feat: noise-native measurement layer + chassis policy seams (QILC port, public substrate)#7
Merged
Merged
Conversation
Diagonal fast path for the whitened-cost statistics from the QILC noise-native spec: noise_floor = tr(WΣWᵀ), debiased_cost J̃ = Ĵ − floor, cost_std via Var[Ĵ] = 2tr((WΣWᵀ)²) + 4rᵀWᵀWΣWᵀWr (plug-in r), and the difference scale σ_Δ = hypot(σ_A, σ_B) for acceptance/stall comparisons. Also adds test/run_filtered.jl, a local helper for name-filtered testitem runs (this TestItemRunner version has no run_tests function). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…se models
W = W_task·Σ^{-1/2} assembled per element from the model's noise types
(ShotNoise → plug-in covariance_fn(y,n) floored at var_floor/n;
KnownCovariance → diag(Σ), off-diagonals dropped by the diagonal fast
path; Deterministic → σ²=0, identity scaling). Returns (w, σ2) flattened
in model element order; Σ is returned because acceptance statistics need
tr(WΣWᵀ), not just W.
Deviation from plan: default var_floor is 0.05, not 0.25 — the plan's own
acceptance test requires the y=0.9 wigner element (σ²=(1−y²)/n=0.19/n) to
be UNfloored at defaults, which contradicts a 0.25 default (0.25/n floor).
0.05 engages only for |y| ≳ 0.975, matching the stated intent (cap weights
as |y|→1); callers can raise it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…E parity reconstruction Ports the stanford-bosonics ILC measurement observables into the package: - displaced_parity/displaced_parity_at gain a composite transmon⊗cavity variant via SINGLE-ENTRY-POINT kwarg routing: Julia does not dispatch on keyword arguments, so one method accepts n_max (cavity-only, exact pre-existing behavior) XOR (Nq, Nc) (composite kron(I_q, D†ΠD)) and raises ArgumentError unless exactly one is given. - qubit_sigma_z(_at): ⟨kron(σz_eff, I_c)⟩ with σz_eff = diag(+1, −1, −1, …), the leakage-defense convention (−1 for ALL excited transmon levels). - density_encoding.jl: rho_triangle canonical order, rho_to_measvec (type generic, ForwardDiff-safe) / measvec_to_rho, reduced_cavity_rho, and rho_measurement_functions closures for reconstruction-mode campaigns. - parity_reconstruction.jl: reconstruct_rho_from_parity, an emulation of the lab's MLE cavity-state reconstruction. DIAGNOSTICS-ONLY contract: the PSD/trace projection is a nonlinear state-dependent bias the linear ILC Jacobian cannot track, so it must never be a feedback observable. - test/run_filtered.jl: TEST_FILTER-driven filtered @run_package_tests entry. Coherent-state anchor note: the package/demo convention op = D†(α) Π D(α) displaces the state by +α, so P(α) = exp(−2|β+α|²) for coherent |β⟩; the tests assert this and the equivalent Wigner-point form at −α. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
# Conflicts: # test/run_filtered.jl
…_shots override SimulatedExperiment gains an rng slot (constructor kwarg, default nothing = exactly the legacy deterministic behavior). With an AbstractRNG, each noisy measurement element gets ε ~ N(0, Σ(y)) drawn from its noise model at the evaluated y (variances floored as in whiten); deterministic elements draw nothing, keeping the rng stream layout-stable. run_experiment gains an additive n_shots kwarg on both experiment paths: SimulatedExperiment rebuilds ShotNoiseMeasurement value-copies with the boosted n (noise scales exactly 1/√n for a fixed seed); HardwareExperiment forwards to run closures that accept the kwarg, else warns once and ignores. Replaces the runner's deleted add_measurement_noise/SIM_NOISE contract and gives fixtures seeded reproducibility. Full suite green (193/193). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
y_goal::Union{Nothing,Vector{Measurement}} constructor kwarg: nothing
(default) keeps today's behavior — solve! resolves the goal once at solve
start via the strategy's tuning_goal hook; supplied → used verbatim. The
resolved goal is computed exactly once per solve and never mid-solve (the
chained-loop-drift invariant: per-round y_goal recomputation from the
current command made real campaigns diverge). Campaign scripts pass the
