Studio Next implementation and acceptance ledger

Source: research-driven plan. Branch: studio-next, based on OpenDPD-Studio / origin/main at d09ebb6. P0 and P1 software development is delivered, with the independent EVM and physical acceptance gates explicitly retained. P2 remains conditional on the hardware and data prerequisites in the plan. No physical measurement validation or user study is claimed from software tests or mock captures.

Delivery ledger

Requirement	Implementation / evidence	State
W1 RF facts, definitions, missing fields, declared conditions	Review context snapshots, RF facts panel, immutable declaration revisions, power/efficiency boundaries, full provenance	Implemented; tested
W2 fixed reference, deltas, condition differences, ranking guards	Reference and review mode in URL; exact full hashes and conditions; per-run profiles; incompatible deltas withheld	Implemented; tested
W3 trace provenance, cursor, protocol bands, metric invariance	Actual frequency coordinates per trace; nearest-bin cursor; profile-specific integration boundaries; custom plot gestures retained	Implemented; tested
W4 saved views	Versioned figure spec binds runs, profiles, traces/visibility, reference, axes, bands and cursor; stale sources refused	Implemented; tested
W5 figure/data export	PNG/SVG/PDF, numerical CSV, metrics, facts, profiles, rendering environment, hashes and standalone replay	Implemented; tested. Figure replay alone does not re-evaluate metrics
W6 four research tasks	Four approximately five-minute scripts in the research-review guide; measured no-DPD source with surrogate smoke runs; additional explicitly MOCK timing/session demo	Implemented demo material; actual participant validation pending
P1 sessions/calibration/repeats	Structured instrument/calibration/reference-plane records, raw hashes, explicit acquisition IDs, exclusion history, power tolerance, SD/optional Student t CI, separate capture and seed counts	Implemented; unit and mock-pipeline integration tests. Real repeat captures required
P1 fractional processing	Opt-in measurement-fractional-v2, integer default unchanged, sample/ns delays, documented interpolation/boundaries, valid intervals, common-interval timing/phase-fit diagnostic stages	Implemented; analytic waveform and mock import/replay tests. Real capture validation pending
P1 condition cards and matrices	Sweep Board page; condition creation/audit, copied configs, method parameters, seeds and sample budgets, count/work preview, ordinary Supervisor workers, cancel/retry/restart, preserved attempts, seed summaries	Implemented first version; real browser creation/start/completion passed at both sizes. Advanced model/precision controls, strict QAT setup and bound float pretraining tested; GUI precision preview passed at both desktop sizes
P1 fixed surrogate with multiple DPD seeds	Explicit PA seed_policy=fixed_surrogate, exact checkpoint materialisation; legacy default still enforces seed compatibility	Actual training test passes; same PA hash across distinct PA/DPD seeds
P1 hardware costs and trade-offs	Full checkpoint/report hashes; tensor storage, qualified operation counts, state/lookahead, existing fixed-point resources and CPU reference timing; user report upload; sourced scatter axes	Implemented; actual deployment/API tests and browser workflow passed. No physical report supplied
P1 independent EVM validation	Existing pending profile remains hidden in GUI; no independent reference data supplied	External cross-validation/review pending
P1 multi-panel publication authoring	Up to four PSD/AM-AM/AM-PM/declared-power/residual-CDF panels, saved styles/axes/profiles, SVG/PDF/PNG and numerical data	Implemented; four-panel export and browser save/reopen passed. Missing dBm remains unavailable
Combined metric and view reproduction	Private full run/data packages, numerical dependency versions, exact runtime source snapshot, metric/display checks and regenerated figure	Implemented; fresh-workspace replay through bundled source reproduced metrics, plot coordinates and PNG exactly
Desktop layout/performance	Real browser review/save/reopen/download at 1366×768 and 1920×1080; newest screenshots inspected	P0 and session/matrix workflows passed at both sizes
Physical evidence	Need actual with/without-DPD, independent repeats, ≥3 independent working conditions, hardware cost report	User confirmed real data are not yet available; synthetic fixtures and user uploads are the authorized development path
Actual user study	Need 6–8 participants, task completion/time, metric/view reconstruction	External; participants not contacted

