Capability remediation finalization

Historical / superseded · Preserved for transparency. This page does not describe the current HydroCore-v5 finalist.

This branch remediated the demonstrated serving defects without changing the HydroCore-v4 weights: canonical structural identity, calibration/OOD applicability, causal telemetry history, missing-health semantics, the 25-step feature window, calibration group selection, and sampling timing. Historical diagnostic and pre-remediation evidence remain unchanged.

Scientific identity

The final campaigns ran against code commit 41aae8795cd2d25894f9521f5ac7ea17d04256a5. HydroCore-v4 remains a501ad87…; feature schema, normalization, signature policy, fusion, and authority thresholds are unchanged. Calibration was previously refit solely from the designated calibration split after the canonical topology identity change. The locked test and locked topology test remain unopened.

The governed canonical golden hash is 5508d272…; programmatic and EPANET paths resolve the same identity. Supported topology observations are calibrated and OOD-NORMAL; unseen coastal is calibration-invalid, topology-novel (1.0), OOD-CAUTION, and cannot plan.

Causal evidence and model decision

The causal temporal curve remains: top-1 0.15/0.50/0.45/0.80/0.80 at 1/2/3/6/25 steps, versus 0.20 for final latest-only evidence. Controlled validation reproduced top-1 0.7205, top-3 0.8680, and MRR 0.8113.

CAP-REM-01 is therefore a causal-prefix training-distribution limitation. It does not establish a parameter-capacity bottleneck. No retraining was performed here; the appropriate follow-up is a separate causal-prefix HydroCore-v5 experiment.

CAP-REM-02: active sampling

The final paired development experiment used 40 canonical golden incidents, three causal steps, 50% initial sensor coverage, three samples, 30-minute acquisition delay, and matched seeded 0.05 mg/L noise.

EIG had median realized entropy reduction 0.9572 bits and positive realized reduction in every EIG round, with expected/realized Spearman 0.544. It did not demonstrate better operational actionability: EIG was actionable within three samples in 0.375 of incidents; random valid-unsampled was 0.450.

The post-sample decomposition found no stale analysis, calibration-group, timestamp, or API/runtime divergence. In localization-correct-but-suppressed states, conformal candidate breadth was the dominant blocker (48 states), followed by learned model-evidence insufficiency (20). The candidate-region blocker was the sole blocker in 45 states. Thresholds and alpha were not retuned.

CAP-REM-02 is consequently scoped as a current product limitation: active sampling is evidence guidance that can reduce uncertainty, but current development evidence does not show that it reduces samples-to-actionability versus random in the sparse causal regime. It remains advisory/experimental until revisited after causal-prefix training.

Full production-path LIVE campaign

The frozen 264-run API-driven LIVE matrix completed with zero authority invariant failures. A separate six-run clean supported-topology control slice then covered branched-loop and loop-grid without altering that frozen matrix. Together the 270 API-path runs cover nominal, missingness, sensor health, sensor coverage, measurement noise/bias, hydraulic shifts, ambiguity, unseen topology, scale, and all three governed supported networks. Overall analyzed outcomes were top-1 0.9087, top-3 0.9722, MRR 0.9472; supported nominal was 0.8889/0.9444/0.9278. Canonical supported runs had no false calibration-invalid result and were OOD-NORMAL; coastal remained calibration-invalid, topology-novel, OOD-CAUTION, and ineligible to plan.

The combined campaign generated 367 plans, 185 VERIFIED results, 182 ABSTAINED plan results, and 367 exact WNTR calls. It acquired 23 samples; no already-observed recommendation occurred. ROB-LIVE-01 and ROB-LIVE-02 are remediated.

The frozen matrix does not define cumulative actionable within 1/2/3 metrics: it retains final authority and selected post-sample states, not a complete initial-to-round authority trajectory for every incident. Those rates are therefore intentionally null in live-capability.json; the 40-incident paired sampling campaign is the authoritative measurement of actionability within a sample budget.

See the machine-readable final record in reports/evaluation/capability-remediation/, notably sampling.json, sampling-blockers.json, full-live-results.json, supported-topology-controls.json, full-live-summary.json, safety-regressions.json, validation.json, and summary.json.

Local release validation

The final local validation record is published in validation.json: 1,136 Python tests passed with one documented historical skip in 657.36 seconds; the strict self-test passed; and frontend lint, typecheck, format, test (29 files/162 tests), and build gates all passed. The focused remediation suite also passed 31 tests. Following the Docker CI regression, its focused helper/authority regression set passed 24 tests.

Remaining limitation and next step

Do not claim EIG operational superiority or treat an uncommissioned topology's planning suppression as a localization failure. The branch includes only the design for network commissioning in NETWORK_COMMISSIONING_DESIGN.md.

The next model work, if authorized, is a separate causal-prefix training experiment—not a capacity increase or architecture rewrite.