ideal target built from ψ_goal via the public observables.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…avior-preserving) Extracts the chassis's Armijo accept/interpolate block and the γ/tr_scale schedule into an AcceptancePolicy seam: decide(policy, ctx) returns (α, accepted, revert, tr_scale, n_evals); the chassis keeps run-experiment → whitened Ĵ/J̃ (via the noise-stats helpers; identical to raw SSR for deterministic models) → decide → apply α-interpolation (or revert to the last-accepted snapshot) → record. LineSearchAcceptance(γ, line_search) reproduces the pre-seam behavior verbatim — pinned by a hard-coded (α, tr_scale, n_experiments, J) trace regression recorded on the pre-seam chassis across the reject cascade, partial-α accept, and line_search=false paths. PulseTuningProblem gains an acceptance slot (nothing ⇒ legacy kwargs). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…trust scaling The hardware-hardened acceptance policy: no probing (α is always exactly β, one experiment per iteration); reject-and-revert iff Ĵ − J_base > max(ρ_rej·J_base, k·σ_Δ) with σ_Δ = √(σ_J²(trial)+σ_J²(base)) from the plug-in cost statistics — the catastrophe cascade cannot fire on a noise fluctuation, and an all-deterministic model (σ_Δ = 0) reduces to the legacy pure ratio test; tr_scale shrink-only (halve on reject — permanent — and after stall_patience flat trials), floored at tr_scale_min; the first measured trial always establishes the base. The chassis reverts to the last-accepted snapshot on rejection and discards the candidate. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ep, dimension-change reset IterateSelector seam: FinalIterate (legacy default), NoiseCorrectedBestJ (argmin J̃ over accepted iterates — debiased selection removes the documented best-J bias), TopKRemeasure(k, reps_factor) (re-measures finalists via the n_shots override at n·reps_factor and picks the re-measured winner; fewer than k accepted ⇒ re-measure all), PolyakAverage (reproduces the former polyak_avg behavior exactly; the kwarg now forwards with a deprecation note). IterationRecord gains J_hat, accepted, F_model (new last_f_model strategy hook), and θ — the measurement-time trajectory snapshot selectors restore. solve! gains a logger kwarg and calls record!(logger, rec) once per iteration (InMemoryExperimentLogger collects them; other loggers opt in). adapt!(device_model, …) now runs BEFORE the strategy step — recalibration must precede Jacobian assembly (ordering inversion per spec). A mid-campaign measurement-dimension drop (lab σ_z fallback) now restricts the active goal/model, resets the acceptance base, and warns. Full suite green (267/267). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
armijo_line_search gains a cost::Function kwarg (default: legacy raw measurement_error). LineSearchAcceptance passes the plug-in whitened cost so backtracking compares like units with its whitened Ĵ reference — a raw trial cost against a whitened J_ref silently accepts every α = 1 step the moment a ShotNoiseMeasurement is declared, and the campaign diverges (caught by the Intonatissimo shot-noise campaign cells). Deterministic models are unchanged (w ≡ 1 reduces the closure to measurement_error). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…n strategy ctx
Task-importance weights (flattened per element, e.g. the σ_z×40 leakage
defense) become a chassis knob: composed into every whitened cost as
W = W_task·Σ^{-1/2}, restricted alongside the goal/model on a mid-campaign
dimension change, and forwarded to TopKRemeasure's re-measure whitening —
chassis, strategies, records, and selectors speak one unit. The strategy
ctx now also carries the ACTIVE measurement model and the chassis (w, σ2)
so strategies can whiten their Jacobians consistently.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…TuningProblem Model and experiment must share operator structure (frame); violation silently null-spaces the Jacobian and stalls the loop (the feat/lab_frame campaign lesson). Companion of the Intonatissimo QP-ILC loop docs. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Member
|
Review (adversarial pass against the approved spec/plan; statistics and behavior-preservation traces re-derived independently — they check out). Verdict: mergeable after fixes. Blocking
Recommended Also suggest bumping the version (the adapt!-before-step ordering change is behavioral, and downstream compat needs to distinguish this chassis from the released 0.2.0). |
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
feat: noise-native measurement layer + chassis policy seams (QILC port, public substrate)
Implements the public half of the QILC stanford port per the approved spec:
noise_floor/debiased_cost/cost_std/diff_std(diagonal fast path) and
whiten(model, y_exp; W_task, var_floor)—GLS
W = W_task·Σ^{-1/2}from the (previously inert) measurement noisetypes, plug-in per iteration. Deviation: default
var_floor = 0.05(not the plan's 0.25 — the plan's own acceptance test requires the y=0.9
wigner element unfloored at defaults).