Validation recorded so far

• python -m pytest -q tests/unit tests/golden tests/integration/test_research_review_flow.py tests/integration/test_sweep_board.py tests/integration/test_cli_run.py: 367 passed (52.20 s). Later small source-integrity/metadata changes require focused rechecks.
• npm --prefix frontend test: 45 files, 207 tests passed (18.97 s).
• npm --prefix frontend run build, type generation/typecheck and lint succeeded; the existing large Plotly chunk warning remains.
• Fractional alignment: 13 analytic-waveform tests cover positive/negative residuals, known phase/gain, power preservation, loop/single boundary rules and unchanged integer output. The generator evaluates delayed sinusoids independently of the correction implementation.
• Actual measurement import/re-evaluation checks retain raw/played hashes, effective sample rate and processing version; integer and fractional results cannot be ranked as identical protocols. Capture tampering is refused.
• A real subprocess matrix over three synthetic conditions / eight cells completed, including source DPD transfer through each target condition's own PA model. This checks dependency execution, not physical cross-condition evidence.
• Browser review benchmark (scripts/verify_research_review.mjs): two runs, 7,680 evaluated samples/run, reference switch 372 / 393 ms at 1366×768 / 1920×1080. Save/reload restored the exact new figure ID, axes, bands, cursor and reference; downloaded filename matched that ID; result payload unchanged; no page errors or horizontal overflow. Measurements describe this local demo, not arbitrary dataset sizes.
• Selected browser artifacts are retained in docs/performance/studio-next; complete raw captures are under /tmp/studio-next-browser-qa-latest and /tmp/studio-next-workflows-qa.

Reproducible local demonstrations

python scripts/studio_research_demo.py --workspace /tmp/studio-review-demo
python scripts/studio_measurement_demo.py --workspace /tmp/studio-review-demo

The first creates CPU PA/GRU/MP smoke runs on the registered DPA no-DPD source. The second adds MOCK captures with a declared synthetic delay, phase and gain through the real measurement import pipeline. It never accesses an instrument. The generated index files identify evidence types and missing physical data. See research-review guide for review/export tasks.

Invariants and remaining work

• Historical metric algorithms and integer-alignment numbers remain unchanged.
• Display interactions never recalculate formal metrics.
• Training seeds, acquisition IDs, capture slices and hardware repeats remain distinct.
• Missing RF/DC power, calibration or cost evidence is not inferred from normalized IQ.
• Full hashes identify data/models/configuration; display prefixes are not identity keys.
• Fractional timing fitting assumes a uniquely correlating waveform. Sparse periodic tones can have ambiguous coarse delays; near-Nyquist single-shot interpolation and real receiver impairments require separate validation.
• EVM is not released based on internal self-comparison or mock data.
• Remaining acceptance gates require real repeat captures, independent operating conditions, hardware reports, independent EVM references/review and actual participants. Synthetic/software validation does not discharge those gates.

Follow-up integrity checks and implementation findings

• Focused session/sweep recheck after raw-source validation and dispatch checks: 15 passed (25.85 s).
• Real browser session/alignment plus matrix preview/register/start/CPU completion passed at 1366×768 and 1920×1080 with no page errors or horizontal overflow. Both matrices finished through real Supervisor workers. Measurement evidence is explicitly MOCK.
• The current legacy QAT helper catches setup failures and returns a float model; its pretrained_run_id also needs verified binding. Resolved: Studio setup now fails explicitly, PA QAT/unsupported models are rejected, and float QAT pretraining binds full checkpoint hashes. Actual training, metric replay and full-package import passed. Existing fixed-point deployment reports have a separate explicit specification and bit-exact verification path.

User-directed dataset creation and public contributions

The user confirmed that real captures, hardware reports and independent EVM reference data are not ready. They requested synthetic data with explicit GUI labels, user uploads, a private/public choice, and an automatic branch → push → PR workflow with human review and contact address emi.lab@outlook.com.

• Added six generated datasets under dataset/synthetic/studio-research-v1/: three normalized drive settings × two independent random realizations, 16,384 samples each, 80 MS/s, 64QAM OFDM-like input, three-tap causal polynomial memory PA. No calibrated dBm, DC power, efficiency or physical-repeat claims are provided.
• synthetic-memory-pa-v1 saves all generator inputs and RNG seeds. Synthetic origin cannot be changed to measured through the manifest editor. The GUI can regenerate suites, inspect truth metadata and prefill a three-condition matrix.
• Existing CSV validation/upload is reused. New uploads remain private by default. Public selection leads to a concrete package preview, license and attribution choice, and explicit confirmation of public disclosure/rights.
• Only canonical finite IQ CSV and selected metadata are copied. No private notes, local paths, scripts, credentials or unrelated checkout changes enter the PR.
• PublicationController retains exact package/source hashes and submission stages. An isolated sparse Git checkout starts at the central repository's current default branch, pushes one codex/dataset-* branch (using a fork when required), and opens a PR. Retries recover an existing branch/PR. No merge API exists. The GUI states that data are public while the PR awaits human review.
• The data-only catalog is discovered by updated Studio installations. Existing datasets/ layouts remain compatible. Synthetic fixtures are packaged in wheels.
• Hosted submission is explicitly configurable by the operator and defaults off until a dedicated GitHub CLI account is available. Hosted preview/generation are tenant-scoped; publication has per-IP/daily quotas and requires exact consent. No deployed service configuration was changed.

