For the last couple of years I've been building, testing, and — more often than I'd like — demolishing pieces of a physics research program I call SHARD. The full corpus now spans nine versions, dozens of papers, and hundreds of machine-checked “receipts”, all public in the isp repository. This post is the guided tour I'd give a curious friend: what the idea is, what we've actually established (with honesty about the grade of each claim), what is provably missing, and how it sits next to the other roads to quantum gravity.
Almost every physical theory starts by assuming a stage: spacetime, fields on it, maybe a Hilbert space. SHARD starts one floor below, with a single kind of object: the record.
A record is an irreversible commitment of information — the moment something definite happens and cannot be undone. Think of a measurement result being registered, a grain of the past crystallizing. In SHARD there is no pre-existing space, no time coordinate, no wavefunction living anywhere. There are only chains of records, each carrying two kinds of bookkeeping:
Everything else — quantum limits, geometry, gravity's form, matter's structure — has to be derived from the self-consistency of that bookkeeping, or honestly declared an import. The discipline of the program is that nothing gets smuggled in: every claim is tagged as a theorem, a demonstration, a measurement, or an explicit import, and the ledger of imports is public.
If records are the atoms of reality, what law governs when a record fires? Here the program got lucky early: the law isn't postulated, it's forced.
Ask a simple question: if a chain survives without committing through evidence amount I, and then through a further amount J, self-consistency of the bookkeeping demands that survival compose:
There is exactly one decreasing function that does this: the exponential,
And the rate λ? A second self-consistency requirement — the theory must correctly account for its own record-keeping (what we call self-accounting: the evidence spent must equal the evidence logged, −log S = I) — pins it to exactly λ = 1. No dial. The “click law” — when records fire — is the theory's one dynamical law, and it comes with no adjustable constant.
That absence of dials is SHARD's personality. Where the law leaves freedom, we prove theorems about the freedom instead of choosing values — and some of those no-go theorems turned out to be the most interesting results in the program.
Take two record chains and ask what correlations their commitments can carry. From the records’ own consistency (given one disclosed structural premise about how record observables compose), you can derive the fence around all possible correlations — and it lands exactly on the boundary known from quantum foundations: the Tsirelson bound,
together with no-signalling, the exclusion of stronger-than-quantum “PR boxes”, and the finer “almost-quantum” envelope. The records, knowing nothing about Hilbert spaces, rediscover quantum mechanics’ outer wall.
But here's the twist, and it's a theorem, not a shrug: the records provably cannot select the exact quantum point inside that fence. Two independent blindness mechanisms make the entangling strength a free input. The one datum that would close the gap — essentially the choice between complex and real composition of systems, the “tensor-product bit” — lies precisely in the layer the records cannot see. Delightfully, that same bit is currently contested in the experimental literature too: the world's physicists are having trouble fixing from the outside the exact bit our records can't fix from the inside.
This year the program sharpened that freedom to something almost embarrassing in its smallness. We asked: if you assume nothing beyond what the records derive — a maximum-honesty (maximum-entropy) principle under only the derived constraints — how much of the entanglement coupling gets determined?
Answer, verified to eighty decimal places: everything except one bit. The magnitudes of all the correlations, their equal spread, their scaling — all forced. The single thing honesty cannot decide is one sign: whether the coupling's pattern is the Bell-violating one or the Bell-satisfying one. The principle is provably blind, at every order, to exactly the question the Bell experiments answered (reality chose violating — that's the 2022 Nobel). So the theory falls silent precisely where the lab speaks. Our next construction — a two-species matter sector — must derive that one sign. If it derives the wrong one, it dies. That's the healthiest situation a theory can be in: one bit from falsification.
The deepest question for any “it-from-information” program is whether space and time actually emerge, or whether you quietly assumed them. SHARD's answer has two chapters — a failure that became a discovery, and a success that surprised us.
The failure: the action does not build. SHARD derives a cost function (an action) that grades candidate universes beautifully: genuine spacetime-like record webs are provably cheap, forgeries pay penalties that grow with the square of their size. The natural hope — physics has a long romance with least-action principles — was to grow universes by always taking the cheapest next step. Three campaigns and five variants later, that hope is dead as a theorem-grade result: step-wise cost minimization gets squeezed in a pincer (any finished-universe target rewards over-builders; neutralize it and the per-step fees reward do-nothings), and no admissible re-pricing escapes — we proved a ceiling over the whole family. Lesson, now carved into the corpus: the action is a judge, not a foreman.
