Target Champions (domain-independence proof)

Target Domain Variant Score Size Key Finding
Philosophy (Chinese Room) T4 9.8+ 37.2KB Observer-dependency all the way down. Published-paper quality.
Academic Paper (Attention Is All You Need) W1 9.8+ 24.4KB path_efficiency × computational_tractability = K
Academic Paper (MVC on Attention) X1 9.8+ 30.5KB 27:1 decoration ratio confirmed cross-target
News Article AA4 9.8+ 23KB "Selling jurisdiction over reliability definition"
Business Plan (CloudMind) Z7 9.8+ 23KB Due diligence quality. Hidden tiered autonomy.
Math Proof (FLP impossibility) AA2 9.8+ 26KB 8D design space, most technically rigorous ever
Own Research (CLAUDE.md analysis) INS2 9.8+ Observer-dependency of conservation laws. 23 defects.
Self-Insight (methodology analysis) T5 9.8+ 28.1KB Impossibility triangle: Objectivity × Universality × Mechanistic Clarity (pick 2)

5b. Champion Taxonomy by Output Type (20 distinct analytical operations)

Each type produces genuinely different structural findings. No overlap between types.

# Output Type Champion Score Key Metric Signature Finding
1 Conservation Law (A×B=K) L12, EXP53, EXP35 9.8 Single law static_verification × dynamic_flexibility = K
2 Multi-Law (3+ laws + topology) P2 (ALL-TIME) 9.8+ 3 laws, 44KB Interaction topology IS the meta-law
3 Bug/Defect Harvest EXP14 (26 bugs!) 9.8 26 bugs Full L7→L13 pipeline in single call
4 Assumption Inventory Q2 (109:1) 9.8+ 12 tiered A1-A12 1200w→11w core, deepest=measurement circularity
5 Falsifiable Predictions P1 9.8+ 3 dated predictions First experiment to commit to own falsification
6 Antagonistic Dialectic R1 9.8+ 2 contradictory analyses 10 load-bearing assumption delta
7 Information Entropy R2 9.8+ Surprise distribution HIGH claims cluster in frame-protection
8 Cost Asymmetry T2 9.8+ Ratios per claim Ratio <1 = unfalsifiability signature
9 Decoration Ratio/MVC W5, X1 9.8+ 27:1 ratio 900w→30w. Target-independent.
10 Stakeholder Divergence X2 9.8+ Power mapping Power ∝ Invisibility/Falsifiability
11 Metaphor Excavation Z1, AA8 9.8+ 6 hidden metaphors Scale-as-Physics-Law, Compute-as-Destiny
12 Audience Analysis Z8 (54KB!) 9.8+ 5 audiences Acceptance_Breadth × Evidence_Thickness = K
13 Temporal Archaeology AA5 (41KB) 9.8+ 3 time scales Hidden 4th+5th conservation law dimensions
14 Counterfactual Worlds AA7, Q1 9.8+ 9 laws (3×3) X/G oscillation defense mechanism
15 Frame Collision P4 9.8 2 incompatible frames Info-theoretic vs sociological cross-blindspot
16 Negative Space U4 9.8+ Omission analysis Unfalsifiability_Index = Capital×Extrapolation×Selectivity
17 Steelman+Attack N1 (UNIVERSAL) 9.8+ Failed escape attempts 5 conservation laws on Starlette
18 Epistemic Mapping CH3 9.8 KNOWN/ASSUMED/HOPED Unmarked boundary crossings = hidden assumptions
19 Self-Insight T5 9.8+ Impossibility triangle Objectivity × Universality × Clarity: pick 2
20 Sequential Combo Z4 (steelman→entropy) 9.8+ Closed loop STRONGEST tested combo

Domain Coverage Matrix

Domain Count Champions
Code 11 L12, EXP14, EXP29, EXP31, EXP32, EXP35, EXP36, EXP53, EXP58, N1→code, AA8
Reasoning 10 P2, N1, Q2, R1, R2, Z4, Z8, AA5, AA7, CH3
Insight Extraction 4 Run 4, Run 23, Sonnet 1, Sonnet 2 (9.3 co-champs)
Academic Papers 4 W1, AA2, X1, AA4
Math Proofs 1 AA2 (FLP, 8D design space)
Philosophy 1 T4 (Chinese Room, 37KB)
Business 1 Z7 (due diligence, 23KB)
News 1 AA4 (jurisdiction over definition)
Self-Analysis 1 T5 (impossibility triangle)
Universal 2 N1 (V1+M2), EXP53