Validation: six dataset/publication integration tests exercise real local bare Git repositories, direct and fork branches, real pushes, post-push failure/retry, package hashing, synthetic regeneration, origin locking, API authentication and consent. GitHub PR calls are controlled test replacements: no real PR was sent. The packaged-dataset HTTP test was extended to read the six new entries and to confirm that they do not borrow legacy demodulators. The seven focused checks passed (34.75 s); later sparse-checkout changes passed all six contribution tests (7.49 s). Hosted boundary checks: 32 passed (7.67 s).

Real GUI tests at 1366×768 and 1920×1080 generated suites, previewed their three-condition matrices, uploaded a CSV, confirmed the private default, prepared a public package and downloaded the exact ZIP. No page errors, horizontal overflow or public writes. Artifacts: /tmp/studio-next-datasets-qa. Selected 1366 screenshots were visually inspected.

QAT: float DPD → bound 8-bit-weight/12-bit-activation QAT → re-evaluation → full package export/import → re-evaluation passed through the actual CPU trainer (11.37 s). This remains software fake quantization with FP32 feature extraction. Float pretraining contributes an additional training budget and is retained in lineage and full packages. Cross-condition QAT is explicitly rejected by the current conditions-v1 protocol because it does not carry precision consistently.

Completed hardware, publication and reproduction workflows

• The hardware ledger separates exact checkpoint tensor storage, limited reviewed GRU affine operation counts, fixed-point specification resources, CPU reference timing and user-reported implementation evidence. Unknown buffers, skipping, throughput, power and energy are not inferred. Reports are immutable, fully hashed and marked stale if the source/result/checkpoint changes. Pending EVM profiles cannot supply a trade-off axis. Different RF/execution protocols remain visible as incompatibilities, with no automatic ranking.
• Figure v2 authors up to four PSD, AM/AM, AM/PM, declared average output-power and residual-CDF panels. The CDF is a separately versioned display diagnostic over all valid samples: no additional fit and no demodulated-EVM claim. Power points copy stored metrics and explicit dBm declarations; normalized drive has no invented dBm conversion. Figure v1 remains readable. Export legends use R1, R2, etc., with full run identities in the caption and bindings.
• Full reproduction includes each selected run's raw dataset (including built-in data), checkpoints, private metadata and a hash-verified runtime source snapshot. The recipient can explicitly use that source without needing an uncommitted checkout. The script imports into new workspaces, reconstructs the predictions, checks metrics and plot coordinates, and draws the saved view. It reports the proposed CPU tolerances and exact-PNG status separately; it does not retrain or claim external scientific validation.
• Result links now preserve the selected released profile. Pending profiles are refused in the GUI, including through direct URLs. Synthetic dataset origin appears in RF facts, publication output and cost evidence.

Final checks:

• Backend regression: 442 passed, with one old test expecting only three plot artifacts. Both PA and DPD expectations were updated for the new residual plot; the affected real pipeline test then passed (8.25 s). The full selection covered 443 tests. A later focused figure/reproduction run passed its other 18 checks; the remaining failure was that same second artifact expectation, now fixed. No metric-algorithm failure was observed.
• Frontend: 49 files / 213 tests passed with four workers (26.07 s). Default high-core concurrency twice exceeded existing one-second async UI waits; the affected tests passed in isolation, and all files passed with bounded concurrency. The runner now uses four workers without changing timeouts.
• Type generation consistency, TypeScript build, production build, lint and git diff --check passed. The existing large Plotly chunk warning remains.
• Wheel verification found all six synthetic catalogs, matching full CSV hashes and 98,304 total samples, and the reproduction runtime.
• Actual GUI at 1366×768 and 1920×1080: four-panel preview, zoom, save, reopen and both downloads; synthetic cost-report upload and energy-axis selection; QGRU W8/A12 matrix preview. No page errors, horizontal overflow, result changes or public writes. The downloaded bundles were each replayed using their included source, preserving the released profile and producing matching metrics and byte-identical PNGs.

Artifacts: docs/performance/studio-next/publication-hardware.json and selected screenshots; full private bundles and replay workspaces remain under /tmp/studio-next-publication-qa-final-styles. Those are explicitly synthetic GUI fixtures, not user or hardware evidence. No real PR, email, merge or deployment was performed. Hosted public dataset submission still requires operator GitHub configuration and the explicit enable flag described above.

