O0-INFO-006

Null Result: Compression Explains Attractor Convergence

**Version:** 1.0

**Research status:** UNSUPPORTED (NULL RESULT AGAINST PREFERRED HYPOTHESIS)

CLAIM STATUS: UNSUPPORTED (ADVERSE FINDING)
EVIDENCE TYPE: COMPUTATIONAL (ADVERSARIAL ANALYSIS)
PHYSICAL VALIDATION: NONE
INDEPENDENT REPLICATION: N/A (ADVERSARIAL STUDY)
PHILOSOPHICAL PROVENANCE: O/0 ARCHIVE
ARCHIVE ENDORSEMENT: LIMITED TO REPORTED RESULT

Abstract

This study presents an adversarial analysis of the semantic convergence findings in INFO-003 and INFO-005. We construct a simple compression baseline: as questions become more abstract, the vocabulary available for meaningful response shrinks, response diversity decreases mechanically, and embedding-space convergence occurs trivially without any "attractor" mechanism. Results: the compression baseline explains approximately 40% of the observed convergence effect (reducing d = 0.72 to residual d = 0.43 after compression correction). This is a partial null result: the preferred hypothesis (structural semantic attractors) is only partially supported. A substantial fraction of the convergence is parsimoniously explained by vocabulary compression alone. The research program must revise its claims accordingly.

Source proposition

This study is adversarial to the O/0 framework's preferred interpretation of INFO-003. It tests the simplest possible alternative explanation: that convergence under abstraction is a trivial consequence of vocabulary narrowing, not evidence of deep structural features of language or reality.

Note: This is an integrity check. The research program must actively seek to undermine its own findings.

Scientific audit

• The compression hypothesis is the most parsimonious explanation for semantic convergence under abstraction and should have been the first thing tested.
• That it was not tested in INFO-003 is a methodological oversight (corrected here).
• The baseline is deliberately simple—if even a simple model explains substantial variance, the interesting-hypothesis is proportionally weakened.
• The residual (unexplained convergence after compression correction) may or may not be interesting; further adversarial analysis is needed.

Research question

How much of the semantic convergence observed in INFO-003 can be explained by a simple compression model (vocabulary shrinkage under abstraction) without invoking any "attractor" or structural mechanism?

Operational definitions

• **Vocabulary diversity**: Number of unique tokens in response set at each abstraction level (type count).
• **Response entropy**: Shannon entropy of the token distribution at each level.
• **Compression-predicted convergence**: The convergence expected purely from vocabulary shrinkage, estimated by computing convergence among random strings drawn from vocabularies of matching size.
• **Residual convergence**: Observed convergence minus compression-predicted convergence.
• **Compression fraction**: The proportion of observed convergence explained by the compression baseline.

Hypothesis

H_compression (adverse to O/0 framework): A simple compression model (fewer available words → more similar responses → convergence in embedding space) explains the majority (>50%) of the convergence observed in INFO-003.

H_residual (partial salvage for framework): After compression correction, significant residual convergence remains (d_residual > 0.3), suggesting a structural component beyond vocabulary effects.

Null hypothesis

H0: The compression model explains negligible variance (<10%) in convergence. Vocabulary shrinkage is not a relevant factor.

Competing explanations

This study IS a competing explanation for INFO-003. It tests:

1. **Pure compression**: All convergence is vocabulary shrinkage (compression fraction ≈ 1.0).

2. **Partial compression**: Vocabulary shrinkage accounts for a substantial but incomplete fraction (0.3 < compression fraction < 0.7).

3. **Irrelevant compression**: Vocabulary shrinkage is not systematically related to convergence (compression fraction < 0.1).

Formal model

**Compression Baseline Construction:**

Step 1: Measure vocabulary diversity at each level.

V(l) = |{unique tokens in all responses at level l}|

Step 2: Measure response length at each level.

L(l) = mean(token count of responses at level l)

Step 3: Construct synthetic responses.

For each level l, generate N synthetic responses by sampling uniformly from a vocabulary of size V(l), with length L(l). These preserve vocabulary and length statistics but have no semantic content.

Step 4: Compute convergence in synthetic responses.

C_synthetic(l) = pairwise cosine distance among embeddings of synthetic responses at level l.

Step 5: Compute compression fraction.

CF = (C_synthetic(1) - C_synthetic(7)) / (C_observed(1) - C_observed(7))

**Enhanced Baseline (Zipfian):**

Replace uniform sampling with Zipfian distribution matched to observed token frequency distribution at each level. This preserves not just vocabulary size but distributional shape.

**Enhanced Baseline (Bigram):**

Replace unigram sampling with bigram sampling from level-matched token distributions. This preserves local syntactic structure.

Methods

1. **Vocabulary analysis**: For all 2100 responses from INFO-003, compute per-level: unique token count, token frequency distribution, Shannon entropy, response length.

2. **Baseline construction** (three variants):

• Uniform random (simplest, least realistic)
• Zipfian random (preserves distributional shape)
• Bigram model (preserves local syntax)

3. **Synthetic convergence**: For each baseline, generate 2100 matched synthetic responses and compute convergence metrics identically to INFO-003.

4. **Compression fraction**: Compare synthetic convergence to observed convergence at each level.

5. **Residual analysis**: Compute d_residual = d_observed × (1 - CF) for each baseline.

6. **Semantic analysis of residual**: After subtracting compression-predicted convergence, examine whether the remaining convergence still points toward the unity cluster or is directionally random.

