Keeping machine-assisted work
honest
The AI framework we use to maintain consistency in a long-running machine-assisted research programme — truth maintenance against sycophancy, adversarial self-registration against post-hoc rationalisation, and a record-ledger discipline against numerology. The general architecture is now extracted as ClaimLedger: generators propose, a verifier checks admissibility, and only a human commit path applies. Most of these papers also appear in the physics list, because the physics canon is the corpus they are proven on.
New · August 2026 — freezing the corpus before representing it
A Physics Canon as a Replayable Algebraic Object: Verifier-Governed Extraction, Representation, and Minimality Classification
David Elliman, Neuro-Symbolic Ltd · August 2026
The methodology paper behind Algebra Before Bit. A long machine-assisted programme is a corpus, not an answer: definitions move, claims gain and lose support, and a correction in one place can leave an older assertion alive elsewhere. This paper freezes the corpus into a byte-pinned algebraic object before choosing any language to represent it, fixing source inventory, lifecycle grading, equivalence, proof grades and admissible negative landings in advance, so a preregistered test cannot be reverse-engineered. A provenance-tagged classical-kernel construction passes every preregistered test, and the classification then returns a negative: the record and phase skeleton is irreducible and rigid, but an exact proper quotient defeats minimality.
July 2026 — the general architecture
ClaimLedger: Verifier-Governed Claim Maintenance for AI-Assisted Research
David Elliman, Neuro-Symbolic Ltd · July 2026
The general architecture distilled from the truth-maintenance work. ClaimLedger is the reference implementation of Verifier-Governed Claim Maintenance: a generative agent may propose claim-state transitions but cannot promote confidence, erase failures or mutate the authoritative record; a deterministic, evidence-bound verifier checks admissibility, and only a human-controlled commit path applies a transition. The ledger carries claim identities, dependency edges, lifecycle state and an append-only history, under six executable invariants. Eight seeded fault operators are all caught by their assigned verifier layers. It enforces consistency, not truth: a reviewable claim-state machine that stays useful when the generator is unreliable.
Methodology — the runnable tool
Relocating Trust to the Verifier: A Truth-Maintenance System for an AI-Generated Theory of Everything
David Elliman, Neuro-Symbolic Ltd · June 2026
PTMS is a truth-maintenance tool for tracking claims, dependencies, evidence and retractions in an AI-assisted research corpus. Its checks can expose inconsistent references, unsupported status changes and claims relying on retired premises. The reported corpus sizes and check counts describe the paper’s historical snapshot. Passing these checks establishes neither the truth of a physical interpretation nor experimental support for the programme; scientific validation remains a separate task.
Methodology — the measurement rule
From Counts to Observables: A Measurement Discipline for Discrete Record-Based Physics
David Elliman, Neuro-Symbolic Ltd · July 2026
This paper separates a count in a finite model from an observable measured by a calibrated apparatus. Specified instrument models illustrate how a record count can remain invariant while a response changes. A physical application must supply the carrier, interaction, response map and calibration. The fine-structure and gravity examples do not establish those maps: the former numerical agreement depends on a geometry no longer identified with the native substrate, and the gravity identification remains conditional. The contribution is the measurement discipline and its finite examples.
Methodology
Adversarial Self-Registration: A Working Protocol for Keeping a Machine-Assisted Theory Programme Honest
A protocol for recording acceptance criteria, evidence dependencies and failed predictions before interpretation can shift. The paper documents dated examples from the programme. Its procedures make changes and failures inspectable; they cannot certify honesty, correctness or a successful physical theory. The registrations remain historical evidence, while the research branches are frozen.
Methodology
Records Say What Can Be Known: Empirical access, response functionals, and severity in finite information physics
An account of the distinction between a model’s record ledger and empirical access through calibrated response channels. It argues for explicit measurement maps, calibration and tests capable of failing. The fine-structure numerical example belongs to a retired physical identification and does not validate the principle. The contribution is a methodological proposal about assessing finite information models, not a derivation of physical observables or a replacement for quantum theory.