#G-A-S-O-I-R-E (Galilean-Aristotelian-Sampling-Observational-Inductive Reasoning Engine)
Dissolving apparent autonomy into mechanisms — with every published number regenerated by code.
An exercise in computational epistemology: convert "self-generated" behavior into causal mechanism, and hold every claim to the standard of mechanical reproduction.
OIRE is a research architecture and benchmark suite built around one operation: dissolving apparent autonomy. Given a phenomenon described as self-generated — a moon that "rotates on its own axis," a market that "recovers by itself," a bulb that "responds to a distant observer" — the engine produces a causal explanation of exactly three kinds:
| Mechanism | What it means | Example |
|---|---|---|
| Externally forced automaton | An outside driver accounts for the behavior | AC currents drive an "idle" motor |
| Reference-frame artifact | The motion is a coordinate-system illusion | Tidal locking: the Moon has no net rotation in the Earth-fixed frame |
| Poised-critical detector | The system sits near a threshold that amplifies tiny perturbations | A bulb reacting to a hand two meters away |
…or it flags an explicit residual: "information insufficient" — because an engine that never says "I don't know" is not reasoning, it is confabulating.
The repository evaluates this idea three ways, end to end, from pinned seeds:
- Symbolic engine suite — 20 labeled phenomena across five domains, including five insufficient-information cases whose correct behavior is a flagged residual.
- Galileo Active-Observer benchmark — reconstruct Io's orbit from ≤12 noisy angular samples under a simplified 1610 observing geometry; five sampling policies compared in a paired Monte Carlo design (N=200 per noise level × 6 noise levels), including a structural-memory ablation.
- Cross-domain transfer — one causal skeleton (damped driven oscillator) instantiated as pendulum / RLC circuit / population dynamics; warm-vs-cold identification across all ordered domain pairs (N=100 each).
These are the values this repository's code actually produces. Where they contradict the 2026 design drafts, the drafts lost. See PROVENANCE ledger.
- Structural memory is the dominant effect. Removing the analogizer prior degrades orbit-reconstruction accuracy ×1.29 and convergence speed ×1.32 versus the warm baseline — confidence intervals disjoint from every warm-started condition.
- Fancy sampling ≠ better sampling. Extrema-targeted active sampling shows no significant gain over uniform sampling at benchmark scale (Δ = −0.0019°, 95% CI [−0.0042, +0.0003]); uncertainty-driven (bootstrap-disagreement) sampling attains the lowest RMSE but with overlapping intervals — suggestive, not conclusive.
- Naive cross-domain transfer fails instructively. Generalization score 0.00: transferred priors never meet the ≥20%-acceleration criterion and actively harm far-domain pairs (up to 9× longer training) — a clean, quantitative demonstration of the false-analogy failure mode.
- The engine meets its specification exactly: 100% classification on determinate cases, 100% residual-flagging on indeterminate ones — validating implementation fidelity, honestly framed as such.
| Policy | RMSE@30d (deg) | 95% CI | Samples to converge |
|---|---|---|---|
baseline_uniform |
0.0104 | [0.0091, 0.0118] | 6.3 |
varA_active_extrema |
0.0123 | [0.0106, 0.0142] | 5.8 |
varB_conservative |
0.0158 | [0.0142, 0.0174] | 6.6 |
varC_graph |
0.0095 | [0.0084, 0.0106] | 6.3 |
varD_coldstart |
0.0135 | [0.0115, 0.0156] | 8.4 |
Reference noise σ = 0.01° · Active-sampler accuracy gain vs baseline: -0.0019° (95% CI [-0.0042, 0.0003]) · Convergence delta: 0.48 samples (95% CI [-0.07, 1.02]) Generalization score: 0.00 · Engine classification accuracy: 100.0% · Insufficient-info flag rate: 100.0%
All values generated by experiments/ from pinned seeds — see results/*/manifest.json. Reproduce with make reproduce.
Benchmark figures (generated by experiments/run_galileo.py)
 |
 |
| RMSE@30d vs measurement noise, by policy | Mean samples to converge, σ = 0.01° |
Most research repos say "code coming soon." This one makes reproduction a mechanical property:
make setup # pip install -e ".[dev]" (Python ≥ 3.11)
make test # 80 tests, 99% coverage on the package
make experiments # regenerate EVERY result from seed 20260826
make emit # regenerate paper tables/macros + the README block above
make reproduce # re-run everything; fail on any byte-level drift- Every experiment writes a manifest: git SHA, environment versions, seeds, SHA-256 of every output file.
- CI re-runs all experiments from scratch and fails if any published hash drifts.
- Every table and number in the paper is
\inputfrom generated files — zero hand-transcribed values anywhere. - Wall-clock timings live in unhashed manifest metadata only; hashed outputs are deterministic by construction.
| Path | Contents |
|---|---|
oire/engine/ |
Layer 1 — symbolic autonomy-dissolution engine (v2.0 reference implementation) |
oire/observer/ |
Layer 2 — vectorized Kepler propagation, sampling policies, bounded multi-start estimation |
oire/transfer/ |
Layer 3 — causal-skeleton transfer protocol with shrunk priors |
experiments/ |
Seeded runners + artifact generation (tables, macros, README block) |
results/ |
Committed outputs with SHA-256 provenance manifests |
paper/ |
LaTeX source → PDF via CI (\inputs generated tables only) |
docs/ |
Architecture · Methodology · Reproducibility · Provenance ledger |
docs/legacy/, media/legacy/ |
Original 2026 drafts & media, preserved verbatim |
- Architecture — module map, data flow, design rules
- Methodology — the six-step generative design method; corpus-derived provenance stated openly (historical texts used as design stimuli, not sources of claims)
- Reproducibility — seeds, overrides, CI semantics
- Provenance ledger — projected (drafts) vs measured (this repo), claim by claim — including the claims that did not survive
If you use this work, please cite:
@software{robinson2026oire,
author = {Robinson, Michael Forsythe},
title = {OIRE: Autonomy-Dissolver Reasoning Engine},
year = {2026},
version = {0.1.0},
orcid = {0009-0002-8487-759X},
url = {https://github.com/Michaelrobins938/oire}
}See also CITATION.cff.
Code: MIT · Paper & docs: CC-BY-4.0
Michael Forsythe Robinson — Independent Researcher
ORCID 0009-0002-8487-759X · GitHub @Michaelrobins938 · forsythepublishing@gmail.com