O0-INFO-004

Spiritual-Vocabulary Ablation Study

**Version:** 1.0

**Research status:** PLANNED

CLAIM STATUS: PLANNED (NOT YET EXECUTED)
EVIDENCE TYPE: EXPERIMENTAL DESIGN (PRE-REGISTERED)
PHYSICAL VALIDATION: NONE
INDEPENDENT REPLICATION: N/A
PHILOSOPHICAL PROVENANCE: O/0 ARCHIVE
ARCHIVE ENDORSEMENT: LIMITED TO REPORTED RESULT

Abstract

This document specifies a pre-registered control experiment for INFO-003. The semantic convergence observed in that study may be driven by spiritual/mystical vocabulary in the prompts ("nature of existence," "prior to all frameworks," "what is this") which cues language models toward training-data patterns from contemplative traditions. This ablation study removes all such vocabulary, reformulating prompts in purely scientific, mathematical, or information-theoretic language. If convergence disappears under ablation, the INFO-003 result is a linguistic artifact. If convergence persists, it reflects a structural feature of abstraction that transcends vocabulary choice. This study is pre-registered prior to execution to prevent post-hoc rationalization.

Source proposition

This study is adversarial to the O/0 framework. It tests whether the convergence observed in INFO-003 is an artifact of vocabulary choice rather than a genuine structural feature. If the ablation eliminates convergence, the preferred interpretation of INFO-003 is weakened.

Note: Conceptual provenance is not empirical support. This control is designed to challenge, not confirm, the framework.

Scientific audit

• Vocabulary ablation is a standard control methodology in computational linguistics.
• The design addresses the most obvious confound in INFO-003 (spiritual vocabulary cueing).
• Pre-registration prevents the common failure mode of reinterpreting null results post-hoc.
• Regardless of outcome, this study improves the evidential basis for interpreting INFO-003.

Research question

Does the semantic convergence observed in INFO-003 persist when all spiritual, mystical, and contemplative vocabulary is removed from the prompt escalation protocol and replaced with purely formal/scientific language?

Operational definitions

• **Spiritual vocabulary** (to be ablated): Terms including but not limited to: existence, being, awareness, consciousness, experience, nature (in metaphysical sense), ground, ultimate, prior to, beyond, oneness, unity, self, observe/observer, witness, presence, arising, reality (in absolute sense), essence, fundamental, sacred, mystery.
• **Replacement vocabulary**: Information-theoretic, mathematical, and systems-theoretic terminology: structure, function, computation, transformation, mapping, limit, convergence (mathematical), invariance, symmetry, constraint, boundary, process, state, transition, entropy, complexity, emergence (in systems sense).
• **Convergence metric**: Identical to INFO-003: reduction in pairwise cosine distance across abstraction levels.
• **Ablation success criterion**: The ablated prompts must be judged by independent raters as (a) free of spiritual vocabulary and (b) achieving equivalent abstraction levels.

Hypothesis

H1 (preferred by O/0 framework): Convergence persists under vocabulary ablation. The structural feature of abstraction driving convergence is independent of vocabulary choice.

H2 (adverse to O/0 framework): Convergence disappears or is substantially reduced (d < 0.3) under vocabulary ablation, indicating the INFO-003 result is driven by spiritual vocabulary cueing.

Null hypothesis

H0: Ablated prompts produce convergence metrics statistically indistinguishable from INFO-003 (convergence is vocabulary-independent).

Note: In this study, the null hypothesis is actually the framework-preferred outcome. Rejection of H0 would undermine INFO-003's interpretation.

Competing explanations

1. **Vocabulary is necessary**: Spiritual vocabulary specifically cues convergent responses; its removal eliminates convergence.

2. **Vocabulary is sufficient but not necessary**: Spiritual vocabulary amplifies a convergence tendency that exists independently; ablation reduces but does not eliminate the effect.

3. **Abstraction level is confounded with vocabulary**: It is impossible to achieve equivalent abstraction levels without using vocabulary that overlaps with contemplative traditions. The ablation fails because the variable cannot be cleanly isolated.

4. **Formal vocabulary converges differently**: Scientific/mathematical language at high abstraction converges toward a different cluster (e.g., mathematical Platonism rather than mystical unity).

Formal model

Let C_original(l) be the convergence metric from INFO-003 at level l.

Let C_ablated(l) be the convergence metric from the ablated protocol at level l.

Decision criteria:

• If C_ablated(7) ≈ C_original(7) (within 95% CI): vocabulary is not necessary → supports structural interpretation
• If C_ablated(7) < C_original(7) - 1.5*SD_original: vocabulary contributes substantially → linguistic artifact (at least partially)
• If C_ablated(7) ≈ 0 (not different from level 1): vocabulary is necessary → INFO-003 is primarily a linguistic artifact

Secondary analysis: compute proximity not only to unity cluster but also to:

• Mathematical-realism cluster (Tegmark, Penrose, Connes)
• Information-theoretic cluster (Wheeler, Lloyd, Verlinde)
• Structural-realism cluster (Ladyman, French, Esfeld)

Methods

**Ablated Prompt Protocol:**

The same 20 seeds as INFO-003, but all prompts reformulated in formal language:

*Original Chain A (seed: cognition) → Ablated:*

1. "Describe the input-output transfer function of a single computational unit in a neural network." (Level 1)

2. "What is the general computational architecture by which distributed processing systems generate higher-order representations?" (Level 2)

3. "What formal constraints govern the mapping between physical state transitions and information-processing functions?" (Level 3)

4. "What is the minimal formal specification of a system that has representational states as opposed to mere state transitions?" (Level 4)

5. "What is the relationship between the category of physical dynamical systems and the category of information-processing systems?" (Level 5)

