Working tools you can open.

Can human-defined commitments remain effective when an AI agent moves from conversation to tool use?

SERAPH evaluates proposed tool calls before they run, binds permitted actions to execution, and records what the host reports afterward. It makes human direction inspectable where an agent acts.

Evidence and limits

The public repository includes adapters, tests, a 17-case governance corpus, and Lean models. These examine defined behavior; live-host reliability and independent reproduction remain open.

Public record

Public source, adapters, tests, a 17-case governance corpus, and Lean models.

When is a human–AI working session losing clarity or narrowing a person’s choices?

Human–AI work can drift before anyone notices that choices have narrowed. A traceable signal could make that shift visible in time to respond.

Evidence and limits

If an alert cannot be traced to observable behavior or does not help a person choose a corrective action, it has not shown practical value. The public source exposes the rules for inspection.

Public record

Public source and documentation

Can descriptors derived from polymer structures help examine five material property targets?

I built this dataset to test what human intuition and AI reasoning could produce together. Coherence Systems Architecture and perturbation techniques guided the search through latent space; the published features let others examine the result and test its predictive value.

Evidence and limits

The contribution is the published dataset and its feature design. Any claim of predictive improvement needs a reproducible comparison on the target properties.

Public record

Entry for NeurIPS – Open Polymer Prediction 2025. Published on Kaggle as “Crystalline Polymers and Quantum Stability” for five polymer property targets. The features are structure-derived, physics-inspired proxies.

Can patterns of premature agreement and other AI responses be made visible in a conversation?

AI systems can confirm a speculative premise before examining it. CHANDRA makes that behavior inspectable so collaborators can question the response instead of mistaking agreement for evidence.

Evidence and limits

The repository includes code tests, but says empirical validation studies have not yet been conducted. Its behavioral labels do not establish subjective AI states.

Public record

Public Python source, tests, examples, and validation plan

Can computation, meaning, and alignment be described by a shared mathematical condition for when a result becomes settled?

Formal models make broad ideas precise enough to challenge. The Lean source shows what follows from stated assumptions and where an interpretation still needs a separate test.

Evidence and limits

The proofs establish results within their stated definitions and assumptions. They do not establish a universal physical law or a quantum or neural mechanism; those interpretations need distinct predictions and empirical tests.

Public record

Public Lean 4 source, three papers, a claim-status table, and documented counterexamples.

Can changes in a human–AI conversation be measured when the conditions shift?

As AI enters more decisions, people need to recognize when an interaction changes their clarity, intent, or room to choose, and whether they can recover.

Evidence and limits

If these measures cannot be scored reliably or do not behave as registered when conditions change, the proposed interpretation needs revision. The registered plan governs that decision; no confirmatory outcome is claimed here.

Public record

OSF registration · 17 May 2026

Do filed quantum hardware predictions survive the tests written for them?

A physical prediction earns trust by meeting the test filed for it, including when the result is negative. The visible job records show what was predicted, measured, and corrected.

Evidence and limits

The first two Bell fidelity tests were negative under their filed rules. A later verdict was withdrawn; later results are kept as separate findings.

Public record

IBM Quantum job IDs · Test #1: d93t71fu62ks7395a110 · Test #2: d93tv0mvtlqs73ftkh80. Preregistered protocols, outcomes, and corrections remain public.

What has to stay clear as people, AI, and ideas change one another?

When people and AI build together, unclear roles and dependencies make decisions hard to inspect or recover. This method makes purpose, limits, and responsibilities explicit.

Evidence and limits

If the method cannot make a project’s purpose, limits, dependencies, and signs of failure explicit before use, it is not doing the work it promises. The specification makes those steps inspectable.

Public record

Working specification · v1.1 · PDF