Bounded crosstalk noise

• Status: proposed
• Revision: 1 (2026-08-11)
• Target contract: 0.1.0-experimental
• Feature flag: experimental.crosstalkNoise=true
• Owners: compiler, runtime, verification, tooling, product, security
• Depends on: approved NM-RFC-0018 and NM-RFC-0019 timed execution plans
• Does not change: independent noise, target topology, leakage, pulse, or hardware contracts
• Implementation gate: section 15 approval is required before implementation

1. Summary

Independent per-gate and per-idle channels cannot represent an error on one scheduled operation that depends on a simultaneous neighboring operation. This RFC proposes a small, explicit crosstalk profile resolved against a frozen timed plan and executed by the bounded density backend.

The feature is a local sensitivity model. A supplied coupling number is not evidence that a physical device has that coupling.

2. Decision set

1. Crosstalk is defined over pairs of overlapping scheduled operations, never inferred from source adjacency or wall-clock execution.
2. Profiles are named exact-version package exports with immutable digests; inline matrices, URLs, callbacks, and runtime discovery are forbidden.
3. Revision 1 supports only a closed coherent ZZ phase interaction on two distinct qubits in the approved five-qubit density backend.
4. Candidate matches are computed and bounded before matrix allocation.
5. Independent gate/idle channels run first; crosstalk interactions then run in a frozen order at each overlap boundary.
6. No profile is selected from calibration unless a later revision defines an explicit mapping. NM-RFC-0019 snapshots do not imply crosstalk data.

3. Profile artifact

An exact-version package export resolves to nm-crosstalk-profile@0.1:

format: "nm-crosstalk-profile"
version: "0.1"
identity: package specifier, export, package digest, profile id
rules: ordered list of
  aggressor native gate
  victim native gate
  ordered aggressor/victim qubit selectors
  required target edge relation
  zzRadiansPerNanosecond
limitations: local sensitivity model, not calibration bound
integrity: sha256 over canonical artifact

There are at most 32 rules. Every rate is finite, non-zero, and has absolute value at most 0.01 radians/ns. Selectors name concrete target qubits or one closed adjacent relation. Wildcards over gate names, regex, functions, inheritance, and unknown fields are rejected.

4. Match and ordering semantics

Two native operations match when their half-open schedule intervals overlap, their gates/selectors match one rule, their qubits are distinct, and the required target relation exists. The overlap duration is:

max(0, min(endA, endB) - max(startA, startB))

Each unique (rule, aggressorOperation, victimOperation) matches at most once. Self matches and reversed duplicates fail profile validation unless represented by separate explicit rules.

Applications are ordered by (overlapEnd, overlapStart, ruleIndex, aggressorOperationId, victimOperationId). At overlapEnd, after independent gate/idle channels ending at that timestamp, apply:

U = exp(-i * theta/2 * Z tensor Z)
theta = rate * overlapDuration

The exact matrix and qubit order are embedded in the carrier. Global phase is retained because later controlled or comparative contexts may make it relative.

5. Bounds

• density qubits: 2..5;
• native operations: existing NM-RFC-0019 maximum 256;
• profile rules: 1..32;
• candidate operation pairs inspected: at most 4096;
• accepted crosstalk applications: at most 128;
• absolute accumulated angle per qubit pair: at most 4 * PI;
• artifact: at most 65,536 UTF-8 bytes; and
• result extension: at most 131,072 UTF-8 bytes.

All pair counts are checked with bounded sweep-line indexing. A quadratic unbounded scan is non-conforming.

6. Carrier and evidence

The timed density carrier embeds the validated profile, exact package identity, all matches, overlap durations, matrices, operation bindings, numeric policy, and integrity. Public runtime recomputes matches from the timed plan and rejects any mismatch before density allocation.

Result evidence records each application, rule/package provenance, operations, qubits, overlap, angle, ordering, and numeric drift. It states:

• bounded local sensitivity simulation;
• profile supplied by the user/package;
• not derived from live calibration;
• no leakage, pulse distortion, non-Markovian, mitigation, hardware-fidelity, or fault-tolerance claim.

7. Diagnostics

8. Product surface

After approval, the Noise Lab may display an opt-in crosstalk lane only when a valid timed plan and exact profile are selected. It must show a schedule overlap graph, ordered matched pairs, angle/provenance, and a comparison against the identical plan with crosstalk disabled. Color is never the sole indicator.

Experiment Studio may compare the two local models but cannot label their difference as predicted device error. There is no automatic profile selection.

9. Export and compatibility

Circuit-only exports cannot preserve schedule overlap or package provenance and therefore fail closed in strict mode. A versioned sidecar may carry the complete timed/crosstalk carrier. Flag-off execution and independent noise remain unchanged.

10. Security

• Exact package artifacts are data only; they cannot execute code.
• Validate byte/count/rate limits before pair generation.
• Reject sparse, inherited, accessor-backed, unknown, or mutable inputs.
• SHA-256 provides integrity, not authorization or scientific credibility.
• Telemetry excludes raw profiles and source.

11. Non-goals

• More than two-body interactions, arbitrary Hamiltonians, continuously driven evolution, pulse envelopes, leakage levels, readout crosstalk, or memory.
• Inferring rates from a calibration snapshot or fitting experimental data.
• Stabilizer/statevector execution, real QPU submission, mitigation, or device fidelity prediction.

12. Rollout

1. Approve dependencies and review record.
2. Freeze profile schema/reader, sweep-line matching, and numeric fixtures.
3. Produce compiler carrier and fail-closed public-runtime validation.
4. Integrate the closed ZZ operation with density execution.
5. Add CLI/tool/product evidence surfaces.
6. Close fuzz, cross-platform conformance, maximum-pair performance, browser, accessibility, and claims tests.

13. Acceptance criteria

• No timed plan means no crosstalk execution.
• Hand-computed two-operation fixtures match the frozen ZZ matrix and order.
• Reordering or resizing an overlap changes identity and execution.
• Unsupported, oversized, forged, or mutated profiles fail before allocation.
• Same carrier is byte-deterministic across pinned Windows/Linux Node 22.
• Every surface states that the profile is a local sensitivity model.

14. Relationship to later work

NM-RFC-0029 defines leakage with a different local dimension and backend. NM-RFC-0030 defines pulse artifacts. Neither is implied by this qubit-only scheduled interaction model.

15. Required review record

This proposal creates no flag, parser syntax, runtime, capability, or UI lane. A merge or passing CI is not implementation approval.