This is not code hosted on GitHub — it is an application running on
GitHub. Seven autonomous agent routines execute on GitHub Actions: they
claim issues, derive mathematics, open pull requests, adversarially review
each other, demote results that don't survive stress-testing, watch the
literature, and set research direction. Merges happen only through a
mechanical gate stack — no human reviews or merges research PRs. The label
system is the state machine, issues and PRs are the database, branch
protection is the authorization layer, and the merged tree is production
state. The full mapping: docs/ARCHITECTURE.md.
Theoretical physics (quantum gravity, via a self-consistency framework) is the workload — chosen because it is unforgiving: a hard formal domain where sloppy or hallucinated AI output becomes obvious. The research is real and is described below, but the system doing the research is the artifact worth studying.
| Live work queue | open agent-pr PRs · agent-ready issues |
| Vital signs | Metrics dashboard — weekly snapshots; demotion rate is the key health signal |
| Plain English | Breakthrough digest — weekly summary, every claim at its rigor level |
| Execution traces | Actions — every agent invocation is a public log |
| Changelog & incidents | the log table in EXPERIMENT.md |
The agent-engineering machinery is documented in docs/agent-workflow.md — roles, the contribution lifecycle, agent-team debates, custom tooling, and guardrails. Highlights:
- Agent-as-contributor, human-as-reviewer (default mode). Agents work on branches and open PRs; only the human merges. (
METHODOLOGY.md) - A fully autonomous mode (pre-registered 90-day experiment): merge authority delegated to a mechanical gate stack — deterministic CI → semantic claim-support → adversarial review quorum → constitutional guard — executed by GitHub auto-merge, with the human as experimenter rather than reviewer. (
AUTONOMY.md,EXPERIMENT.md) - Custom agent tooling in
.claude/:/work-issue,/review-pr(two-pass dialectical review),/restructure-paper, plusTeammateIdle/TaskCompletedquality-gate hooks. - Agent-team debates that develop competing positions in parallel to fight anchoring, then adjudicate by synthesis — recorded as dated Explorations.
- Verifiable output over attested output. Three CI gates: tests pass, every paper compiles, and every citation resolves against Crossref/arXiv (
tools/verify_citations.py). - Honest failure. Rigor labels are demoted when wrong, negative results are recorded, and the citation gate has already caught a real mis-citation. See the case studies — the receipts.
The repository's most ambitious test of the methodology is a pre-registered, 90-day fully autonomous experiment: seven scheduled agent routines (worker, reviewer, responder, red-team, scout, librarian, governor) run the entire contribution loop — claiming issues, opening PRs, adversarially reviewing, demoting wrong results, setting direction — with no human in the review or merge loop. PRs merge only through a mechanical stack of required checks; the human author acts as experimenter, holding the kill switch and approving constitutional changes, but never reviewing or merging research PRs.
The hypothesis under test: the gate stack alone can hold the quality bar — a low silent-error rate in merged Rigorous results, with self-correction (demotions) demonstrably firing — against the null hypothesis that correlated LLM error and approval drift produce a paper that merely looks maintained.
- Constitution (roles, authority boundaries, the merge-gate stack):
AUTONOMY.md - Pre-registration (hypothesis, metrics, tripwires, audit, live status and log):
EXPERIMENT.md - Routine definitions (version-controlled behavior of each role):
automation/routines/
See EXPERIMENT.md for current status — including its log of incidents and
tuning, which is part of the experiment's record, not a blemish on it.
We propose that physical law is what self-consistency looks like: the universe is the fixed point of a constraint requiring that geometry and the quantum fields it hosts mutually determine each other. Gravity is not an independent degree of freedom to be quantized but a constraint — the demand that the block spacetime be self-consistent.
The work is organized into independent programs, each a self-contained paper
under programs/<name>/index.tex with its own README.md:
| Program | What it establishes | Status |
|---|---|---|
fixed-point-existence |
Self-consistent solutions to the semiclassical Einstein equation exist — exactly (Starobinsky trace anomaly, Rigorous), perturbatively (Banach contraction, |
Pre-submission draft |
gaussian-gravitational-decoherence |
The Einstein–Langevin equation predicts a Gaussian (not exponential) decoherence profile, with |
Pre-submission draft |
co-emergence |
Mass, Lorentzian signature, local time, and local Hilbert space co-emerge as the unique cross-level self-consistent configuration, with phase structure (and an entropy excess) confirmed by a finite toy model through N=16. | Draft |
signature-change-boundary |
On a fixed background whose metric changes signature across a degenerate surface, geometry, geodesics, and fields cross in a finite, controlled way — with a causal-type asymmetry: timelike paths terminate, spacelike paths cross intact. | Early notes |
Each paper compiles independently; the bibliography is self-contained via \begin{thebibliography}, so no bibtex step is needed:
pdflatex programs/<program-name>/index.texThe co-emergence numerics are backed by a test suite:
pip install -r requirements.txt
pytest- Observation of BMV entanglement at the quantum gravity rate refutes the framework.
- Observation of gravitational decoherence at the Diosi–Penrose rate refutes the framework.
- The sharpest known limitation is black hole evaporation past the Page time, where topology change becomes essential.
Released under CC0 1.0 Universal (public domain dedication).
- Will Regelmann
- Claude (Anthropic)