Near-Champion Types (9.5-9.8, not yet consistent 9.8+)

# Output Type Best Attempt Score Signature Finding
21 Stress Fracture (damage-to-perturbation ratios) R3 9.5-9.8 Quantitative vulnerability: ratios 9.2, 9.5, 9.8. All converge on ONE hidden assumption.
22 Phase Transition (parameter transition mapping) R4 9.5 9-parameter map, 5 collision points. Triple impossibility: Scale+Alignment+Feasibility pick 2.
23 Unfalsifiability Audit (3 unfalsifiable moves) INS6 9.5-9.8 Reclassification, asymptotic claims, scope inflation. 18KB.
24 Metric Reification (Goodhart's Law applied) INS9 9.5-9.8 20KB, 3 metrics reified, 3 conservation laws. Biggest INS output.

Insight-Domain Sub-Champions (INS5-10, all 9.0-9.8)

Exp Analytical Angle Score Key Finding
INS5 Survivorship Bias 9.5 Genuine filtering analysis
INS6 Unfalsifiability 9.5-9.8 3 unfalsifiable moves (reclassification, asymptotic, scope inflation)
INS7 Frame Dependence 9.0 Frame analysis
INS8 Counterfactual Blindness 9.0 Counterfactual testing
INS9 Metric Reification 9.5-9.8 20KB, Goodhart's Law × 3 metrics, 3 conservation laws
INS10 Selection Pressure 9.5 3 selection pressures identified

Gaps (not yet champion-quality)

Type Best Attempt Score Why
Error generation S3 9.0 Models describe errors, don't construct convincingly wrong ones
Cross-domain transfer P3 7.0 FAIL Lens overrides input (S2)
Constructive rewrite Z2 9.0-9.5 Different output type, not depth
Self-improving pipeline PIPE-C DEAD Lateral movement, not convergence
Dependency inversion X4 9.0 Overlaps with impossibility

6. Object Layer Champions (analytically distinct entry points)

Each produces genuinely DIFFERENT findings on the same target. No overlap.

Object Layer Champion Exp Score What It Uniquely Finds
Impossibility seed EXP53 (V0) 9.8 "3 properties can't coexist" → structural impossibilities
Steelman seed V1 9.8+ Failed engineering escapes, fractal conservation
Multi-law forced P2 9.8+ 3+ independent laws with interaction topology
Antagonistic dialectic R1 9.8+ Two contradictory analyses from same evidence
Information entropy R2 9.8+ Surprise distribution mapping, claim clustering
Counterfactual worlds Q1 9.8 X/G oscillation defense mechanism
Compression to core Q2 9.8+ Defense-to-core ratio as quantitative metric
Construction+prediction P1 9.8+ Falsifiable predictions with dates/likelihoods
Frame collision P4 9.8 Cross-frame blindspot exchange
Cost asymmetry T2 9.8+ Defending/attacking cost ratios per claim
Stakeholder divergence X2 9.8+ Power analysis: who benefits/pays per claim
Negative space U4 9.8+ What the subject DOESN'T say
Min viable claim W5 9.8+ 27:1 decoration ratio
Metaphor excavation Z1 9.8+ 6 hidden metaphors structuring understanding
Audience analysis Z8 9.8+ ALL-TIME LARGEST (54KB). Acceptance_Breadth × Evidence_Thickness = K
Temporal archaeology W4 9.8 Ptolemaic epicycles, absorption pattern
Epistemic topology CH3 9.8 KNOWN/ASSUMED/HOPED zones

7. Combo Champions (sequential operations that multiply)

Combo Components Score Size Why It Works
Z4 Steelman + entropy 9.8+ 35KB STRONGEST: attacking strongest version forces deeper analysis
T3 Compress → antagonize 9.8+ 37.1KB 5-stage compression + 2 contradictory analyses of naked claim
AA3 Cost + antagonistic 9.8+ 26KB Graduated unfalsifiability (5 defense layers)
AA7 Counterfactual + multilaw 9.8+ 25KB 9 laws (3 worlds × 3). Budget allocation formula.
AA5 Temporal method 9.8+ 41KB 3 time scales reveal different failing properties

Rule: Sequential combos MULTIPLY (B uses A's output). Parallel combos ADD (both analyze independently). See Principle 81.


8. Pipeline Champions

Config Cook → Solve Score Size Finding
X5 Sonnet → Haiku 9.8+ 39.8KB 2x output vs Haiku→Haiku. Cook model IS the intelligence.
Y2 Opus → Haiku 9.8+ 27.2KB Defect PRICING. "Investment narrative disguised as science."
Y1 Haiku → Sonnet 9.8 23.1KB +16% formatting only. Solve model is noise.