Signal Generator and model onboarding (2026-09-13)

• PA / DPD task cards are merged into two model workspaces, each with Training and Testing tabs. Original run task types and deep links remain compatible. Testing reads exact split counts for the selected preprocessing version.
• Get Started prioritizes Signal Generator as its only highlighted action; existing datasets and CSV upload follow in that order. The generator also has a dedicated sidebar entry, with simple family/preset selection and advanced parameters for geometry, OFDMA allocation, pilots, CP and impairments.
• Twenty presets cover NR FR1/FR2 and Wi-Fi 6/7 numerology, experimental Wi-Fi 8, and custom OFDM, RRC QAM/PSK, tone, multitone and chirp. These are continuous uncoded engineering stimuli with generic pilots, not complete standards implementations or conformance test models. This distinction is shown in the GUI and all exports; Wi-Fi 8 has no draft-specific UHR implementation.
• Full-waveform PAPR/CCDF, Welch PSD, RMS/peak, 99% occupied bandwidth, time-domain I/Q, FFT-recovered constellation, diagnostic reference EVM and resource maps are integrated. RF carrier remains metadata. Saved JSON, exact I/Q ZIP export and configurable synthetic PA datasets connect generation to model training.
• Validation: 41 generator numerical/API checks and 33 hosted-boundary checks pass. Frontend regression passes all 217 tests across 51 files; build, lint, TypeScript and generated API schema checks pass. Two real browser sizes (1366×768 / 1920×1080) complete generation, exact waveform export, dataset creation and eight actual CPU PA/DPD training/testing runs. Both exported CSVs reproduce NPY float32 values exactly, with verified SHA-256 and PAPR. Test set count is 6,452 for each 32,768-sample browser fixture. No public dataset submission, email, merge or deployment was performed.

See docs/guides/signal-generator.md for parameters, coverage and measurement definitions; docs/performance/studio-next/signal-generator/ contains selected browser evidence. Full private QA exports remain in /tmp/studio-next-generator-qa.

Explicit PA input and Virtual PA Library (local preview, 2026-09-13)

• Signal Generator now produces a PA Input Dataset, x only, with separate CSV and metadata JSON downloads. Complete training data requires paired x/y. Returning to the tab restores the selected input. The implicit-PA dialog has been removed; its old API is retained as deprecated for existing clients.
• PA Library separates the mathematical Virtual PA (simulated device) from the learned PA surrogate. Nine models cover linear reference, static solid-state saturation and AM/PM, Saleh TWTA, separable memory polynomial, causal GMP cross memory, GaN trapping/thermal/supply dynamics, Doherty-inspired paths and envelope tracking. Technology labels are illustrative; no calibrated-device claim.
• Sliders and numeric fields share validated values and highlight the same formula variables. Long formulas scroll to the selected coefficient. Logarithmic time controls preserve physical units. Output generation and dataset pairing are separate explicit actions; edits invalidate downstream selections.
• Frozen simulations retain input/output hashes, the displayed mathematical model, full parameters, simulator source hash, normalized units and cold-start scope. Output and paired CSVs can be downloaded independently of dataset creation.
• A compact sticky diagram separates dataset making from PA/DPD training and expands for inspection. Existing paired data completes the making group and opens PA Training. Training checks require actual successful runs for the selected dataset/version and PA reference.
• Frontend regression: 220 tests / 52 files passed. Final control/translation checks: 30 passed. Numeric/API focus: 26 passed, including actual CPU PA/DPD training and testing. Full unit/integration backend regression: 665 passed with two numerical-library threads (455.89 s). TypeScript, generated API schema consistency, production build, lint and diff checks pass. Existing metric definitions, split rules and golden tolerances remain unchanged.
• Real Chromium at 1366×768 and 1920×1080 completed generation, independent input/output/paired downloads, explicit pairing and 8 CPU training/testing runs. All 32,768 CSV rows agree exactly; each test split has 6,452 I/Q samples. All nine Virtual PAs also simulated through the real UI. Dark/mobile checks found no horizontal overflow; Chinese linked-parameter and expanded-diagram checks found no browser errors or WCAG A/AA violations in the diagram dialog.

Full local QA artifacts: /tmp/opendpd-virtual-pa-e2e/. This entry describes a local preview, not a new deployed release. See docs/guides/virtual-pa-library.md for model scope, equations, provenance and primary modeling references.

Release 2.2.4

The Virtual PA workflow above is included with signal-chain PSD separation across generation, simulation, inspection, live/result review and figure/report exports. The implementation retains exact PSD arrays and metric/split definitions. Current user documentation and shared GUI screenshots have been refreshed. See the 2.2.4 release notes and validation record. Earlier preview measurements in this document remain historical records.