Research

Public summaries of the pre-registered program

Our research stance, restated: we pre-register hypotheses and experimental designs before measuring; no numeric claim below is a result — none exist yet, and we say so. Negative results are first-class outcomes and will be published here alongside positive ones.

Executor context custody & harness economics

Motivation: Agent harnesses spend heavily on coordination bookkeeping executed as model tokens, and the context that would make each call good is too heavy for users to supply.
Hypotheses: A verified obligation can replace transcript reconstruction when it is a sufficient statistic for the call; deterministically projected, role-scoped context improves first-pass acceptance; a versioned presentation profile can improve phrasing without weakening semantic quality; the combined system lowers coordination tokens/retries enough to justify its fixed cost.
Pre-registered design: Four matched arms — a frontier model with a conventional transcript loop vs a small model given (i) minimal verified-obligation context, (ii) deterministically projected dependency- and role-scoped context, (iii) the same plus a hand-engineered, versioned presentation profile (no training/optimization loop in this round; learning is deferred). Blinded task-quality scoring independent of the acceptance gate; separated token accounting; anti-gaming controls (frozen gate, held-out scoring, immutable context manifests).
Status: Pre-registered; execution awaits two internal prerequisites — the kernel-realization milestone and the frozen execution-binding anchor/reveal protocol. No measurements yet.
Falsified by: No material token/retry reduction at equal blinded quality; projected context failing to beat minimal context; presentation gains that coincide with held-out quality loss (gate exploitation).

Neuro-symbolic bottleneck & agent economics

Motivation: If a closed, verified coordination vocabulary lowers the entropy of what a model must produce, smaller specialized models may suffice for authoring — changing the economics of the whole path.
Hypotheses: Closing the output vocabulary makes constrained decoding useful; a small specialized model can reach a fixed intent-fidelity threshold at competitive end-to-end cost; constrained decoding's effect on semantic (not just structural) accuracy must be measured and may be negative; there exists a traffic volume above which the specialized path wins on total cost of ownership.
Pre-registered design: Matched arms over the same typed-interpretation → compiler path; a total-cost model including amortized training, ontology maintenance, clarification, and repair; the break-even volume is a reported output, not an assumption.
Status: Proposed research plan; not ratified; no measurements.
Falsified by: Small-model arms missing the fidelity threshold at competitive cost. (A break-even viability ceiling will be pre-registered with the plan's ratification.)

Kernel adequacy (GO/NO-GO)

Motivation: The coordination kernel must be adequate for coordination beyond value settlement before it is generalized — and that question deserves a pre-registered answer, not an accumulation of intuitions.
Hypotheses: The combined kernel (typed operations + verified plan + semantic layer) can represent a defined non-settlement coordination corpus within pre-registered thresholds.
Pre-registered design (drafted): An open classifier over a held-out corpus, baselines against the current kernel, numeric GO/NO-GO thresholds proposed for advance commitment; both outcomes first-class — only a later GO decision would authorize generalizing the kernel.
Status: Draft pre-registration prepared; not ratified; execution parked. No measurements.
Falsified by: The corpus exceeding the kernel's representational envelope beyond the proposed thresholds once ratified — a NO-GO, which would be published as such.

An orthogonal coordination algebra

Motivation: Recurring coordination semantics — undertakings, conformance, non-interference, temporal validity and forfeiture — were being flattened into guards and ledger entries, losing distinctions that matter across domains.
Hypotheses: A small orthogonal operator basis (obligation · conformance · independence · validity window · deadline with scoped consequence) covers these semantics across independent domains without domain- or backend-specific exceptions.
Pre-registered design: Candidate operators nominated from evidence in independent domain families; remove-one tests (an operator that reduces to compositions of the others is dropped — one already was); behavioral witnesses with positive/negative contrasts; a reference evaluator whose expected traces are frozen and reproducible; held-out validation on a genuinely untouched family before any promotion.
Status: A provisional five-operator proposition is frozen with its reference evaluator and traces; its first compiler realization is in progress. Held-out validation has not run.
Falsified by: The held-out family requiring a domain-specific exception; an operator failing remove-one; realization forcing semantics the proposition cannot express.

Ontology-first domain research (documentary-credit pilot)

Motivation: Domain semantics should be derived from primary-source evidence, not intuition — and the language corrected from what domains actually require.
Hypotheses: Statute-grounded behavioral witnesses (each an exact source excerpt, an explicit interpretation, a discriminating behavioral test, and a modeling-pressure hypothesis) are a sound, honest unit for domain-to-language pressure.
Pre-registered design: A bounded witness packet on one domain (documentary letters of credit under one enacted statute), each witness carrying provenance digests and positive/boundary/negative contrasts; explicit epistemic labels (direct vs interpretive; open vs settled); coverage claims limited to what the sources actually support.
Status: The pilot packet is complete and merged after bounded review; it fed the coordination-algebra nomination above. Broader, practice-source coverage is scoped as follow-up work with independent re-sealing.
Falsified by: Independently authored and re-audited witnesses repeatedly failing to preserve source-to-interpretation traceability, or failing to yield stable discriminating outcomes — a single witness failing re-audit invalidates that witness, not the method.

Execution-binding adequacy

Motivation: Binding verified semantics to real execution — libraries, processes, remote APIs, durable jobs — must not smuggle semantic decisions into adapters; whether the current binding contract suffices is an empirical question.
Hypotheses: A small, technology-neutral set of binding properties (identity, authority, routing, parameters, invocation, timing, failure, evidence) either reduces to the existing declarative contract or forces a minimal typed extension — decided by evidence, not preference.
Pre-registered design: Executable specimens across genuinely independent execution classes (verified non-wrapper relationships; distinct failure ownership); frozen current-contract baselines; a staged protocol where candidate extraction is frozen before a sealed, independently-authored substitution case is revealed; admission requires the same property forced by at least two independent classes.
Status: A preparation draft has pinned reproducible candidate baselines; the protocol and result phase remain unratified and blocked on the kernel-realization milestone. No adequacy verdict exists.
Falsified by: (For the extension hypothesis) every candidate property reducing to profile data on the existing contract — a RETAIN verdict, published as such. (For the method) non-reproducible classification, reveal-order leakage, inability to distinguish profile data from semantic extension, or a candidate admitted from a single execution class.

← Back to the overview