The success: the churn builds. The process that does work is almost comically humble: chains grow, accumulate content, and occasionally die and restart (“churn”). No optimization, no foresight. And this recycling process, run from the click law with nothing tuned:
Combined with a sufficiency theorem proved earlier in the corpus (a web is a faithful piece of 2D spacetime if and only if it is 2D and its discrepancy vanishes), these pieces chain into what we call the discharge: growth ⟹ spacetime, as a theorem. Two of its three pillars are done; the third (upgrading the statistical layer from measured to proved, via a Poisson-approximation argument) is in progress as I write.
Each record web is 2D — one time direction, one spatial. Doesn't that kill the program? No, and the reason is the most modern idea in the stack: the other dimensions are made of entanglement. Each 2D sheet is a thread; the couplings between threads — who is entangled with whom, how strongly — define nearness, and the pattern of nearness is transverse geometry (this is the record-native version of the “entanglement builds space” idea associated with Van Raamsdonk and ER=EPR). We've demonstrated the reconstruction end-to-end at pilot scale: from pure record data across many sheets, the machinery rebuilds transverse geometry, and in the 3+1 configuration the reconstructed “sky” comes out spherical to under 1%. What we have not done — and say so loudly — is derive why three transverse-plus-time rather than some other number. Dimension selection is an open problem, not a result.
Gravity. Following a thermodynamic route (à la Jacobson), the form of Einstein's equations emerges from record bookkeeping — conditionally, with the conditions documented. But Newton's constant G does not emerge, and this is not a failure of effort: it's a proved no-go. Records carry relative structure; one absolute scale (equivalently, the fundamental step length) provably cannot be manufactured from inside. SHARD predicts the shape of gravity and honestly imports its strength.
Matter. The record substrate turns out to be surprisingly good at fermions. It evades the classic Nielsen–Ninomiya obstruction to putting chiral (handed) fermions on a lattice — record-natively, with an exact index theorem — and supports a mass-generation mechanism without a Higgs field (dynamical chiral symmetry breaking in an interacting flow), demonstrated at toy scale. There's also a closed-form “gap law” on the record ledger, proved as a global-optimality theorem for every size — a piece of pure math that grew its own life: the key inequality is now machine-formalized in Lean 4 in its own repository, walsh-delta. What matter does not deliver: no particle masses or mass ratios are predicted — which gap belongs to which particle is an import (the “mode” assignment), stated as such on every page that touches it.
Experiments. The program maintains falsifiers you could actually run: the sharpest is a universal Gaussian onset prediction for gravitationally-induced decoherence (mechanism-blind to all orders — if the onset is measured to be non-Gaussian, a load-bearing piece dies), alongside the one-sign prediction above.
A theory should display its debts as prominently as its assets. SHARD's, today:
Causal sets (Bombelli–Lee–Meyer–Sorkin; Rideout–Sorkin growth): the closest cousin — discrete partial orders, “order + number = geometry”. Differences: SHARD's order isn't postulated, it's the dominance structure of derived record bookkeeping; the growth law isn't axiomatized (no sequential-growth axioms — the click law is forced by self-accounting); and the entropic-junk catastrophe that haunts counting measures over causal sets is, at least at exactly-computable sizes, beaten by the derived builder rather than assumed away.
Collapse models and flash ontologies (GRW/CSL; Bell's “flashes”, Tumulka's relativistic models): records are kin to flashes — pointlike actual events as the furniture of the world. But collapse models introduce new constants (collapse rate, localization width); SHARD's click rate is fixed at 1 by self-accounting, and the program derives limits from the event structure rather than adding it to quantum mechanics.
The stochastic-quantum correspondence (Barandes): SHARD's moves are defined seal-to-seal — the indivisible grain — and the program has receipts on which layers of the stochasticity are irreducible (the sequencing randomness provably cannot be removed; it can only be relocated between channels). SHARD reads this correspondence as load-bearing physics and builds geometry on top of it.
Wheeler's “it from bit”: SHARD is a concrete, executable descendant — with a twist Wheeler might have enjoyed: we can now prove that some its don't come from bits. The scale, the mode, one sign: it from bit, except for the bits that only an “it” can supply.
CDT, loop quantum gravity, strings: causal dynamical triangulations also grows spacetime from causal building blocks, but postulates the simplices and the action — and SHARD's sharpest growth result is that its own derived action cannot be the builder, which we'd gently offer as a caution to any action-driven growth program. Loop quantum gravity quantizes geometry; SHARD never has a geometry to quantize. String theory adds structure to unify; SHARD subtracts, to see how much of physics survives on bookkeeping alone.
I want to end on the part of the program I'd defend even if the physics dies. Every claim in the corpus is built like an experiment:
Whether or not records are what the world is made of, I'm convinced this is what theoretical physics done in public with machine assistance should look like.
v8/LEDGER.md — the corrections ledger, which is the program's conscience.The edge of the program right now is exactly one bit wide. I find that a very exciting place to stand.