Hypothesis record · Draft protocol, not yet registered
Can quantum assistance halve the time to a verified iron-sulfur result?
One molecule, one fixed Hamiltonian and one test, to be registered before any result that counts; results already public stay outside it. Three charts show what has been posted on this problem, how soon classical work has followed hardware results, and how the draft rule would grade the claim under every possible outcome on 31 December 2027; a fourth shows the dated ledger records like this one run on.
2026-10-31 · Amendment 1, before registration: binding is per team, and every quantum dossier includes an ablation. Read the record’s history
- System
- [Fe₄S₄(SCH₃)₄]²⁻CAS(54, 36), one pinned file
- Pass line
- a verified variational upper bound; provisional until registration
- The test
- Speed score ≥ 2the slowest of ten quantum runs is at least twice as fast as the fastest classical run that reaches the line, with timing uncertainty counted against the claim
- The budget
- US$100,000 and 168 hoursceiling per run, at public prices; donated access valued the same way
- The clock
- 2027-07-04 → 2027-12-31quantum dossier public by, then judged once
The problem
54 electrons in 36 orbitals
The singlet (S = 0) state of an iron-sulfur cluster whose four iron centres are open-shell. The benchmark fixes its electronic problem in one public file, and its target is a variational upper bound, not the exact energy.
[Fe₄S₄(SCH₃)₄]²⁻ · schematic, not the pinned geometry, hydrogens omitted
Ms = 0 determinants, before spatial symmetry
8,862,957,169,158,400
C(36, 27)²: every way to place 27 spin-up and 27 spin-down electrons in 36 orbitals. The ground state is a weighted mix of a great many of them.
Singlet (S = 0) states among them
1,171,176,483,067,360
C(36, 27)² − C(36, 28)·C(36, 26). The target is a singlet, so the space a spin-adapted solver works in is about 7.6 times smaller. No solver enumerates either space: DMRG, selected configuration interaction and SQD each restrict the search differently. The size of the space is not evidence that a quantum computer is faster.
Largest subspace diagonalized on this model
1.0 × 10¹⁰
determinants, about one in 890,000 of the Ms = 0 space: the SQD run behind the tracker’s quantum entry (Shirakawa et al., 2025).
Twice a second, one configuration drawn at random: a lit cell is an occupied orbital, spin-up above, spin-down below.
Schematic illustration. Idealized cubane geometry with methyl groups drawn as carbon only; not the benchmark’s coordinates. The count is the number of determinants with 27 spin-up and 27 spin-down electrons in 36 orbitals, C(36, 27)². Random configurations are drawn for illustration and are not a computed wavefunction.
Where the field stands
Posted energies on one Hamiltonian
Posted energies for this same Fe₄S₄ problem, measured above the reference: the lowest published variational bound on the pinned file. The 2017 model files list a DMRG value 7.4 mEh above it, and the tracker files this instance as a baseline benchmark with no advantage candidate. Lower is a tighter upper bound. The runs did not share a budget, and none is an attempt under this test. Choose Pass line to see the line up close.
Posted energies, CAS(54, 36)
Posted 2025-11-11 · quantum hardware
−326.913 Eh
IBM and RIKEN, sample-based quantum diagonalization on 77 qubits, first public as a preprint on 2025-10-31 (Shirakawa et al.), 278 mEh lower than the team’s 2024 result. As posted: 10 hours of quantum processor time and 10 hours on 76,032 Fugaku nodes; the submission and its preprint describe the two phases as overlapping. At IBM’s public pay-as-you-go rate, which starts at US$96 a minute, the quantum time alone would cost US$57,600 or more, over half this test’s US$100,000 ceiling per run, before any Fugaku time.
Posted 2025-11-20 · classical
−327.175 Eh
TrimCI (University of Wisconsin–Madison), a random-determinant method. As posted: the SQD energy in 15 minutes on a 20-core Apple M1 Ultra, and this one, 262 mEh lower, in 14 hours on 1,536 cores.
Posted 2026-01-08 · classical reference
−327.247036 Eh
Spin-adapted DMRG with entanglement-minimized orbitals, Zhai et al., on the pinned file: the lowest published variational bound on it, and zero on this chart. Later, GPU-accelerated DMRG (Legeza et al., 2026-03-30) reached −327.24466 Eh, 2.38 mEh above it, in about 12.6 hours on one DGX B200 node.
Pass line
The reference plus 1.6 mEh (0.0016 Eh, about 1.004 kcal/mol). Any verified energy at or below it passes. Violema chose this absolute tolerance on this file; it is not a claim of experimental accuracy.
