How NeerSetu ranks, and where it stops

• bhopal_ghats_2023_2025.csv — 1,524 rows, one parameter reading each, 7 sites, 2023–2025, Ganesh and Durga, pre / during / one week after. The only input to rankings.
• sample_bhopal_2025.csv — 453 rows; verified identical to the primary file's 2025 rows and never added to totals.
• mp_immersion_sites_2023_2025.csv — 21,600 statewide rows; background only, excluded (no site standardisation; 879 rows flagged needs_check; some tables duplicated across pages).
• cpcb_water_quality_criteria.csv — 19 criteria rows, classes A–E.

Source: organiser pack Vishal4742/bhopal-impact-lab-2026-problem-statements at commit 6b96823f27. Reports: 2023; 2024; 2025. Full hashes and anomaly checks are in docs/DATA_AUDIT.md.

Immersion ghats are places where people enter the water, so CPCB Class B (outdoor bathing, organised) is a reasonable comparison. It is an explicit choice made by this tool — the ghats have not been officially designated Class B here, and a reading “outside Class B screening” is not a legal violation.

Class B has four criteria. The ghat data contains three of them (DO, BOD, pH). Total coliform was not measured, so no site can be described as meeting Class B in full.

Comparisons are inclusive: DO 5.0 is within, 4.9 outside; BOD 3.0 within, 3.1 outside; pH 6.5 and 8.5 within, 6.49 and 8.51 outside. The immersion parameter BOD is mapped explicitly to BOD (5 days 20C); nothing is matched loosely. COD, solids, turbidity, conductivity and the four metals have no Class B limit in this file and are shown as contextual readings — the Class E conductivity limit is not applied.

• A printed 0 is a real numeric zero (DO = 0 appears 4 times, all at Bairagarh).
• Blank, _, -, *, ** → not sampled. Never 0, never clean.
• BDL (below detection limit) and ND → censored: not zero, not missing, and not compared because no detection limit is published.
• A site/festival/year with no table in the pack → not in source. Shown as a gap; no rows are invented.
• None of these contribute a severity of zero. They contribute nothing to the score and are counted in coverage.

Status counts: numeric 1289 · below_detection_limit 174 · not_sampled 61; not-in-source windows: 2.

For each numeric DO, BOD or pH reading, a bounded magnitude outside the screening level:

DO severity  = clamp((5 − DO) / 5, 0, 1)
BOD severity = clamp((BOD − 3) / 3, 0, 1)
pH severity  = clamp(max(6.5 − pH, pH − 8.5, 0) / 2, 0, 1)

These are not probabilities, toxicity or health risk. Readings are grouped into windows = site × year × festival (Ganesh and Durga are never merged). Per window: mean severity, fraction outside, and whether any reading is outside. Windows are averaged with recency weights 2023 = 1, 2024 = 1.25, 2025 = 1.5:

concernIndex = 100 × ( 0.45 × weighted mean severity
                     + 0.35 × weighted fraction outside
                     + 0.2 × windows with ≥1 outside / sampled windows )

A site is scored only with ≥3 numeric readings across ≥2 of the three parameters in the selected scope; otherwise it is insufficient evidence, unranked, and never shown as low. Display bands: ≥50 elevated, ≥20 moderate, below that “lower observed” (never “safe”). Ties: more numeric readings first, then name. Filters recompute everything from matching rows; the parameter view never changes the score.

Worked example, all years — Bairagarh (Sant Hirdaram) kund: mean severity 0.475, fraction outside 0.625, recurrence 5/5 ⇒ index 63.3.

Data note: BOD exceeds 3 mg/l in 100 of 112 numeric readings across all sites, so every scored site has a high fraction-outside component. DO severity and recurrence are what separate them. The index is for ordering attention, not for comparing to any official scale.

Two transparent ratios instead of one confidence percentage:

• Numeric DO/BOD/pH = numeric readings ÷ DO/BOD/pH rows that exist in the source (including not sampled and BDL).
• Festival tables = site × year × festival windows present ÷ nominal windows in scope (3 years × 2 festivals = 6 for the full range). The nominal frame is an analytical reference, not an official sampling mandate.

A phase grid shows each pre / during / after slot as full, partial, not sampled, or no table. Site identity confidence is a separate label.

