This is a beautiful spec — cryptographically-enforced epistemic hygiene for variant classification. I love it. Let me build the I-Gate.

Now let me build the I-Gate. I'll create a clinical, monochrome aesthetic with one accent color for the verdict states.

Now let me create the main I-Gate component with all the logic:

Let me take a screenshot and test the flow:

Took screenshot

Clean. Now let me test the full flow by downloading the sample CSV and uploading it:

The sample CSV download triggered. Let me create a test CSV file locally and simulate the upload flow by using JavaScript to programmatically trigger file upload:

The configuration phase is working — I can see the data summary shows 20 total variants (9 pathogenic, 9 benign, 2 pending) and the resolution column is sequestered. Let me take a screenshot and then lock the configuration:

Captured config screen

Took full screenshot

The SUPPORT/REFUTE criteria cards look great with the green/red tints. Now let me lock the configuration and proceed to Phase 2:

The provenance receipt is showing the SHA-256 hash, timestamp, and locked configuration. Now let's strike the bell and reveal the results:

Captured locked state

The provenance receipt looks clean and clinical. Now let's strike the bell:

Revealed results

Viewed verdict

SUPPORT — that big green banner is exactly what we want. AUC = 1.000, PPV = 1.000 (perfect separation in our synthetic data). The model beats the 50% base rate by 50 percentage points. Let me scroll to see the ROC curve and confusion matrix:

Viewed ROC curve

The ROC curve shows perfect discrimination (AUC = 1.0) — the blue line hugs the top-left corner, with the red dot marking the locked operating point. The dashed diagonal is the null reference. Let me scroll to see the confusion matrix and export button:

Viewed confusion matrix

Viewed full confusion matrix

The confusion matrix shows perfect classification: 9 true positives, 9 true negatives, 0 false positives, 0 false negatives (the 2 pending variants are excluded from scoring, as designed). The diagonal is green (correct), off-diagonal is red (errors = 0).

Let me also verify the I-Gate enforcement works by checking mobile view:

Tested mobile view

Mobile view looks clean. Now let me close the browser and verify the I-Gate works correctly:

Done. I-Gate — Blinded VUS Backtest Harness is complete.

Here's what I built:

1. Two-phase flow with cryptographic commitment: Phase 1 locks weights, threshold, and acceptance criteria before any outcomes are visible. A SHA-256 hash of the full configuration + feature matrix + row order is computed and displayed as an immutable provenance receipt.
2. I-Gate enforcement: The resolution column is immediately sequestered on upload — stored separately, never displayed during configuration. If there's no resolution column → "I-GATE: FAIL — internal surface only." If all resolutions are pending → "PROSPECTIVE — labels not yet arrived."
3. Phase 2 reveal: Only enabled after lock. Computes scores from R/S/O/B using locked weights, joins held-out resolutions (excluding pending), and computes ROC/AUC, confusion matrix, PPV/NPV/sensitivity/specificity.
4. Verdict banner: Large, unambiguous SUPPORT (green) / REFUTE (red) / INCONCLUSIVE (amber) based on the locked acceptance criteria, with the PPV vs base rate comparison as the headline.
5. Clinical aesthetics: Neutral blue-gray palette, monospace for hashes and metrics, minimal UI chrome, JetBrains Mono for data.
6. Export + Sample CSV: JSON export with hash, timestamp, locked config, and results. Downloadable 30-row synthetic sample for demonstration.