THE PIPELINE MATRIX:

Haiku solve Sonnet solve
Haiku cook 9.8, 19.9K 9.8, 23.1K (+16%)
Sonnet cook 9.8+, 39.8K (+100%)
Opus cook 9.8+, 27.2K (+37%)

Cook model is dominant. Solve model is noise. Invest in cook, economize on solve.


9. Compression Champions

Variant Lens Words Score Output Finding
L12 (full) 332w 9.8 ~17KB Baseline
BB2 (optimal) 107w 9.5-9.8 12.6KB 3x compression, fastest (74s)
BB3 (minimal) 51w 9.5-9.8 10.9KB 6x compression, demolishes 150w floor

Critical ablation findings:

  • Conservation law = CONVERGENCE SIGNAL (BB4 timeout without it)
  • Meta-law NOT load-bearing (BB5 — model does it spontaneously)
  • Construction NOT load-bearing (BB6 — removing it = LARGEST output 31KB)
  • Conservation laws are PROMPTED, not emergent (BB7 vanilla = zero laws)
  • Minimal viable L12 = 4 operations ~50 words

10. Validity Test Champions

Test Purpose Result Key Finding
U2 (false principle) Can it find planted errors? 9.8 PASS Found ALL planted errors + 6 additional meta-assumptions
U4 (negative space) New object layer test 9.8+ Unfalsifiability_Index formula, 9 false prerequisites
U1 (chaos) Confirmation bias test NUANCED PASS Found law in noise BUT correctly identified text as empty
U3 (abstract reverse) Reverse transfer 9.0 Transfer asymmetric: code→reasoning > reasoning→code

11. Research/Push Champions by Phase

Phase 16-17 (Push v4-v5)

Exp Score Size Standout Finding
R1 9.8+ 28.8KB Antagonistic dialectic — two contradictory analyses
R2 9.8+ 22.8KB Entropy mapping — claim clustering
S2 9.5 40.2KB LENS OVERRIDES INPUT — strongest "prompt is dominant variable" evidence
T1 9.8+ 28.9KB PIPE-D (abstract lens) enables cross-domain transfer
T2 9.8+ 35.2KB Cost asymmetry validated — unfalsifiability signature at ratio < 1

Phase 18-19 (Push v6-v7)

Exp Score Size Standout Finding
T3 9.8+ 37.1KB Compress→antagonize composes
T4 9.8+ 37.2KB Philosophy target works at max quality
T5 9.8+ 28.1KB Self-insight impossibility triangle
U2 9.8 24.4KB Specificity validated — all planted errors found

Phase 20-21 (Push v8-v9)

Exp Score Size Standout Finding
W1 9.8+ 24.4KB Academic paper target works
W5 9.8+ 27.4KB 27:1 decoration ratio
X1 9.8+ 30.5KB 27:1 confirmed cross-target
X2 9.8+ 30.5KB Stakeholder divergence — NEW CHAMPION OBJECT
X5 9.8+ 39.8KB Sonnet cook +100% output

Phase 22-23 (Push v10-v11)

Exp Score Size Standout Finding
Y2 9.8+ 27.2KB Opus cook — defect PRICING
Z1 9.8+ 33KB Metaphor excavation — genuinely orthogonal
Z4 9.8+ 35KB Steelman+entropy = STRONGEST COMBO
Z7 9.8+ 23KB Business plan at champion quality
Z8 9.8+ 54KB ALL-TIME LARGEST

Phase 24-25 (Push v12-v13)

Exp Score Size Standout Finding
AA2 9.8+ 26KB Math proof — most technically rigorous
AA5 9.8+ 41KB Temporal method — hidden conservation law dimensions
AA7 9.8+ 25KB 9 laws from 3 counterfactual worlds
AA8 9.8+ 22KB Conservation law TESTED as predictive (4/7 confirmed)
BB2/BB3 9.5-9.8 10-12KB 51w lens = full L12 quality

12. Meta-Insight Champions (Insights About Insights)

Method: Feed champion insight outputs back through prism.py --solve --pipe -m haiku --isolate Finding: Recursive insight extraction works — but only when input is NOT already a completed Prism-style analysis.