SimulatedExperiment(...; rng)drawsε ~ N(0, Σ(y)) per noisy element;
run_experiment(...; n_shots)value-copyoverride on both experiment types (TopKRemeasure contract).
measurement_functions/):displaced_parity(x, α; Nq, Nc)(kwarg-routed beside the cavity-only path),qubit_sigma_z, ρ triangle encode/decode,reconstruct_rho_from_parity(docstring: MLE reconstruction is diagnostics-only, never ILC feedback).
PulseTuningProblem:acceptance::AcceptancePolicy—LineSearchAcceptance(behavior-preservingextraction of the Armijo + γ-schedule path, pinned by hard-coded pre-seam
traces) and
OneShotAcceptance(β-damped apply, σ_Δ-awarereject-and-revert, shrink-only trust scaling, stall anneal).
selector::IterateSelector—FinalIterate,NoiseCorrectedBestJ(argmin debiased J̃),
TopKRemeasure(k, reps_factor),PolyakAverage(subsumes the deprecated
polyak_avgkwarg).y_goalkwarg (resolved once; the chained-loop-driftinvariant) and
W_task(task importance composed into every whitened cost).IterationRecordextended (J̃ + Ĵ, accepted, F_model via the newlast_f_modelstrategy hook, θ measurement-time snapshot, timings) +solve!(...; logger)seam.adapt!(device_model, …)now runs BEFORE the strategy step (recal mustprecede Jacobian assembly).
W_task restriction + acceptance reset + warning.
armijo_line_searchgains acostclosure soprobes compare in the same (whitened) units as
J_ref— a raw trial costagainst a whitened reference silently accepts every step once shot noise is
declared (caught by the Intonatissimo campaign cells).
Suite: 270/270. Every extraction is pinned by a behavior-preservation
regression recorded on the pre-seam chassis.
Verification: suite 270/270 at
58b9ecd. Behavior-preserving extractionsare pinned by hard-coded traces recorded on the pre-seam chassis (reject
cascade, partial-α accept, line_search=false, Polyak). Downstream, the full
Intonatissimo suite (which consumes every seam here) runs 427 pass / 0 fail /
2 expected B-spline skips, and the end-to-end acceptance gate through the new
stack reached F_truth 0.99718 with ≥ 0.99 sustained over the final 20 of 30
iterations under χ×1.15 + K_c×2.2 + 1000-shot noise + 1 %/trial drift.
Stack: this is the base of the QILC-port stack — Intonatissimo
feat/qp-ilc-strategy(harmoniqs/Intonatissimo.jl#44) and stanford-bosonicsport/intonatissimo-chassis(harmoniqs/stanford-bosonics#6) build on theseseams. Merge order: this PR first.
Spec:
armonissima/specs/spec-20260701-161246-qilc-stanford-port-noise-native.md·Plan:
armonissima/plans/plan-20260701-235146-qilc-stanford-port-noise-native.md🤖 Generated with Claude Code