cjc0013/ouroboros-ooi-ctd-benchmark

Ouroboros OOI CTD benchmark

Ouroboros autonomously converted a bounded slice of the NSF-funded Ocean Observatories Initiative Gold Copy into a compact, AI-ready, provenance-graph-backed benchmark on a $500 laptop.

This release covers one CE02SHSM CTD stream across the selected deployment window. It deliberately favors auditability over scale: 70,106 delivered native rows, 17,528 deployment-bounded hourly rows, 509,920 forecast rows, 337 separately labeled Ouroboros quality events, and 668,016 provenance edges. Scale is reported but is not used as a quality score.

Why it is different

• Native OOI observations and delivered QC remain intact beside the hourly derivation.
• Deployment, serial, calibration, and source-file boundaries are explicit; no interpolation or averaging crosses a hardware swap.
• Each forecast target links into a graph that can be traced to its hourly observation, deployment, calibration, source file, and source row.
• Legacy OOI bitmap QC and categorical QARTOD flags use different, semantically correct reducers. Per-hour QARTOD counts preserve frequency information.
• Partial hours remain visible but cannot enter benchmark inputs or targets.
• Direct OOI salinity and density means remain the targets; GSW 3.6.23 recomputations and direct-minus-recomputed deltas expose nonlinear aggregation effects.
• Every input and output is SHA-256 manifested; source-byte and altered-output falsifiers are included.
• The source-acquisition, coverage-scan, calibration-resolution, build, verification, package-verification, and forecasting-audit code paths are shipped with their receipts.

Benchmark

Inputs are fixed 24-hour and 168-hour windows. Forecast horizons are 1, 6, and 24 hours for five named targets: sea_water_temperature, sea_water_practical_salinity, sea_water_pressure, sea_water_electrical_conductivity, sea_water_density. Splits are chronological. Windows crossing a split, deployment, serial, missing hour, or partial hour are absent. Normalization is fit without test leakage.

The benchmark_windows config exposes train (358,530), validation (75,680), and test (75,710) splits. A row stores the input start/end references rather than duplicating the input array; join those timestamps to observations_hourly for a lossless input sequence.

Forecasting value and protocol audit

The fully specified evaluator shipped here is a post-predecessor exploratory analysis, not a continuation of the earlier Ridge protocol. On the corrected data it has 28 positive test point estimates, 24 supported wins, 0 supported losses, and 6 inconclusive cells. Its median improvement over the validation-selected comparator is 5.79%. That is the exact claim; it is not a claim of beating the best test-set baseline.

The feature family and RMSE block-bootstrap statistic were chosen after earlier results were visible. Execution keeps test rows out of alpha selection, comparator selection, fitting, and refitting, but the design choices cannot be proven validation-only. They are labeled post-hoc exploratory and frozen prospectively by pipeline/forecasting_protocol_spec.json; the evaluator refuses to run if either its source hash or the protocol hash changes. The validation-selected comparator is also the best test-set RMSE baseline in 30 of 30 cells. The earlier 29/30 diagnostic used test MAE, which is not the frozen current estimand. The test-best RMSE comparison is reported separately as a descriptive sensitivity analysis: 28 positive cells, 24 supported wins, and 0 supported losses. It is not used for selection or a confirmatory claim.

The same code was run on the frozen 510,040-row predecessor benchmark and produced 28 positive point estimates, 24 supported wins, and 0 supported losses. Removing 120 rows changed no selected alpha, comparator, or bootstrap outcome; median improvement moved from 5.7892% to 5.7908%.

The frozen method was also applied to a newly acquired, immediately preceding 24-month cohort. That time-disjoint historical replication produced 22 positive point estimates, 11 bootstrap-supported wins, 0 supported losses, and 19 inconclusive cells. The preregistration predates every acquisition receipt and is hash-bound through scan and acquisition, so this cohort is labeled a confirmatory historical replication. It shares deployment 18, serial 16-7243, and calibration identity with the current cohort; it is not described as fully independent. See HISTORICAL_REPLICATION_ADDENDUM.md.

The archived predecessor claim was 22 positive cells, 19 bootstrap-supported wins, and 2 supported losses. Its 23-feature model source was not retained, and 14 reconstruction candidates failed exact replay, so that protocol cannot honestly be rerun on the corrected data. Those archived numbers remain visible but are not presented as worse than, or improved by, the current 28/24/0 result: the methods differ. The headline difference is a protocol change, not an effect of the 120-row correction. GSW-derived columns are not model inputs. See evidence/protocol_reconciliation.json and evidence/forecasting_protocol_leakage_audit.json.

The independent release verifier checked 509,920 benchmark rows, replayed 192,808 hourly QC cells, replayed GSW for 17,528 hours, and passed both corruption falsifiers.

Start here

• QUICKSTART.md: load all six configs, run a baseline, and trace provenance.
• SCHEMAS.md: every Parquet column, type, and meaning.
• DATASHEET.md: purpose, scope, exclusions, limitations, intended and unsuitable uses.
• COMPARISON.md: capability comparison with raw OOI and common benchmark packaging.
• VALUE_REPORT.md: forecasting and nonlinear-aggregation evidence.
• HISTORICAL_REPLICATION_ADDENDUM.md: the preregistered time-disjoint replication, boundary correction, results, and frozen identities.
• CHANGELOG.md: the in-place correction record.
• pipeline/forecasting_protocol_spec.json: the hash-frozen exploratory forecasting method and prospective comparability rule.
• pipeline/: current build/evaluation/package code plus the source-acquisition chain and package verifier.
• evidence/: source acquisition, build, validation, baseline, normalization, verification, falsifier, trace, archived protocol, reconciliation, leakage audit, current exploratory value, and historical replication receipts.

Citation and acknowledgment

Data source: Ocean Observatories Initiative (OOI), NSF-funded OOI Data Portal / Gold Copy. Cite OOI and the original source URLs recorded in evidence/input_manifest.json. This derived package is released under CC BY 4.0 and retains the OOI/NSF acknowledgment and provenance required for reuse.