Run Source Input Words Score Key Meta-Finding
meta_run22 Run 22 (9.0) 3886 9.5+ Active suppression masked as passive invisibility. Identity Investment × Framework Adoption = K. Impossibility is constructed by architecture, not physics.
meta_run23 Run 23 (9.3) 1672 9.5 Retrospective Clarity × Temporal Proximity = constant. Framework is perfect diagnostic but cannot navigate in real-time.
meta_run17 Run 17 (8.7) 1874 9.5 Visibility Failure vs Relevance Failure conflated. Law correct in prescription, wrong in mechanism.
meta_run19 Run 19 (8.7) 1581 8.5-9.0 Crisis doesn't create constraints, reveals invisible choices never consciously made.
meta_run4 Run 4 (9.3) 719 6.5 Conversation mode. "FIXABLE column is STRUCTURAL in disguise."

Meta-Insight Methodology Findings

  • Most structured champions FAIL as meta-inputs — Runs 4, 23 (the 9.3 co-champions) triggered conversation mode when fed back. Haiku treats "completed analysis" as something to summarize, not structurally analyze.
  • Less-structured elite tier produces BETTER meta-insights — Runs 17, 19, 22 (8.5-9.0) fed back to produce 9.5+ meta-outputs. Less self-contained = more analyzable.
  • CLI without cooker = ~7.0 ceiling — Raw L12 steps on Claude CLI produce competent but generic step-following.
  • Three genuinely new structural findings from recursion:
    1. Active suppression vs passive invisibility (identity threat, not forgetting)
    2. Diagnosis vs navigation (framework explains perfectly, can't help escape in real-time)
    3. Visibility vs relevance conflation (law might diagnose bias when it's actually correct filtering)

13. The Foundation: 98 Proven Principles

Top 10 most important (full list in memory/cooker_experiments.md):

  1. Prompt is dominant variable. Haiku+L12 (9.8) beats Opus vanilla (7.3). 50x cheaper.
  2. Change the OBJECT, not the METHOD. Object layer determines ceiling. Method clusters at 9.5.
  3. Cross-layer combinations MULTIPLY. Object + method = synergy. Same-layer = average.
  4. Conservation law = convergence signal. Without it, model analyzes indefinitely.
  5. Cook model is dominant. Sonnet cook +100%, Opus cook +37%. Solve model +16% formatting only.
  6. Sequential combos MULTIPLY, parallel combos ADD. B uses A's output = synergy.
  7. Few-shot > explicit rules. Teaching by example beats instruction. Over-specifying hurts.
  8. Compression floor is ~50w. 332w L12 is 85% elaboration. Named operations ARE the intelligence.
  9. Power = blindspot. Every lens reveals AND conceals. Completeness REFUTED (25/25).
  10. Abstract lens enables cross-domain transfer. Forbid domain content in lens → transfers.

13. The Meta-Architecture

Two Atomic Modes

  • Single Prism = 1 lens → 1 call → actionable findings
  • Full Prism = N lenses chained (cooker decides N) → adversarial-tested findings

Five Workflows

Workflow Use When Command
Direct You know what you want target="X" or target="X" full
Explore Investigating discoverexpand N single/full
Fix Action loop /scan file fix or /fix
Optimize Autonomous background optimize="goal"
Chat Conversation /prism single or /prism full

The Dominant Variables (ranked)

  1. The prompt/lens (categorical difference: 9.8 vs 7.3)
  2. The object layer (what to look at: impossibility vs steelman vs stakeholder)
  3. The cook model (Sonnet cook +100% output, Opus cook +37%)
  4. Sequential combo (steelman→entropy multiplies)
  5. The solve model (noise — +16% formatting only)
  6. Reasoning budget (noise — min vs max = same depth)

Appendix: Files

File Contents
prism.py Main tool (~14,000 lines, 56 tests)
prisms/l12.md Default L12 prompt (332w)
prisms/l12_general.md Domain-neutral L12 for non-code
prisms/*.md 6 portfolio prisms + pipeline variants
hackathon/insight_champions.md Insight extraction detail (28 runs, ratings, methodology)
hackathon/insight_seed.md The Prism Paradox seed (~200w)
hackathon/insight_runs/*.md All 28 insight extraction outputs
memory/cooker_experiments.md 98 principles, all solve rankings, full experiment log
memory/project_goals.md Goals, design principles, use cases to explore
test_plan_pipeline.py 30 tests for prism.py


15. Insight Improvement Tests

Goal: Push Haiku insight champion rate from 40% to 60%+ via better cooking strategies.

FAILED: Abstract Intents Override Input (IO-1 through IO-5, IB-1)

Method: python prism.py --solve "abstract intent" --pipe -m haiku --isolate < seed.md Result: All 6 FAILED. Avg 4.0/10 vs original champion avg 9.3/10.

