Real recorded data Synthetic tests labelled inside

Klein's Pollution Solution — Benchmark Report

Project Promised Land · Detection Instrument v5.0

Paul Samuel Guarino · Independent Researcher · East Northport, New York [email protected]

Instrument locked: 9 September 2026 SHA-256 root fingerprint: d75583934b6947a3d42983193eda07e544af5d6b4d0a785dacca87f581557204 Patent status: U.S. Non-Provisional Application No. 19/700,318 — pending Report date: 15 September 2026

1. What the system is

A deterministic five-layer detection engine for contamination and surveillance time series. It reads a stream of measurements and identifies the point at which the signal's behaviour changes — in several cases well before the readings cross a regulatory action limit.

It does not diagnose, does not declare an emergency, and does not set policy. It says: something changed here, investigate now.

Three properties define it:

• Deterministic. No neural network, no training, no weights, no random seed in the detection path. Identical input produces identical output on every run, by anyone, on any machine. Every alert traces to the rule that fired it.
• Air-gapped capable. No cloud, no GPU, no external service. It runs in a browser on ordinary hardware.
• Sealed. Instrument, scoring code, reference data and measurement record are hash-frozen under the root fingerprint above, with per-file SHA-256 in LOCKED_v5.0_2026-09-09/MANIFEST.txt. Any evaluation is provably against the same artefact.

The five layers

Layers 2–5 operate on readings below the action limit. They are the early-warning function. Underneath, the engine runs a 256-point Hann-windowed spectral transform on every sample, reading three defined frequency bands and the ratio between them, plus phase-locking measurement against a reference waveform.

A parallel spatial engine separates a localised source from a regional drift using grid binning, neighbour-relative z-scoring, and Moran's I spatial autocorrelation.

2. Portfolio result

22 tests · 16,746 records · zero false negatives in every test ever run.

14 datasets (5,546 records) and 8 synthetic stress series (11,200 records), measured 3 September 2026 against the locked build. Full per-test detail: BENCHMARK_v5.0_RESULTS_2026-09-03.json.

Two columns are reported throughout. Tier 1 is the action-limit guard: the alarm sounds on every record at or above the regulatory limit, and never sounds spuriously. The all-tiers column adds the four sub-threshold layers, which fire below the action level — that is where the early-warning function lives.

3. Flagship result — New Delhi PM2.5

27,820 consecutive hourly records. Station 8118, New Delhi. The largest continuous series in the portfolio and the reference measurement behind the preregistered blind protocol.

Severe onset is defined by the locked scorer as a 24-hour rolling mean reaching ≥ 250 µg/m³ after at least 24 hours below it — episode_analysis(dets, gt=250.0, pre_h=48, post_h=48) in klein_port_v2.py, code unchanged.

Episode detection — Air Mode, all tiers

Sample-level, ground truth = value ≥ 250 µg/m³

False alarms reduced 91.1% with recall unchanged at 1.000. The 2,115 Tier-1 violations independently match a manual recount of rows ≥ 250 (7.6% of 27,820).

Air Mode's baseline is a trailing 720-hour (30-day) causal window — it uses only data already seen, never future data, and is operational after roughly 30 days of prior record. The entire series was fed through the instrument's own loadScenario() → ingest() path in a real browser: 27,820 hourly records in 84 seconds.

The early-warning layers, measured without the threshold

The engine's detection does not depend on the action-limit rule. On Yap Island dengue 2011 (185 daily observations, 978 cases), with the absolute-limit layer switched off entirely — ABSOLUTE_LIMIT = 0, Layers 2–5 only — the non-threshold layers alone first flagged the epidemic on 16 September 2011: 70 days before the peak, at 7 cumulative cases, with 99.3% of the outbreak still in the future.

With the threshold layer active, the same series is flagged 85 days before peak.

On Delhi, the phase-lock layer (Tier 4) isolated on its own warns 28 of 29 onsets at a 44-hour median lead — the longest lead of any layer in the instrument, and is what carries the all-tiers lead to 46 hours in legacy mode.

The blind protocol therefore runs a threshold-off arm as standard, so an evaluator can measure the early-warning layers independently of the regulatory limit.

4. Environmental contamination datasets

These are the results closest to soil, groundwater and site-remediation practice. Every one is real recorded regulatory or monitoring data.

Zero false positives on these four in both the Tier-1 and the all-tiers column — the sub-threshold layers did not manufacture alarms on real environmental records.

Sensitivity sweep. Flint, Bethpage and East Palestine were re-run at σ = 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 4.0 and 5.0. F1 = 1.000 at every setting. The result is not the product of a tuned sensitivity.

Detection latency: immediate on every event in all four datasets — 3,521 of 3,521 events flagged on the sample they occurred, average latency 0.0.

Sources

5. Spatial engine — localised source vs regional background

Nurdle Patrol, Texas coast. 4,254 GPS-tagged citizen-science beach surveys of plastic pellet counts. NCEI Accession 0259486, DOI 10.25921/ccz7-mq57.

