Skip to content
FoldEngine Italiano

The register, at a glance.

Every tile is generated by a script from the project's own register, at each publication. Nothing here is typed by hand: a figure without its limits or its act stops the build. Generated from Nexus commit 95d42145148b of 21 October 2026.

V2Part CCCLVIII

Value
495.0ns/day
698.2 ± 1.7 µs per step at 4 fs
Status
Promoted · headline
Limits
  • Conformational preparation not demonstrated: the starting state comes from a preparation that was run, but its conformational equilibrium is not proven.
  • Energy drift certified on CK, not on V2: +2.204e−4 ± 0.162e−4 kT/ns/dof on window A, samples [10:410) after 1.2 ns of preparation, standard error across 8 replicas. V2 reproduces CK bit for bit from s1-e01 for 240 ps (measured); on the certification trajectories the identity is an argument from construction, not a measurement.
  • Across the four states of test C, on samples [10:410), the drift is equivalent within ±1e−4 kT/ns/dof (90% Welch interval). This does not mean the drift is independent of the conformation.
  • The equivalence margin is ±1e−4 kT/ns/dof: a difference smaller than that is not excluded, and is not measured.
  • Without a thermostat the system warms by +0.053 ± 0.008 K per ns.
  • Electrostatics are DSF with a 9 Å cutoff, not PME: this is not the same calculation as the Amber and OpenMM figures published for JAC.
  • One session only: six arms in one night. It is not an uncertainty across sessions.
  • Prediction G1 failed: per marginal watt V2 delivers 0.989 of CK, in a session that was not quiet. In energy per step V2 spends about 352 µJ more than CK, and the host is not excluded as the cause.
  • The promotion does not certify conformational equilibrium, the accuracy of observables, equivalence with other engines, or performance outside the protocol: another starting state, long runs that write samples, another machine.

Promoted by an act of the project owner on 5 October 2026, Part CCCLVIII of the register. Full record: digests and the command that reproduces it

CKPart CCCXXXVII

Value
475.8ns/day
726.4 ± 3.9 µs per step at 4 fs
Status
Promoted · no longer the headline
Limits
  • Conformational preparation not demonstrated: the starting state comes from a preparation that was run, but its conformational equilibrium is not proven.
  • Energy drift certified on CK: +2.204e−4 ± 0.162e−4 kT/ns/dof on window A, samples [10:410) after 1.2 ns of preparation from the input coordinates, standard error across 8 replicas.
  • Across the four states of test C, on samples [10:410), the drift is equivalent within ±1e−4 kT/ns/dof (90% Welch interval). This does not mean the drift is independent of the conformation.
  • The equivalence margin is ±1e−4 kT/ns/dof: a difference smaller than that is not excluded, and is not measured.
  • Without a thermostat the system warms by +0.053 ± 0.008 K per ns.
  • Electrostatics are DSF with a 9 Å cutoff, not PME: this is not the same calculation as the Amber and OpenMM figures published for JAC.
  • The promotion does not certify conformational equilibrium, the accuracy of observables, equivalence with other engines, or performance outside the protocol: another starting state (from state s4 CK is 15 to 18 µs per step slower), long runs that write samples, another machine.

Promoted by an act of the project owner on 30 September 2026, Part CCCXXXVII of the register. Full record: digests and the command that reproduces it

Errata124 entries

The most recent corrections in the Nexus register, newest first. Errata are written in Italian, the working language of the project.

  • Part DXXIRevisione delle cifre (DXVIII–DXX e le verifiche di Nautilus), ed errata, 14 ottobre 2026
  • Part DXXErratum alla DXIX: che cosa vuol dire «motore delle corse di produzione», 14 ottobre 2026
  • Part DXVIIRevisione delle cifre (CDXXXVI–DXVI), ed errata, 12 ottobre 2026
  • Part CDXXXVRevisione delle cifre, sei rilievi: ERRATUM alla CDXXXIII, alla CDXXXIV, alla CDXXVII, alla CDXXIX, alla CDXXX e alla CDXXXII (il rimedio del barostato è provato, la causa esclusiva no; gli S² vengono da 3 blocchi su 4 per replica)
  • Part CDXXVERRATUM alla CDXIX: le periodicità negative di FactorIX.prmtop non sono un difetto di OpenMM da segnalare, ma un comportamento voluto (issue 3391, chiusa nel 2022); e il prmtop di Factor IX è vecchio, le sue energie non sono un riferimento di Amber
  • Part CDXXIIILa revisione delle cifre sulle undici Parti: quattro rilievi, nessuna predizione cambia esito. ERRATUM alla CDXIV, CDXIX, CDXX e CDXXII: un'attribuzione causale e una proprietà del codice sono non valutate, una precisione è ridimensionata, un rimando è corretto

All errata, with the sha256 of each text

Daily rootRFC 6962

Closed on 2026-10-21: 242 leaves.

64714918167154272acd5b797a596a13dbeea91d3e69c6006aff9f11ae4461bb

OpenTimestamps: attesa

How to check it yourself

Three Macs, one resultM4 Pro · M2 · M1

M4 Pro

same bits

The development Mac. Reference for every comparison.

Attestation root (validator of the production path, edition 2): 3bb28046d9fac28ae54cc082ed724c33, registered 3 October 2026

M2

same bits

Reproduces the three digests of the M4 Pro run of the production engine, 1 ns at 4 fs.

Attestation root (validator of the production path, edition 2): 3bb28046d9fac28ae54cc082ed724c33, registered 5 October 2026

M1

same bits

Reproduces the same three digests, on the same build of macOS.

Attestation root (validator of the production path, edition 2): 3bb28046d9fac28ae54cc082ed724c33, registered 6 October 2026

Limits: one machine per chip, macOS 27.2, the builds and starting states that were tested. Not other builds, other states, other units, or the Pro and Max variants. The bits depend on choices the Metal compiler is allowed to make, such as fusing a*b+c into one FMA: identity across chips is observed, not guaranteed by the code, and a new chip is tested before anyone says «the whole family». Source: Nexus 95d42145148b, RISULTATI.md; Nexus 95d42145148b, reference-digests-jac.txt.

In progress

In progress

A twin of Nexus on the CPU

A second program, in C on the CPU, recomputes the forces of the production engine from the Metal specification alone. In 120 of 200 samples the forces match the GPU byte for byte on all 23,558 atoms: the samples taken right after the neighbour list is rebuilt. The other 80 differ by a few fixed-point quanta, because the list there is a few steps older. The trajectory is not done yet: integrator and constraints still have to be written.

Source: Nexus 95d42145148b, SCHEDA.md.

Prototype

A local server for agents

A software agent can ask the Mac for a short molecular dynamics run with Nexus and get the result back with a signed proof. Local only, over stdio: no open port, no network call. Four tools, none of which runs commands or reads files; a long run waits for a person to confirm it. It is a prototype, not part of the app.

Source: Hypoform c4350c67cbc6, LEGGIMI.md.

Internal testing

The app, on TestFlight

Build 202610210924 was uploaded on 21 October 2026 for Mac and for iPhone and iPad, for internal testing only. There is no public beta and the app is not on the App Store yet.

Source: FoldEngine app 5c71ffc3e75d, testflight-build.tsv.