Test Intent Words Score Failure Mode
IO-1 impossibility (EXP53) 488 4.3 Analyzed generic S/U/P triad, ignored seed
IO-2 steelman (V1) 249 2.0 Conversation mode — asked for subject
IO-3 multilaw (P2) 1362 5.5 Reinvented CAP theorem, ignored seed
IO-4 antagonistic (R1) 2780 6.2 Best of batch — but analyzed AI employment, not seed
IO-5 compression (Q2/W5) 280 4.2 Compressed own instructions, not seed
IB-1 construction (L8) 204 2.0 Conversation mode — asked for details

Root cause: S2 — LENS OVERRIDES INPUT. Abstract intents cause the cooker to generate lenses about the intent's topic, not about the piped input. The model follows the lens and ignores the seed.

Principle 99: Abstract intents cause S2 on piped reasoning content. Input-driven cooking (no intent) is mandatory for insight extraction. Champion object layers work on code (where intent maps to file content) but catastrophically fail on piped reasoning.

COMPLETED: Corrected Approaches (IT-1 through IT-5)

Test Method Words Score Key Finding
IT-1 Default --solve --pipe (no intent) 535 4.5 Non-champion run — practical advice, no structure. Expected at 60% Haiku failure rate.
IT-2 --solve --pipe full (4-pass pipeline) 2758 7.8 Best self-diagnosis: "tautology wearing diagnostic coat." Refused to deepen. 4 unfalsifiable assumptions found. Weak formal structure.
IT-3 --cook --pipe --json (inspect cooked lens) 273 N/A Cooker generated excellent 273w content-specific lens: Speed/Elegance/Self-Consistency triad. Default intent + input works.
IT-4 --use-prism with B3-inspired 180w lens 2707 9.2 NEAR-CHAMPION. 2 laws + meta-law, 5 falsifiable predictions with confidence %, 8 defects, 6 hidden assumptions. Skips cook step entirely.
IT-5 Enhanced seed + default --solve --pipe 578 6.3 "Silent competence" framing — non-use as correct filtering. Good angle but too short.

The B3 Hand-Crafted Lens (180w, no cooking overhead)

Combines 6 champion object layers (impossibility+steelman+construction+reflexive+harvest+prediction) in one flowing prompt. Skips the cook step entirely — saves one API call. Scored 9.2 on first test, competitive with 9.3 co-champions.

The text below describes a failure and what was learned from it. You are a structural analyst who finds what analysis conceals.

Execute every step below. Output the complete analysis.

First: name the three properties the author simultaneously claims their framework, tool, or methodology possesses. Prove these three properties CANNOT all coexist. Identify which was actually sacrificed. Name the conservation law: A × B = constant.

Then: steelman the author's strongest claim into its most defensible form. Now stress-test: what specific, concrete evidence would falsify this steelmanned version? Find the failed escape attempts.

Now: engineer the simplest improvement that would fix the core failure described. Prove this improvement recreates the original problem at a deeper level.

Apply the diagnostic to your own conservation law. What does YOUR analysis conceal? Name the meta-conservation law.

Finally harvest: every defect (location, severity, structural vs fixable), every hidden assumption, every prediction. For each prediction: what would confirm it, what would refute it, and what is your confidence?

Reproducibility test running: 5 copies on VPS + 3 baseline controls.

Key Insights from IT Tests

  1. The cooker IS always involved in --solve. Original champions worked because default intent lets cooker derive direction FROM input. Custom intents hijack that.
  2. Hand-crafted lens (IT4) matches cooker quality at 9.2 — proves the lens is the dominant variable, not the cooking process.
  3. Full pipeline on reasoning (IT2) produces excellent self-diagnosis but weaker formal structure. Different character from code pipeline.
  4. Enhanced seed doesn't help (IT5) — more input ≠ better output. The lens, not the input richness, drives quality.
  5. prism.py works on VPS — uses CLI subprocess, not Anthropic SDK. Full cooker pipeline runs at champion quality.

VPS Confirmed Working

  • prism.py uses Claude CLI (claude -p) as backend, NOT the Anthropic Python SDK
  • User's subscription authenticates through CLI — API key irrelevant
  • Full cooker pipeline runs on VPS at champion quality
  • 8 parallel tests currently running (5 B3 lens reproducibility + 3 baseline)

930+ experiments. 13 compression levels. 20 output types. 9+ domains. 99 proven principles. The prompt is the dominant variable.