Papers on AI

Relocating Trust to the Verifier: A Truth-Maintenance System for an AI-Generated Theory of Everything

David Elliman, Neuro-Symbolic Ltd · June 2026

The runnable companion to the methodology that produced the canon above. A fluent, approval-seeking LLM pointed at an open-ended theory-of-everything by a researcher who wants it to succeed forms a mutual-confirmation loop with no reality-check on either side — the maximally dangerous case, because the mathematics can always be made internally consistent and the one decisive check (experiment) is decades away or absent. The cure is structural rather than a better model: relocate trust from the generator to a verifier. PTMS is a truth-maintenance system that enforces consistency, not truth — it inverts the classical TMS assumption that justifications are ground truth, treating every AI-supplied justification as untrusted until it binds to checkable evidence: a cited script that exits 0, a retired claim's signature absent at every surviving site, a cross-reference that resolves. Three layers of increasing cost — check (read-only consistency over the canon), verify (re-executes the cited self-asserting scripts), and graph (resolves the cross-reference dependency graph) — run over a machine-readable sidecar registry that never edits the human-facing ANCHOR.md / DRIFT.md prose. Reports its behaviour on a real six-month AI-assisted physics corpus, mechanically catching named failure classes: un-propagated multi-part retractions, evidence regressions, seductive numerical coincidences, and live claims standing on retracted foundations. Explicit about the limit — the tool makes such research auditable and constrained, not correct; only forward prediction closes the remaining gap, and no apparatus can manufacture it.

From Counts to Observables: A Measurement Discipline for Discrete Record-Based Physics

David Elliman, Neuro-Symbolic Ltd · July 2026

The methodological companion to the truth-maintenance tool above: a measurement discipline that separates what a discrete model can count from what an experiment can measure. Record-based physics naturally throws off integer counts — states, channels, syndromes, loops, contacts — and their nearby coincidences are exactly where numerology creeps in. The rule: a count becomes an observable only after a response map proves that a real apparatus reads it, expressed through a closed-record response functional Z[J₊, J₋] whose derivatives are the measured currents, residues, fluxes, cross-sections, spectra, pole masses, and likelihoods. Its purpose is to make the programme easier to refute, not to shield it — a count with no response map is demoted, and a response calculation that disagrees with experiment kills the branch. The discipline now has an operator core proved on explicit finite models: recorded insertions carry exactly zero retarded (commutator) component, recorded coincidences factorize at diagonal Born weights, and nothing applied after a record is written can change the recorded count, while the same operation shifts the response kernel at order one. The sharp consequence — counters are dressing-blind, only response kernels acquire radiative dressing — is exactly why the bare α₀ = 1/137 is a count while the dressed 137.036 is a kernel, and it arms a counter/kernel split for Newton's constant, black-hole Hawking flux, the CMB/halo sector, strong-sector Wilson loops, and the electroweak pole-mass ledger. The conclusion is methodological: discrete record-based physics becomes scientific only when it states not just what the substrate can count, but which response a real experiment measures.

Adversarial Self-Registration: A Working Protocol for Keeping a Machine-Assisted Theory Programme Honest

The working companion to the truth-maintenance paper above, documenting the discipline the programme's registered predictions actually run on. Large language models make theoretical claims cheap: a machine-assisted physics programme can generate plausible derivations, matching coefficients, and post-hoc rationalisations faster than any referee can check them, so the classical failure modes of self-deception — forking paths, silently adjusted acceptance criteria, tautologies presented as successes, quietly superseded claims — now operate at machine speed. The countermeasure is a protocol of adversarial self-registration whose seven disciplines are mechanical rather than aspirational: acceptance bars committed to version control before data are touched; every quantitative claim living in a self-asserting gate whose draft expectations, when killed by the computation, stay in the code as a visible fossil; superseded claims corrected loudly, never overwritten; a linter that blocks any push contradicting the claim ledger; analysis conventions amendable only before unblinding; registered bars never adjusted after data; and predictions frozen in public, timestamped, machine-readable form. The evidence offered is not a virtue claim but a receipt set: six public registrations with frozen kill rules published in eight days; one dark-energy registration already killed internally under its own rules and left frozen rather than refitted; a registered acceptance bar missed by 0.005 dex on a single point and honoured — the paper shipped with its verdict section vacant; a convention amendment committed while the deciding scan was still running; a sign-convention bug found by a designed eight-variant control matrix, with every contaminated result voided by name, including an apparent 2.6σ "detection"; and a failure museum of the programme's own withdrawn mechanisms. The protocol is situated in the blind-analysis and preregistration traditions, its threats to validity stated — including the card-stuffing attack that cheap registration invites — and a minimal kit extracted that any machine-assisted research effort can steal. It certifies honesty, not correctness: the note is timestamped while most registered discriminators remain unresolved and after the first internal kill, so its claims about how verdicts will be faced are themselves auditable.

Records Say What Can Be Known: Empirical access, response functionals, and severity in finite information physics

The epistemology companion to the measurement-discipline papers: what a finite information model is allowed to claim about knowledge. Such models naturally invite a confusion — they specify a substrate ledger (admissible records, repair operations, forbidden states, internal counts), but experiments never read an arbitrary internal ledger entry; they read calibrated response channels: currents, residues, poles, susceptibilities, cross-sections, spectra, fluxes, and likelihoods. The resulting principle — records say what can be known; responses say what can be measured — is deliberately not a new interpretation of quantum mechanics competing on the same axis as QBism, relational quantum mechanics, or Everett: it is a constraint on empirical access in any finite, record-bearing substrate theory. A record is a stable, copyable, dynamically protected fact; a response is the closed-time-path or equivalent operational map by which an instrument couples to such facts; and internal ledger structure may be real in a model while remaining unknowable in principle unless it is written into stable records and exposed by a response. The paper formalises the distinction with a record algebra, a response functional, and an empirical-access equivalence relation; compares the view with QBism, relational quantum mechanics, Everett, hidden-variable realism, quantum Darwinism, and engineering calibration practice; and proposes a Mayo-style severity criterion for finite information physics — a numerical claim is credible only when the record object, response map, calibration constants, inherited readouts, and possible falsifiers are stated before the test. Recent applications in the programme serve as case studies: the QED α(0) boundary and Thomson readout, black-hole flux, CMB/halo likelihoods, and electroweak pole matching. The intended contribution is methodological: the record–response split does not prove a substrate true, but it sharply limits how such a substrate may make empirical claims.