As posted, not rerun here. The tracker stores the same Hamiltonian in a different orbital basis, which leaves the energies unchanged, so its entries compare directly; Zhai et al. name the pinned file. Legeza et al. do not name their integral file: that they used the same Hamiltonian is inferred from agreement with results on the pinned file. 1 mEh is a thousandth of a hartree. Marks are not joined: no line implies what was posted between them. Runtimes are as each group reported them, not rerun; none follows this test’s ten-run protocol or budget accounting. The reference is set by rule: the lowest published, non-extrapolated variational upper bound for the S = 0 state on the pinned file, frozen when the test is registered. Lower bounds published after that enter the comparison and do not move the target. It stays provisional until registration confirms the spin sector and the file’s hash. Later entries may exist; this is the record as checked on 2026-10-24. The test itself is about time to reach the pass line.
How fast the field moves
Ten years, in the order results became public
Select any mark to read it. Two arcs show how soon classical work was posted after a hardware result: once for a spin model in 2023, once for this Hamiltonian in 2025. The 180-day window exists so a later classical answer still counts.
Evidence time, 2016 to 2026
Shirakawa et al., IBM and RIKEN, 21 authors
Closed-loop sample-based quantum diagonalization on IBM Heron processors with up to 152,064 nodes of the Fugaku supercomputer. For this [4Fe-4S] (54e, 36o) model it used 77 qubits and a 10¹⁰-determinant subspace, reaching −326.912 hartree, 278 mEh lower than the team’s 2024 result. The authors describe its accuracy as comparable to some all-classical approximation methods.
Dates are first public versions (preprint or announcement), not journal years. The cost estimates measure different things under different hardware assumptions, so they are placed in time, not plotted against each other. The axis changes scale at 2023. Each mark is an observation; in a full map it would be a claim with its own tier.
How the future looks
The rule for every outcome is written now
This is not a forecast. A tier is the record’s grade for the claim: it starts at Speculation and moves only on results under this test. Choose what is true at the deadline, and the draft shows the tier its rule would assign. It becomes a commitment only when the full text has a public timestamp.
Draft outcome rule, to be applied once on 2027-12-31
On 2027-12-31 the draft rule gives
Stays Speculation; the positive result is shown on the record with its confirmation count
Positive, fewer than two independent positive confirmations, no decisive contradictory reproduction of it. Confirmations shown: 1.
- Positive, at least two independent positive confirmations, no decisive contradictory reproduction of it.Established
- Positive, and at least one decisive contradictory reproduction exists.Active debate
- Positive, fewer than two independent positive confirmations, no decisive contradictory reproduction of it.Speculation
- Every bound dossier fails: its speed score’s upper bound is below 2.0, or it misses the ten-run rule.Speculation
- No bound dossier is positive, and at least one is eligible but inconclusive.Speculation
- No dossier from any team was complete, eligible and public by 2027-07-04: no result, never a negative one.Speculation
Draft protocol, not yet registered. It becomes a commitment when its full text receives a public timestamp; results before that cannot count. The energy margin, budget, run count, speed margin, 180-day window, deadline, independence rules and tier moves are Violema’s policy choices, not field consensus.
The ledger behind it
The same engine already holds a real map’s history
Records like this one sit on a dated event ledger. Here it is on PsychDoc’s psilocybin evidence map: every claim and relation, when it entered, and when its meaning changed.
PsychDoc map events
- 88 events
- 10 revisions
- 36 added
- 48 restated
- 4 sources changed
86 events were reconstructed from the map’s git history; the 2 on 24 October were recorded at the time, for a correction found when two AI reviewers from other providers checked the map’s claims against the papers. The engine’s freeze rules stop a published entry from changing without an event that says so.
In a release being prepared; not yet on psychdoc.ai. Counts are from the map’s exported ledger on the release branch.
Not a forecast, and no probabilities.
The outcomes of this test map to tiers written in advance, the way a funder’s prize criteria already work. The historical marks are observations, not tiered claims. Facts are from the hypothesis record, checked against primary sources on 2026-10-24; no domain expert has reviewed it yet.
Cited work includes IBM, RIKEN, Algorithmiq and Cleveland Clinic. IBM and Algorithmiq also sell hardware and software used in that work. That is context, not grounds to exclude anything: the same eligibility and comparison rules apply to every team, and naming an institution is not a claim that anyone there has joined this test. Violema builds evidence records and has a commercial interest in this one being read.