Showcase: From Reading to Verdict (Delta Basin)

This showcase walks every layer of the engine using a real build of the riverflow pack against the Delta Basin tenant (slug: delta). The numbers, signals, verdicts, and validated observations below are not mocked — they come from riverflow_delta_kls.duckdb, the same KLS JinDesk reads at delta.jinflow.io.

Delta Basin is a synthetic but realistically-shaped freshwater sensor-network operation: an RK-class fleet across three watersheds, on firmware ranging from deprecated 1.3.x to current 1.4.x, producing the three-layer reading families a real sensor company delivers — raw signal, calibrated measurement, derived environmental variable. It is deterministic: the whole basin regenerates byte-for-byte from a seed, so every number here reproduces exactly.

The tenant

The Gold contract — the consumption layer signals operate on:

Entity	Rows
Sensors	60
Deployment sites	31
Watersheds	3
Reaches	20
Deployments	48
Raw readings	607,248
Measurements	607,248
Environmental variables	202,416
Bench calibration tests	271
Calibrations	192
Firmware flashes	79
Maintenance events	18
Alarms	435

The medallion beneath (Bronze → Silver → Gold) is invisible to the analytical instruments — signals see only Gold. On the published cloud KLS the raw and silver layers are purged; only Gold travels, and the River page says so honestly rather than pretending the fleet was never built.

Step 1 — Signals detect

The pack ships hand-written and typed signals. On Delta, six of them have findings, plus one perspective that aggregates:

Signal	Findings	What it catches
signal_deprecated_firmware_in_field	31	Sensors still on retired firmware
signal_suspect_reading_rate	11	Sensors self-flagging elevated quality doubt
signal_datastream_inactive	3	Streams that have gone dark
signal_site_fouling_burden	2	Sites carrying persistent fouling load
signal_watershed_fouling_burden	2	Watersheds concentrating the sediment pattern
signal_calibration_drift	1	A bench test drifting beyond tolerance

The suspect_reading_rate signal is worth a note: it does not compute doubt, it reads it. A sensor stamps its own quality_flag at the source — fouled optics, a bad packet — and that flag travels untouched through the medallion. The signal turns the fleet’s self-reported honesty into a diagnostic. Delta carries 694 self-flagged suspect readings, and they cluster exactly where the fouling signals fire.

Step 2 — A perspective aggregates

perspective_sensor_health rolls the sensor-scoped signals — deprecated firmware, inactive datastreams, offline, and suspect-reading-rate — into one health score per sensor (32 findings). A sensor flagged by several independent signals at once is degrading on multiple fronts — a compound pattern no single signal can see. These are the field-visit shortlist.

Step 3 — Theses confirm

Five theses, each a plain business question. Their evidence scores spread honestly:

Thesis	Status	Score
thesis_site_service_burden	confirmed	1.00
thesis_bench_calibration_currency	confirmed	1.00
thesis_firmware_field_discipline	confirmed	1.00
thesis_watershed_data_continuity	confirmed	0.67
thesis_fleet_reliability_risk	plausible	0.60

Note the last row. fleet_reliability_risk is plausible, not confirmed — and so, by the cascade gate, it fires no verdict. The platform will not explain a “why” for a “what” it has not confirmed.

Step 4 — Verdicts explain

Each confirmed thesis walks all the way to a named root cause and a recommendation — across four distinct categories:

Verdict	Thesis	Root cause
verdict_firmware_rollout_stalled	firmware discipline	process_failure — the OTA campaign closed its queue, not the fleet
verdict_flat_schedule_vs_site_load	service burden	structural — a flat calendar against site-shaped fouling
verdict_datastream_gaps_fragment_basin	data continuity	system_failure — inactive streams in the neglected-sensor tail
verdict_recalibration_interval_lag	calibration currency	data_quality — drift correlating with the deprecated tail

Every recommendation folds into one operational answer: a single field visit that flashes the firmware, cleans the fouled optics, and recalibrates the drifting sensor.

Step 5 — Observations ratify

This is the chapter the other showcases do not have. Alongside the diamond, Delta carries field observations — signed, dated notes from the operation’s team:

“After a cold night I stop trusting the first five hours of v3.2 readings.” “The sensors below the cattle crossings are always the first to drift.”

Each enters the record as a hunch (suspected). It becomes validated only when Understanding actually backs it — through a Contributing Factor, a piece of signed reasoning that cites the specific finding. Delta stands at 8 validated, 3 suspected.

The three that stay suspected are deliberate: a farmer’s decade-long memory the data cannot reach, an uncomputed baseline, and one scientist respectfully disputing another’s interpretation. A platform that validated all five would be flattering its users and lying to them. The refusal is the point — and it is what makes the validated ones defensible. See the note “How a field hunch becomes a validated claim” on the Observations page.

The full chain

One reading, walked all the way up:

quality_flag='suspect'              a sensor doubts its own optics
  → gold_measurements               carried untouched through the medallion
  → signal_suspect_reading_rate     11 sensors flagged
  → perspective_sensor_health       compound degradation, 3 high-severity
  → thesis_firmware_field_discipline confirmed (1.00)
  → verdict_firmware_rollout_stalled process_failure, named + recommended
  → obs_cattle_crossing_drift        a field hunch, now validated against it

Facts before interpretation, interpretation before explanation, explanation before narrative — and every step clickable back to the reading underneath.

What this showcase teaches

• The graphic is navigation. The overview tiles are not decoration; click a severity band and you land on the filtered list.
• The narrative has to earn itself. No verdict without a confirmed thesis; no validated observation without real evidence.
• Defensibility by construction. Every finding walks back to a reading that already carries its own quality flag — the answer to the reviewer’s “how do you know?” is attached, not asserted.
• Everything is declared. Signals, theses, verdicts, tiles, and whole pages — including Delta’s Calibration and Fleet pages — are YAML the pipeline compiles, in four languages.

Compare

• Showcase: From Case to Verdict (Inspire) — the numetrix hospital pack.
• Showcase: From Person to Verdict (Val d’Oria) — the hrcentral workforce pack.