Research overview
Constraint graphs, tension, and information transfer.
The research program is evidence-first: build small mechanisms, expose their telemetry, run controlled comparisons, and report measured results.
Research discipline
Claims move only as far as the receipt allows.
The archive is intentionally built around pressure: what the system claims, what the setup actually tested, what the result says, and what would make the claim weaker.
Current pressure
Strongest public evidence is currently mechanism-scale and benchmark-bound. Broad architecture claims remain hypotheses until paired with exact baselines, replay commands, and artifact hashes.
Focus
Graph-Based Reasoning
Represent reasoning work as nodes, edges, activation, constraints, pressure, and revision history.
Focus
Constraint Propagation
Study how local updates change neighboring claims, model states, or proof traces over time.
Focus
Tension Dynamics
Track mismatch, contradiction pressure, and relaxation as measurable telemetry rather than hidden vibes.
Focus
Interpretable AI Systems
Prefer inspectable fields, replayable traces, provenance, and benchmark receipts over loose claims.
Claim discipline
Architecture advantages are treated as hypotheses until backed by benchmark receipts, replayable experiments, or source-linked evidence.
Archive section
Core question
Can small graph/tension mechanisms make reasoning state, contradiction pressure, and repair choices more inspectable without overstating capability?
Archive section
Current hypotheses
Local relaxation can reduce bounded graph tension; provenance can improve claim handling; learned proposal models may be useful when verifier traces stay explicit.
Archive section
What would falsify / weaken the claims
Matched baselines that erase the advantage, receipts that fail to reproduce, telemetry that does not localize errors, or benchmark gains that vanish outside toy setups.
Archive section
Current receipts
Proof Bank entries currently support narrow toy and mechanism claims. Draft entries remain weaker until public artifacts and replay commands are attached.
Archive section
Next benchmark pressure
Matched softmax-vs-sigmoid comparisons, stricter CIG contradiction tests, exact seeds/commits, and verifier-loop receipts for proof-ranker integrations.