Evidence and public references
This is the clean public proof surface: the flagship benchmark, supporting materials, public code, archives, and the shortest references that substantiate the core claims.
Use this page when you want the papers, the benchmark, the artifact ledger, and the verification tools in one place.
Program evidence by maturity level.
The public evidence is proof-graded: one resolved flagship, one transfer calibration, an active validation lane on a new substrate, an active adapter foothold, and formal objects that preserve the measurement contract.
Resolved flagship: random 3-SAT
Measured constructive accessibility transition at n=1000, confirmed finite-asymptote prediction, oracle bridge, rewind depth, and late visible failure.
Transfer calibration: graph coloring
A second oracle-backed CSP shows process-side collapse inside an alive landscape, with a different morphology from K-SAT.
Active validation: AI reasoning
A first case study on a math-tuned reasoning model measures scaffold-lattice confinement: 1,840 of 1,842 numeric continuations land in the wrong arithmetic family while the target lies outside it. Prompt-baseline and replication runs are in progress. See case study ↓
Active adapter: scheduling
Scheduling is an internal adapter foothold where forward construction commits irreversibly. It should be read as an extension lane, not a completed flagship result.
The research-facing benchmark and public proof tools.
The interactive benchmark, the flagship citation block, the artifact ledger, and the verification tool live here so the homepage does not have to carry the full proof surface.
Nine system sizes through n = 1,000 anchor the flagship result. Oracle-verified trap analysis, bounded rewind, trajectory-space diagnostics, a confirmed scaling prediction, and a second combinatorial domain all support the same deeper claim: the goal can remain valid after the process has lost the route.
One place for the core public references.
Use this page for the public references: the flagship paper, supporting materials, public code, archives, and the primer.
Constructive Accessibility from Committed Prefixes in Random 3-SAT
The first resolved benchmark: a measured constructive accessibility boundary, late visible failure, exact-prefix deadness, trap depth, and strong within-instance path dependence.
Constructive Accessibility Instrument
The public reference implementation and release track for the software line. This is the code-side anchor, separate from the website and separate from the service documents.
Supporting materials and large-n diagnostics
Finite-size scaling, hazard diagnostics, bridge and rewind analyses, robustness controls, and the transfer calibration lane sit here as the public understructure for the flagship result.
Graph coloring as a second oracle-backed adapter
The graph-coloring lane shows a process-side collapse inside a still-alive colorability landscape. It is strong evidence that the measurement contract travels, but it is not yet claimed as a full mechanistic twin of the K-SAT flagship.
The shortest evidence summary
A process-indexed boundary exists.
The flagship benchmark resolves a constructive accessibility transition well below the classical satisfiability threshold.
Failure has a trajectory-space fingerprint.
Late visible collapse, bridge-verified deadness, rewind depth, and strong path dependence all support the same route-side interpretation.
The measurement contract transfers.
A second oracle-backed combinatorial domain already shows the same deeper process-versus-landscape split with a different surface morphology.
The instrument reads structure without an oracle.
A first reasoning-substrate case study shows committed-prefix futures confined to a wrong arithmetic family that excludes the target. The signal is structured, finite-M, and model/policy-indexed; it is not a non-extendability certificate.
The reasoning lane has its first case study.
A reasoning trace is a path of commitments. The reasoning lane measures the future field induced by a committed prefix — rollouts continued from the exact state, with the resulting answer distribution treated as the observable. This is the same conceptual object the SAT flagship measures, run on a different substrate where no exact oracle exists.
Scaffold-lattice confinement on an arithmetic reasoning trace
On a floor-sum target with correct answer 532, a 7B math-tuned reasoning model produced 1,840 of 1,842 numeric continuations inside the wrong arithmetic family 8ℤ. The target lies outside that family. Across three checkpoints and 30 rewind conditions — 2,112 continuations total — no correct continuation was observed. At 25% progress the field was answer-diverse but already family-locked; by 50% it had consolidated onto the wrong attractor 56.
Prompt-baseline and replication measurements
The next measurement is early-prefix tomography: a prompt-level baseline at higher rollout count plus dense early checkpoints, designed to localize whether correct mass exists before the wrong arithmetic family appears. A predictive cross-problem replication panel tests whether the wrong-family modulus is structural or coincidental. Either outcome sharpens the measurement program.
Claim status
The case study supports finite-M, model/policy/budget-indexed claims about scaffold confinement and non-recovery under the declared instrument. It does not yet establish a live-to-dead transition, absolute irrecoverability, or model-general behavior.
The reasoning substrate has no exact oracle analogous to the SAT bridge check. Reasoning-side claims are persistence claims, not non-extendability certificates. The full reasoning of this distinction is in the case-study draft.
Public artifacts, versions, and verification tools.
This is the control surface for public references and downloadable artifacts. Use it when you need the ledger, the mirror list, or in-browser SHA-256 verification.
Where to start
Artifact ledger
Public artifacts, versions, and mirrors. The site is the hub — OSF, GitHub, Zenodo, and later arXiv are the public destinations.
Software release version, release link, and SHA-256 hashes are read from /artifacts.json. Deploy this file from GitHub release automation instead of editing version strings by hand.
| Artifact | Version | Status | Primary | Mirrors |
|---|---|---|---|---|
| Flagship paper Constructive Accessibility from Committed Prefixes in Random 3-SAT |
v1.0 | Public | OSF | Zenodo · arXiv forthcoming |
| Public demo Constructive Accessibility Instrument (public reference implementation) |
v0.3.2 | Public after tag | GitHub Release | Repository · Zenodo release forthcoming |
| Interactive figures Epistemic, process diagnostics, scientific benchmark, and trajectory atlas figures |
v4–v6 | Public | Website | GitHub |
| Supplement Extended data, validation runs, coloring adapter |
v1.0 | Public | OSF | Zenodo |
| Primer The Builder's View — a plain-language introduction to constructive accessibility |
v4.0 | Public | Website | — |
| Primer (audio) The Builder's View — spoken narration (~20 min) |
v1.0 | Public | Website | — |
Verify a download (SHA-256)
Pick a file you downloaded. This computes SHA-256 in-browser and checks it against the expected hash. Nothing is uploaded.
Concept first
If you need the shortest conceptual path before looking at papers, start with the concept pages.
Constructive accessibilityPlain language first
If you want the least technical entry point, start with the primer.
The Builder's ViewVerified? Start intake.
If the evidence above maps to the failure mode you are already seeing, the next step is the intake form. It is fine if some answers are still rough.
Open the intake form →Code path
If you need the public software and release track, use the repository and release pages directly.
GitHub repository