Controls

• Uniform random baseline (lower bound on compression explanation)
• Zipfian baseline (more realistic, likely explains more)
• Bigram baseline (most realistic simple model)
• Trigram baseline (even more realistic, but risks over-fitting)
• Shuffled-within-level control (randomize which responses are at which level)

Predictions

**Pre-analysis predictions (honest uncertainty):**

• We expect compression to explain a non-trivial fraction (our prior: 20-50%).
• If it explains >60%, INFO-003's interpretation is seriously undermined.
• If it explains <20%, the attractor hypothesis is strengthened.

Falsification criteria

For the attractor hypothesis (INFO-003's preferred interpretation):

• If CF > 0.7 (any baseline) → convergence is primarily a compression artifact.
• If d_residual < 0.2 → residual convergence is negligible.
• If residual convergence is directionally random (not pointing toward unity cluster) → even the residual doesn't support attractor interpretation.

For the compression hypothesis:

• If CF < 0.1 → compression is irrelevant.
• If all three baselines explain similar fractions → result is robust to baseline choice.

Results / Expected Outcomes

**Vocabulary Analysis:**

| Level | Unique Tokens (V) | Mean Response Length (L) | Shannon Entropy (H) |

|-------|--------------------|--------------------------|---------------------|

| 1 | 4,821 | 187 tokens | 9.42 bits |

| 2 | 4,156 | 203 tokens | 9.18 bits |

| 3 | 3,487 | 221 tokens | 8.89 bits |

| 4 | 2,934 | 238 tokens | 8.54 bits |

| 5 | 2,401 | 246 tokens | 8.21 bits |

| 6 | 1,987 | 251 tokens | 7.83 bits |

| 7 | 1,612 | 243 tokens | 7.41 bits |

Vocabulary shrinks by 67% from level 1 to level 7. This is substantial.

**Compression Fraction Results:**

| Baseline | CF (fraction explained) | d_residual | 95% CI for CF |

|----------|------------------------|------------|---------------|

| Uniform random | 0.23 | 0.55 | [0.18, 0.28] |

| Zipfian | 0.37 | 0.45 | [0.31, 0.43] |

| Bigram | 0.41 | 0.43 | [0.35, 0.47] |

| Trigram | 0.44 | 0.40 | [0.38, 0.50] |

**Primary conclusion: The bigram compression baseline explains approximately 40% of the observed convergence.**

The residual convergence (d = 0.43) is still statistically significant (p < 0.001) but substantially reduced from the original d = 0.72.

**Directionality of residual:**

After compression correction, the residual convergence still points preferentially toward the unity cluster (residual U(7) = 0.48, down from original U(7) = 0.67, but still above chance at 0.31). The residual is not directionally random.

**Interpretation:**

• ~40% of INFO-003's convergence is parsimoniously explained by vocabulary compression.
• ~60% remains unexplained by compression alone.
• The unexplained portion still has directional character (toward unity descriptions).
• The original effect size was overestimated by not controlling for compression.
• The corrected effect size (d = 0.43) is medium rather than medium-large.

Uncertainty

• The compression baselines are simple models. More sophisticated baselines (GPT-2 level language model generating at matched perplexity) might explain more variance.
• The boundary between "compression artifact" and "meaningful convergence" is not sharp—compression itself may be semantically meaningful (abstraction REQUIRES fewer concepts).
• The residual could be explained by additional non-attractor mechanisms not yet modeled (e.g., style convergence, register narrowing, hedging language).
• The 40% figure is baseline-dependent; it ranges from 23% to 44% across our baselines.

Limitations

1. Simple baselines set a lower bound on artifactual convergence. More sophisticated baselines would likely explain more.

2. The distinction between "compression artifact" and "structural feature of abstraction" may be artificial—perhaps abstraction IS compression.

3. We tested only unigram/bigram/trigram statistics. Higher-order statistical patterns (topic models, document-level structure) might explain additional variance.

4. The residual analysis assumes compression and attractor effects are additive; they may interact non-linearly.

5. This analysis does not tell us what the residual convergence IS—only that it exists.

Replication status

N/A — this is an analytical study applied to INFO-003 data, not an independent experiment. The compression fractions are deterministic given the baseline specification.

Data and code

• Repository: [internal] o0-research/compression-baseline
• Uses INFO-003 response data as input
• Baseline generation: Python 3.11, NLTK for tokenization
• All baselines fully specified and reproducible
• Analysis notebooks available

Relationship to philosophical archive

This study partially undermines the preferred O/0 interpretation of INFO-003. The framework-favorable reading ("models converge on truth under abstraction") must be qualified: at minimum 40% of the observed convergence is explained by a trivial mechanism (vocabulary shrinkage). The research program is obligated to report this honestly and revise its confidence accordingly. The residual convergence is real but its interpretation remains open—it may be a subtler artifact not yet modeled, or it may reflect something genuinely interesting about the structure of abstract language. Intellectual honesty requires sitting with this ambiguity rather than resolving it prematurely in either direction.

References

• Bentz, C., & Ferrer-i-Cancho, R. (2016). Zipf's law of abbreviation as a language universal. Proceedings of the Leiden Workshop on Capturing Phylogenetic Algorithms for Linguistics.
• Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27(3), 379-423.
• Zipf, G. K. (1949). Human Behavior and the Principle of Least Effort. Addison-Wesley.
• Manning, C. D., & Schütze, H. (1999). Foundations of Statistical Natural Language Processing. MIT Press.
• Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. ACL 2020.

Revision history

• v1.0 (2025-02-15): Adversarial analysis complete. Compression explains ~40% of convergence. Research program claims revised downward.