Run through the instrument's own runSpatialAnalysis(): 10×10 grid, hotspot threshold z > 2, neighbour radius 1. No labels, no ground truth, no tuning — an unsupervised structural test.

One stretch of coastline averages 142.5 pellets per survey while the adjacent stretches average 32.9 and 33.8 — a 4.3× elevation, 4.55σ above its own geographic neighbours, on 2,288 surveys against 836 and 1,033.

Moran's I is ≈ 0 across every cell (−0.003 to +0.003). There is no smooth regional gradient. The elevation is localised, not diffuse — the signature of a point source rather than a basin-wide background — and it lands on the stretch of coast carrying the pellet-handling and petrochemical infrastructure.

This is the local-source-against-regional-background discrimination the spatial layer exists to perform. It is an unsupervised structural result; there is no ground truth for where pellets were actually released, so it is not scored as TP/FP.

Data note: 31 of the 4,254 rows carried sign-flipped positive longitudes, placing Gulf Coast surveys on the far side of the world and distorting the grid. Corrected before analysis; the original file is untouched.

6. Adversarial stress tests

Eight synthetic series (11,200 records) built specifically to defeat the engine — borderline oscillation, slow ramps, heavy noise, poisoned baselines, and a 10,000-record throughput run.

Recall is 1.000 in all eight. Nothing was missed anywhere, under any attack.

Two results are worth naming.

The poisoned-baseline test. The engine's first 30 readings are themselves contaminated — the failure mode that defeats any system that learns "normal" from what it sees first. Perfect score: the contamination is flagged rather than absorbed into the baseline.

The gradual ramp. The classic slow-onset case a fixed threshold misses entirely, because no single reading is alarming. Caught at recall 1.000.

Every figure above is displayed by the instrument itself on screen, not only in this report.

7. Method and scoring

Applied identically across every dataset:

• Ground truth positive: any record at or above the defined action limit.
• True positive: a Tier-1 violation on such a record.
• False negative: such a record with no Tier-1 violation.
• Sub-threshold detections from Layers 2–5 are reported separately in the all-tiers column and are not counted as Tier-1 false positives. They fire below the action level, which is what they exist to do.
• Per-dataset input: the action limit, set to that contaminant's real-world regulatory level. It is the only per-dataset input. Every other parameter is fixed and identical across every run in this report.
• Execution: each dataset fed through the instrument's own loadScenario() → ingest() path in a real browser — not through a separate analysis script — and scored from the instrument's own detection log by its own runBenchmark() logic.

8. Provenance and reproduction

LOCKED_v5.0_2026-09-09/ contains, hash-frozen:

To verify:

Get-FileHash -Algorithm SHA256 "software\Klein_Pollution_Solution_DEFINITIVE_v5.0__INSTRUMENT_OF_RECORD.html"

shasum -a 256 software/Klein_Pollution_Solution_DEFINITIVE_v5.0__INSTRUMENT_OF_RECORD.html

Compare against MANIFEST.txt. To confirm the root, hash the sorted list of the manifest's <hash> <path> lines.

Operating notes. Font Awesome, KaTeX and Google Fonts load from CDNs for presentation only; the detection path never calls them, and run fully offline the detection output is identical. An optional live-API input mode accepts a monitoring endpoint when a user configures one, and plays no part in replaying a file.

Every figure in this report is reproducible from public data without the inventor's participation.

9. Independent evaluation — the open invitation

A preregistered blind protocol is committed and hash-locked (PREREGISTRATION_Blind_PM25_Test_2026-09-14.md, committed before any evaluation data changes hands).

How it runs:

1. The evaluator selects the data — a series the instrument has never seen, meeting stated minimum criteria for size and completeness. If a first draw fails a criterion, the evaluator redraws under the same rule. The criteria are fixed in advance; the dataset is not.
2. The identity of the data is withheld from the inventor. It arrives as a plain two-column series.
3. The inventor returns the detection log — flagged timestamps and layer, nothing else — in three arms on identical data: legacy mode, Air Mode, and threshold-off (ABSOLUTE_LIMIT = 0, Layers 2–5 only), which directly answers the "a conventional threshold would do the same job" objection.
4. The evaluator scores every arm against ground truth the inventor never sees.

Pass and fail criteria are deliberately excluded from the sealed materials. They are the evaluator's to set. Writing them in advance would defeat the point of an independent test. The inventor commits the claim; the evaluator commits the scoring.

The root fingerprint is sent to the evaluator before their data is received — their copy, their inbox, their timestamp — so the artefact cannot be altered afterwards.

Paul Samuel Guarino · Independent Researcher · East Northport, New York, USA [email protected]

Two systems by this inventor hold Awardable designation on the United States Department of Defense Tradewinds Solutions Marketplace, awarded 2026, following evaluation by Chief Digital and Artificial Intelligence Office assessors. Multiple United States patent applications are filed and pending.