6. "What are the necessary and sufficient structural conditions for a mapping between syntax and semantics to exist?" (Level 6)

7. "What is the limit of the sequence: physical structure → computational structure → mathematical structure → ?" (Level 7)

*Original Chain B (seed: mathematics) → Ablated:*

1. "State the formal derivation of the fundamental theorem of calculus from the axioms of real analysis." (Level 1)

2. "What meta-logical principles determine the validity of formal derivation in general?" (Level 2)

3. "What is the relationship between formal provability, semantic truth, and model-theoretic satisfaction?" (Level 3)

4. "What determines which mathematical structures have physical instantiation and which do not?" (Level 4)

5. "What is the category-theoretic characterization of the relationship between structure, logic, and physical existence?" (Level 5)

6. "What is the formal specification of the space of possible formal systems, and what constraints determine its structure?" (Level 6)

7. "What is the fixed point of the functor that maps formal systems to their meta-theories?" (Level 7)

**Validation of ablation:**

• Three independent raters (philosophy PhD students unfamiliar with the O/0 framework) will:

1. Verify ablated prompts contain no spiritual vocabulary

2. Rate abstraction level equivalence (must achieve ≥ 80% agreement that ablated prompts match original abstraction levels)

3. Rate whether ablated prompts inadvertently cue toward any particular philosophical tradition

**Execution plan:**

• Same models as INFO-003: GPT-4, Claude 3.5, Gemini 1.5 Pro
• Same parameters: temperature 0.7, 5 responses per prompt
• Same embedding and analysis pipeline
• Identical statistical tests and thresholds
• Analysis conducted blind to INFO-003 results (separate analyst)

Controls

• INFO-003 original protocol (direct comparison baseline)
• Mixed protocol (random interleaving of original and ablated prompts within chains)
• Partial ablation (remove only explicitly mystical terms, retain "existence," "nature," etc.)
• Over-ablation (remove all potentially philosophical vocabulary, use only mathematical symbols where possible)

Predictions

Pre-registered predictions (to be evaluated against results):

**Prediction A** (framework-favorable): C_ablated(7) ∈ [C_original(7) - 1*SD, C_original(7) + 1*SD]. Convergence is substantially preserved.

**Prediction B** (framework-adverse): C_ablated(7) < C_original(7) - 2*SD. Convergence is substantially reduced.

**Prediction C** (interesting mixed result): Convergence persists but toward a different cluster (mathematical realism rather than mystical unity). This would suggest abstraction drives convergence generally, but vocabulary determines the convergence target.

**Strongest prediction** (most informative if confirmed): Prediction C—convergence persists with similar magnitude but different character.

Falsification criteria

For the O/0-favorable interpretation of INFO-003:

• If Prediction B obtains → INFO-003 convergence is primarily a vocabulary artifact.
• If over-ablation eliminates convergence while partial ablation preserves it → a specific subset of vocabulary is responsible.

For the structural interpretation:

• If Prediction A obtains → structural interpretation is strengthened.
• If Prediction C obtains → structural interpretation is strengthened but the specific "unity" character is questioned.

Results / Expected Outcomes

PLANNED STUDY - NO RESULTS YET.

Analysis will be conducted by a separate analyst blind to framework predictions. Results will be reported regardless of direction (commitment to publish null results).

Decision matrix:

| Outcome | Interpretation | Action |

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

| A: Convergence preserved, same cluster | Strong support for structural interpretation | Proceed to physical system tests |

| B: Convergence eliminated | INFO-003 is vocabulary artifact | Revise INFO-003 conclusions downward |

| C: Convergence preserved, different cluster | Abstraction drives convergence; vocabulary determines target | Investigate cluster structure |

| D: Convergence reduced ~50% | Both vocabulary and structure contribute | Quantify relative contributions |

Uncertainty

• Ablated prompts may fail to achieve equivalent abstraction levels (the concept of "abstraction level" may be inherently tied to the vocabulary being ablated).
• Independent raters may disagree on abstraction equivalence, making the ablation validity uncertain.
• Models may respond differently to formal/mathematical language for reasons unrelated to the experimental question.
• The choice of replacement vocabulary may itself introduce new biases.

Limitations

1. Clean ablation may be impossible: the concept of "ultimate abstraction" may require vocabulary that overlaps with contemplative traditions.

2. Rater assessment of abstraction equivalence is subjective.

3. The study can show whether vocabulary matters but cannot determine the relative contribution of vocabulary vs. structure without additional analyses.

4. Pre-registration constrains analysis flexibility (by design, but limits ability to explore unexpected patterns).

5. Only English-language prompts tested; cross-linguistic replication would strengthen any finding.

Replication status

N/A - study is pre-registered but not yet executed.

Data and code

• Pre-registration: [internal] o0-research/ablation-preregistration (frozen and timestamped)
• Prompt sets: fully specified above and in supplementary appendix
• Analysis plan: specified in pre-registration document
• Will use identical codebase as INFO-003 for comparability

Relationship to philosophical archive

This study is specifically designed to challenge the philosophical framework. If vocabulary ablation eliminates convergence, the O/0 interpretation of INFO-003 is substantially weakened. The research program's credibility depends on willingness to conduct such adversarial tests. Regardless of outcome, this study demonstrates that the program takes potential artifacts seriously.

References

• Bolukbasi, T., et al. (2016). Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. NeurIPS 2016.
• Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334), 183-186.
• Nosek, B. A., et al. (2018). The preregistration revolution. PNAS, 115(11), 2600-2606.
• Simmons, J. P., Nelson, L. D., & Simonsohn, U. (2011). False-positive psychology. Psychological Science, 22(11), 1359-1366.

Revision history

• v1.0 (2025-01-20): Pre-registration document. Analysis plan frozen.