wai-quantum 0.6.14: design critique, equivalence, alternatives and classical shortcuts across every surface; contraction trees that sweep along the line #427

Merged

dcharlot merged 1 commit from wai/quantum-0.6.14 into main 2026-10-22 09:00:55 -04:00

Conversation 0 Commits 1 Files changed 53 +3345 -143

2026-10-22

Transaction Science

wai-quantum 0.6.14: design critique, equivalence, alternatives and classical shortcuts across every surface; contraction trees that sweep along the line ...

This carries 0.6.13, which was not published on its own, and 0.6.14.

## 0.6.13

The engine is the microscope and the judge, a local model is a muse that waits to be
asked, and the reader is the author. Every figure is the engine's. Nothing changes the
reader's program until they apply a change.

engine::critique gives a program's measures before it runs:
- gates by kind; one-, two- and three-qubit gates; T gates; gates off the Clifford
  angles;
- depth and two-qubit depth, and whether it is a Clifford circuit;
- a fidelity at reference error rates, named as a model.
It also gives the program's smells, each with its lines and operations: gates that
cancel, rotations that merge or do nothing, idle qubits, and qubits never measured.
critique::variants rewrites the program's own text with the smells fixed, all of them
to a fixed point and each one alone. Each candidate carries its measures and the
equivalence verdict; one the check finds different is never offered.

engine::equiv compares two programs up to a global phase:
- by tableau for Clifford circuits, at any size;
- otherwise on every basis state, to 8 qubits;
- otherwise on random product states, to 22 qubits, which is evidence, not proof;
- otherwise unchecked, with the reason.
qasm::parse_lines gives each operation's line, and draw::layout each gate's operation
index.

engine::shortcuts lists every exact classical route the engine knows, priced on the
program, and names the cheapest for each question:
- the stabilizer tableau;
- the cut-rank sum;
- an exact matrix product state, at a bond bounded by the operator Schmidt ranks of
  the gates crossing each cut. The bound is held by a test against the true Schmidt
  ranks of random circuits;
- the light cone;
- a tensor-network contraction, when asked to be thorough;
- the state vector.
A cheap route proves a design classically easy; none being cheap proves nothing. On
the 70-qubit random graph-state circuit the quick check finds the cut-rank sum (widest
cut 30) in under 2 s. The thorough one finds widest cut 29 in about 4 minutes.

Quantum Studio gains:
- the critique, always on under the diagram, with the smells flagged on it, Show and
  Fix;
- the classical cost, on request;
- Alternatives: Improve (the engine's rewrites and the model's, each checked) and
  Explore (the model's different circuits, each measured and compared), with an atlas
  and Apply;
- Predict first, which sets the reader's expected outcome against the run;
- an assistant that asks questions first and speaks only when asked, offering (never
  sending) a hint when allowed. Its new tools are critique_qasm, check_equivalent and
  classical_shortcuts.
wai-web exports the four engine calls. The desktop app answers them natively.
wai-quantum-cli 0.1.1 adds critique, equivalent (exit 0/1/2), improve and shortcuts,
with four MCP tools.

examples/cutrank_qasm gains probs: the exact 2^n·p of given bit strings, with their
linear cross-entropy.

## 0.6.14

A deep circuit on a line, contracted qubit by qubit along time, keeps a boundary of
only the bonds between two neighbouring qubits: about half the depth for a
controlled-Z brickwork. The contraction search rarely found that tree on its own.

- Network::amplitude_split_placed builds the same network as amplitude_split, index
  for index, with each index's place in the circuit: a qubit and a time.
- tensor_places, sweep_orders and sweep_path turn the places into the trees that sweep
  qubit by qubit, or time by time.
- find_path_seeded and find_path_restarts_seeded score such trees beside the trials and
  polish the best of them all alike. A seed that is not a tree of the network is
  ignored.
- The engine's tensor_network method and the thorough shortcut lens take both sweeps as
  seeds.

On the 70-qubit random graph-state circuit, one amplitude's largest intermediate falls
from 2^48 to 2^36, and its multiply-adds from 2^56.1 to 2^48.7, in the thorough lens.
With 16 trials and no seeds the search finds 2^59. The pinned plans in
tests/bit_identity.rs call find_path itself and are unchanged.

The job service answers four design checks: POST /v1/critique, /v1/equivalent,
/v1/improve and /v1/shortcuts, with the matching MCP tools. They run in its own process
within bounds:
- each counts against the client's submissions;
- at most two run at once, and one that waits ten seconds is refused with a 503;
- equivalence work past 4e9 gate amplitudes, and a thorough search, are refused with a
  422 naming the local command.
The smoke script checks each one. It now follows a job that outlasts /v1/run's wait,
as a client does, instead of failing on a loaded host.

wai-quantum-cli 0.1.2 carries the faster contraction.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

7cef125325

• View at this point in history

2026-10-22 02:40:04 -04:00