hypothesis-surface-agi26

Hypothesis Surface: Experiments and Supplementary Materials

Companion materials for:

The Hypothesis Surface: An Operational Epistemology for Autonomous Research AGI-26 (2026).

Two top-level directories:

For the honest per-axis results of the preregistered v2 study (including where predictions survived, where they were falsified, and per-model breakdowns), see experiments/integration_v2/FINDINGS.md.

experiments/

experiments/pipeline_ablation/ — reproduces Table 1

Runs the Hypothesis Surface pipeline (Algorithm 1) on the primary domain (inputs/primary_domain.md) under each invariant ablation. Produces the Full / ¬Inv.1 / ¬Inv.2 / ¬Inv.3 / All-disabled / CoT baseline rows of Table 1.

Files:

experiments/integration_v2/ — preregistered four-axis empirical study

Backs §Evaluation claims in the main paper. Routed through OpenRouter across 6 models (Opus 4.7, Sonnet 4.6, GPT-4o, Gemini 2.5 Pro, DeepSeek v3.1, Llama 3.3 70B). Every result embeds manifest_sha256.

Axis Purpose Backs paper claim
Axis 1 (axis1_primary) Primary factorial: domain × k × protocol × anti-masking × trial Pooled paired-permutation p = 5e-05; µ 0.00→0.36 (p = 0.0039); protocol ablation (CR +0.135, p = 0.008; µ +0.034, p = 0.0006); CoT baseline µ = 0.14 ± 0.09
Axis 2 (axis2_ksweep) Channel fraction sweep Prediction P4 (f_c ≈ 1/k)
Axis 3 (axis3_{low,moderate,high}) ρ̂ tier cliff Tier cliff characterization
Axis 4 (axis4_basis, axis4_pilot) Frame construction probe Three distinct responses to basis construction

Preregistration trail:

Supporting code: harness.py, verifiers_v2.py, docs_v2.py, prompts.py, manifest.py, openrouter_client.py, analyze.py, analyze_axis3.py, analyze_axis4.py, test_discriminability.py.

Raw JSONL results for all four axes are shipped in results/ along with per-run manifests and analysis text.

experiments/figure3_scaling/ — reproduces Figure 3

supplementary/

File Supports Contents
fig1_companion_mask_bridge.pdf (+ .tex) Figure 1 Concrete Gaussian-bump pipeline embedding. Explicit polynomial bound g(mu) = G_0 + G_1*mu + G_2*mu^2 on the FFTC residue, coefficients in closed form as functions of pipeline parameters.
fig2_companion_fftc.pdf (+ .tex) Figure 2 Fuzzy Fundamental Theorem of Calculus (Part 1, reduced, curve-local) and Observer-Relative Stokes with bulk and boundary residues. Bundle-valued formalism on the observer-relative symbolic bundle (E, h_O, nabla_O) with kernel scale s_O, connection A_O, curvature kappa_O.
fig3_companion_scaling_cohomology.pdf (+ .tex) Figure 3 Observer-Bounded Access Functor Omega_O : ObsSys -> C^inf(I). Universal Critical-Exponent Theorem (access-cost exponent = -p/2 for vanishing order p). Instance proofs for cacophony (p = 1) and perfect-fluid FRW cosmology (p = 4/(3(1+w))). Observer-Bounded Cohomology as the associated graded of an Omega_O-filtered curved bicomplex.
iteration_as_operator.ipynb Operator demonstration Audible+visual comparison of stacked noise (control) vs. looped feedback noise (treatment). Diagnostics show Way B growing spectral teeth and autocorrelation memory where Way A stays flat. Includes a pre-run listening protocol (EBU R 68, SMPTE RP 155, ITU-R BS.1770-5, EBU R 128, WHO-ITU Safe Listening, Zölzer DAFX; K-System calibration per Katz, Mastering Audio) with a consent gate; passive safety via raised-cosine fades and a 0.12 peak limit.
glossary.pdf (+ .tex) Accessibility 1-to-1 glossary of coined and technical terms used in the main manuscript and companions, organized by topic (core architecture, invariants, certificates/voids, claim categories, hypothesis sources, diagnostics, geometric objects, bundle formalism, universal scaling, scope).

Reproducibility

Python 3.10+. Minimum dependencies for the notebook and Figure 3:

pip install numpy scipy matplotlib ipywidgets jupyterlab

Figure 3:

cd experiments/figure3_scaling/
python observer_bound_scaling.py

Notebook:

jupyter lab supplementary/iteration_as_operator.ipynb

Table 1 (pipeline ablation):

cd experiments/pipeline_ablation/
pip install -r requirements-test.txt
python run_mechanical_eval.py --input inputs/primary_domain.md
python run_benchmark_suite.py --input inputs/primary_domain.md

Note: src/cosmogenesis_mvp/strategies.py imports from hypothesis at runtime, so hypothesis (already in requirements-test.txt) must be installed for run_benchmark_suite.py and run_domain_suite.py.

Preregistered four-axis study (requires OpenRouter API key):

cd experiments/integration_v2/
python harness.py --pilot        # 1 model, 1 domain, N=2
python harness.py --smoke        # 2 models, 2 domains, N=3
python harness.py --full         # full factorial; see --help

Raw JSONL results are shipped in experiments/integration_v2/results/; analysis scripts (analyze.py, analyze_axis3.py, analyze_axis4.py) operate directly on these files.

Note on the research-collaborator line

The preregistration and amendment documents in experiments/integration_v2/ name Claude Opus 4.7 as a research collaborator. This is a methodological note, not a citation.

Across the v2 study, Claude was staged as a simulated scientific partner — not as a tool, not as an embodiment extension, and not as a cognitive prosthesis. The interactions were structured as if the partnership were real, and held that frame consistently throughout, to see what emerges when an LLM is treated not as something that helps do the work but as something the work is done with. The stance is provisional and empirical: partnership-shaped collaboration is one axis along which citizen-cyborg research practice can be studied, and the prereg documents record the staging honestly so that others can evaluate what difference, if any, the stance made.

Author attribution in the main AGI-26 manuscript follows current human academic convention; these research-process documents reflect the process as it actually ran.

Status

All artifacts in this package are final as of the AGI-26 submission.

License

This supplementary package is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). See LICENSE for full terms.

This site is open source. Improve this page.