QEC deneyleri ve tur izleri

• Status: proposed
• Revision: 1 (2026-08-09)
• Target contract: 0.1.0-experimental
• Feature flag: experimental.qecExperiments=true
• Owners: language, runtime, verification, product
• Depends on: NM-RFC-0011 feed-forward, Algorithm Pack 0.4 QEC correction artifacts
• Does not change: stochastic @noise, stabilizer/statevector ceilings, hardware claims, or fault-tolerance readiness

1. Summary

N/M already has two separate QEC-adjacent surfaces:

• stochastic, global @noise(model, probability) simulation; and
• exact, algebraically verified single-error correction maps in Algorithm Pack 0.4.

Neither surface defines a structured error schedule, a round trace, decoder identity, or a logical-state invariant result. This RFC adds a closed, explicitly negotiated experiment declaration that evaluates the existing exact correction artifacts without pretending to be a noisy threshold simulation:

qec experiment steane_rounds {
  code: qec.steane7_code@0.4;
  rounds: 3;
  inject round 1 X(q[2]);
}

The declaration produces a deterministic artifact containing every round's injection, syndrome, selected correction, residual class, logical invariant, and decoder provenance. It is a discrete algebraic experiment, not a quantum trajectory or hardware run.

2. Goals

• Bind an experiment to one exact-version, built-in QEC correction artifact.
• Define one to eight compile-time rounds.
• Permit at most one supported single-qubit Pauli injection per round.
• Produce a complete round trace with exact syndrome/correction provenance.
• State the logical invariant only inside the artifact's published single-error model.
• Carry the result through Core, token structural frontend, CLI, Worker, Playground, docs, and the QEC Lab.
• Reject exporters and combinations that cannot preserve the experiment.

3. Non-goals

• Stochastic sampling, threshold estimation, logical error rates, decoder benchmarking, correlated noise, leakage, crosstalk, erasure, or calibration.
• Multiple physical errors in one correction round.
• Repeated syndrome extraction circuits, ancilla reset, measurement latency, or hardware feed-forward.
• Claiming that a correction-map result is a decoded quantum state.
• Extending @noise or allowing @mitigation to reinterpret this artifact.
• Publishing a stable QEC language surface.

4. Closed grammar

qec-experiment = "qec" "experiment" identifier "{" experiment-item* "}" ;
experiment-item = code-item | rounds-item | injection-item ;
code-item       = "code" ":" qec-code-id "@" "0.4" ";" ;
rounds-item     = "rounds" ":" integer ";" ;
injection-item  = "inject" "round" integer pauli "(" "q" "[" integer "]" ")" ";" ;
pauli           = "X" | "Y" | "Z" ;

The closed code identifiers are:

• qec.perfect5_code@0.4
• qec.steane7_code@0.4
• qec.surface_d3_patch@0.4
• qec.repetition_code25@0.4

Every integer is a source literal in revision 1. Runtime expressions, aliases, package ranges, omitted versions, and imported user decoders are invalid.

5. Schedule rules

• rounds is required and must be in 1..8.
• code is required exactly once.
• Each round index is in 0..rounds-1.
• At most one injection may target a round.
• The qubit index must address a data qubit in the selected artifact.
• The 25-qubit repetition-code artifact accepts only X; the other frozen artifacts accept X, Y, and Z.
• A round without an injection has the all-zero syndrome and no correction.
• Experiment declarations cannot coexist with @noise or @mitigation.

The one-error-per-round restriction is semantic, not a UI default. A future multi-error contract requires a new revision and new verification evidence.

6. Normative evaluation

For every injected error label Pq, the compiler looks up the unique verified correction-map entry whose corrects set contains that label. It records:

1. the exact syndrome;
2. the artifact-selected representative correction;
3. that the injected error times the correction is in the stabilizer class, as already proven when the correction artifact was built; and
4. logicalInvariant = holds-within-published-model.

No matrix/statevector execution is inferred. A missing entry is a compiler error, not an uncorrected runtime result. Clean rounds record the zero syndrome and logicalInvariant = holds-clean-round.

7. Typed artifact

The program metadata is closed and contains:

• contract version and RFC id;
• experiment name, source line, exact code/package identity, round count, and error model;
• ordered injection schedule;
• an ordered, gap-free trace for all rounds;
• decoder provenance: correction contract, verification kind, artifact hash, generator rank, coverage, and publication state;
• the logical invariant scope;
• a deterministic experiment identity hash; and
• an explicit limitation string.

The artifact contains no raw source and is recomputed from the built-in correction map before use.

8. Lowering and export

The declaration is removed before ordinary executable statements reach the simulator. Runtime output may display the compiler-produced trace, but may not recompute the correction map.

NM JSON IR may carry the typed experiment metadata. OpenQASM 2, OpenQASM 3, framework exporters, and target-native transpilation must fail closed until a versioned carrier explicitly includes this contract. They must not emit a concrete circuit and silently discard experiment provenance.

9. Diagnostics

Diagnostics point to the declaration item and describe the bounded accepted form. None recommends raising runtime or hardware limits.

10. Product surface

The QEC Lab may show a three-round compiler-produced trace for the selected correction-map row. The trace must visibly distinguish:

• scheduled injection;
• syndrome;
• selected correction;
• residual stabilizer class;
• logical invariant scope; and
• decoder artifact hash.

The Playground deep link includes qecExperiments=1. The UI labels the result as an algebraic artifact and repeats that it is not noisy, multi-round decoder, fault-tolerance, or hardware evidence.

11. Compatibility and rollout

1. Freeze flag-off behavior and negotiation diagnostics.
2. Land the closed parser, artifact builder, validator, and Core/token parity.
3. Add CLI and Worker negotiation plus deterministic output.
4. Add QEC Lab round trace and Playground deep links.
5. Add fail-closed IR/export/transpilation behavior.
6. Close maximum-carrier, fuzz, EN/TR, browser, accessibility, security, and named language/runtime/product review evidence.

The RFC remains proposed and the feature remains experimental until the full evidence matrix is reviewed. Implementing revision 1 is not approval.

12. Acceptance criteria

• Disabled syntax produces NM-QEC-EXP-001 and no executable residue.
• Core and token structural frontends produce identical metadata.
• Every frozen correction artifact accepts all of its published covered errors.
• Trace length equals the declared round count and indices are gap-free.
• Duplicate-round, out-of-range, unsupported-Pauli, invalid-qubit, non-exact package, @noise, and @mitigation cases fail closed.
• Experiment identity is deterministic and changes when schedule/code changes.
• CLI JSON, Worker, Playground share state, QEC Lab, spec catalog, and EN/TR content expose the same contract and limits.
• Unsupported exporters produce NM-QEC-EXP-008 instead of dropping metadata.
• Published text never calls the artifact a threshold, logical error rate, hardware result, or fault-tolerant decoder.