INDEPENDENT ENGINEERING PROJECT

Image-quality decisions.
Evidence you can inspect.

A photographer workspace and evaluation toolkit that connects rendered-image observations, conservative selection, human review and reproducible regression checks.

Explore the working demo Review the evidence ↓

No account is required for the guest demo. Guest drafts are temporary. You can try your own photos for the visit; saved work requires sign-in.

01 / VERIFICATION

What the checks establish

Recorded 2026-09-08 UTC — integrated release checkpoint. Automated checks use curated expectations and test doubles; the hosted checks below exercise the actual site. Independent real-photo accuracy has not been established.

Decision cases passed: 204 / 208
Decision cases failed: 0
Historical cases blocked: 4
Account / API checks passed: 121

Four frozen historical audit cases lack approved executable expectations. They stay blocked. The API suite covers account isolation, approval gates, persistence conflicts and malformed updates using disposable SQLite and an R2 test double.

Hosted checks preserved two synthetic uploads and their original bytes, saved comparison and Investigation history after refresh, and actual PDF/JSON downloads. The twelve-photo report retained its images and credits in thirteen pages. These checks do not establish performance under load or photographic accuracy.

The lint gate passes with 0 new diagnostics; 113 known findings remain. This is a recorded checkpoint, not a live service monitor. Checked source: d186f1d18042. Download the verification record.

02 / SYSTEM DESIGN

From original to accountable decision

Original preserved
Rendered preview
Measurements + optional models
Context + one culling engine
Human review + report

Browser

Preview processing and optional photo-model inference use the visitor’s device. Model download, WebGPU availability and browser memory affect performance.

Cloud

Worker APIs enforce account ownership. D1 stores collections and revision pointers; R2 stores originals, previews and saved lab evidence. Feature-plan generation sends the entered requirement text to the configured OpenAI service.

Photo-model inference and cloud text-plan generation are separate paths. QA replay calls the same production culling engine and does not write collection decisions.

03 / METHODS

Implemented behavior and its limits

Method	Implemented	Interpretation limit
Rendered-image measurements	Near-white / near-black fractions and regional tonal information. Edge detail and smooth-patch noise are estimates.	Preview pixels cannot establish RAW recovery, calibrated color accuracy, sensor SNR or optical MTF.
Optional local vision models	MediaPipe localization, CLIP content / embedding hints, and SmolVLM observations support inspection.	Pretrained models were not trained by this project. Their output is advisory; confidence is not calibrated.
Conservative selection policy	One production culling engine combines available evidence with photographer intent, protected moments and retained alternatives.	Decision-replay tests check specified behavior. They do not measure agreement with photographers on unseen events.
Before/after repair quality	Aligned pixel measurements and texture proxies localize changes. Optional two-image vision covers twelve quality factors, with matching crops, human review and saved evidence.	Alignment and an independent allowed region are needed for preservation measurements. Visual AI can be wrong; no dataset accuracy or calibrated realism score is claimed.

The SSIM implementation uses non-overlapping complete 8×8 windows on 8-bit rendered luminance. It is not a claim of numerical parity with a library default. See the original SSIM method and reference implementation options.

04 / ENGINEERING INVESTIGATION

A small input exposed a large allocation

A 1 × 10,000 image revealed two sizing defects: its preview width rounded to zero, and a fixed 256-pixel statistics width implied 2,560,000 rows. A browser might reject that canvas before allocating it; an actual device crash was not established.

Correction

The preview now keeps both dimensions between 1 and 1,600. Statistics stay within 256 × 512. Ordinary landscape and 2:3 portrait sampling is preserved. A finally block releases decoded bitmaps after success or failure.

Regression evidence

The production importer is exercised with five aspect ratios, an over-100-MP input, and two failure paths. A Canvas / decoder double checks requested sizes and resource release without creating a huge raster.

Remaining boundary: the browser decodes the input before the application can check its pixel count. This change bounds subsequent canvases; it does not guarantee a maximum browser memory footprint. Native decoding and large real batches still need device testing.

05 / TRY & REPRODUCE

A focused review in five minutes

Open the guest demo. Explore the sample collection. Sample annotations illustrate the workflow and are excluded from real benchmark truth.
Visit Evaluation Lab. Record a separate photographer judgment, reveal the app decision, and inspect a disagreement with its saved evidence. Engineering tools retain the curated automated suite and its expected/actual results.
Inspect GenAI Repair. Compare an aligned original and edited image, inspect measured regions and visual findings, then record your judgment and create a Lab investigation. The guest view supports local measurements; visual AI requires sign-in.
Read the limits below. A visible result is useful only when its evidence supports the claim.

Reproduce from a source checkout

Use Node 24 and pnpm 10 with the checked-in lockfile. The commands below run local checks without requiring paid model generation.

pnpm install --frozen-lockfile
pnpm verify
pnpm qa --output ./outputs/qa-run.json

The GitHub Actions recipe runs the same release checks when hosted on GitHub. The current Sites source host does not execute that recipe. No green remote CI run is claimed.

06 / NEXT VALIDATION

What remains to be demonstrated

Independently reviewed real photographs with event-separated holdout data, error examples and uncertainty intervals.
Detector calibration and task-specific studies of sharpness, noise and exposure, including failures and intentional photographic effects.
Measured browser/device latency, memory behavior, batch capacity and authenticated upload retry / cancellation under real load.
Immutable upstream model revisions in exported evidence, reduction of existing lint debt, and broader hosted browser acceptance.
Production identity-provider configuration and normal session acceptance. The optional Clerk sign-in currently uses a development instance; existing ChatGPT accounts retain their original collections.

HEIF / HEIC, HDR or video evaluation, RAW sensor analysis, camera-pipeline tuning, verified Content Credentials and reliable blur-cause / processing-artifact diagnosis are outside the demonstrated scope.

Image-quality decisions.Evidence you can inspect.