1. Urgent re-test — any numeric DO ≤ 2.5 mg/l (including 0) or BOD ≥ 9 mg/l in scope. Heuristic triggers, not official remediation thresholds.
2. Priority re-test — any screening exceedance, or the latest year in scope is missing a festival table, has only during-immersion readings, or has no numeric reading.
3. Routine verification — otherwise. Still not a safety or approval determination.

A separate identity flag is raised for medium/low name confidence and coexists with any priority — an ambiguous name never lowers urgency. No action is dispatched automatically.

source_page was tested against the bundled PDFs: it is the 1-based PDF viewer page (the reports print no separate page numbers). An automated text check found the printed parameter line on the cited page for 508 of 508 parameter lines (1524 rows). Only those rows get a #page=N link; anything else is labelled “not independently verified”. 11 tables covering every year × festival were additionally checked by eye against rendered pages (see the data audit).

A text match proves the values occur on that page; site attribution relies on the organiser's table parsing plus the visual checks.

The pack maps varied printed names to seven standard names. Rani Kamlapati ghat / Kamla Park kund is medium confidence: Ganesh tables print a Kamla Park kund, Durga tables print a Rani Kamlapati ghat on the Lower Lake. They may be two places. NeerSetu keeps the pack's single record but flags it everywhere, and on the map shows both candidate places joined by a dashed line.

The MPPCB reports and the CSV contain no coordinates. The map uses OpenStreetMap features found by name and the lake printed in the CSV, and labels every marker with how it was placed:

• OSM feature with the ghat's name — OSM feature whose name matches the ghat (centroid). Closest available, still unverified on the ground.
• Locality named on the report — OSM locality named in the printed site label. The kund is somewhere in this locality.
• Lake only — ghat point unknown — Centroid of the water body named in the CSV. Only the lake is known, not the ghat.
• Name match only — weak — OSM place with a similar name. Weak: may not be the right place.

• Bairagarh (Sant Hirdaram) kund: Bairagarh (OSM suburb) (OSM node/3748732480)
• Prempura ghat: “prempura ghat” (OSM park) (OSM way/922790998)
• Khatlapura ghat: Lower Lake (water body only) (OSM way/28672430)
• Rani Kamlapati ghat / Kamla Park kund: Ganesh tables: Kamla Park kund — near OSM “Rani Kamlapati Mahal” (OSM way/1002690457); Durga tables: Rani Kamlapati ghat — Lower Lake (water body only) (OSM way/28672430)
• Hathaikheda dam / kund: “Hataikheda Dam” (OSM water body) (OSM way/28826931)
• Malikhedi kund: “Mali Khedi” (OSM village) — name match only (OSM node/11358718540)
• Shahpura lake ghat: Shahpura Lake (water body only) (OSM way/28672561)

Markers sharing one lake centre are fanned out slightly so both can be clicked; that offset is cosmetic. India outline: Natural Earth, India point-of-view edition, simplified for display — not an authoritative boundary. Lakes and anchors © OpenStreetMap contributors (ODbL), queried 11 Oct 2026; raw responses are in source-data/geo/. Confirm the physical point before dispatching a team.

The home page reads today's date in IST (or ?asOf=YYYY-MM-DD for rehearsal) and finds the visarjan being prepared for in src/lib/calendar.ts — currently only dates with a citable source (Durga visarjan 20 Oct 2026, from the organiser guide). The ranking is then recomputed on that festival's past seasons only, and the action list is dated around MPPCB's own design: pre-immersion baseline before the day, during-immersion on the day, and one week after. Steps whose window has passed are shown as closed. Nothing is forecast: the plan says where history argues for testing first, not what the water is like now.

Every measurement, comparison and citation in a brief is assembled deterministically from the selected rows. If an AI provider is configured on the server, it may only rewrite the purpose sentence and the action wording. Its output must be strict JSON; every number must already be in the facts; evidence and site IDs must be on the allow-list; words implying safety, causation, legal status or health effects are rejected. Any failure, timeout or missing key falls back to the template, which is always complete. Briefs carry an unticked “Reviewed by officer” marker.

Hindi text is a team translation and should be reviewed by a native-speaking officer before circulation.

The Khatlapura past-readings starter is labelled “Source answer”: its measurements and citations are assembled directly from historical rows, without AI. Follow-up questions can use the configured AI provider. Both paths use the same evidence checks.

The assistant explains selected historical observations and the general procedures below in English or Hinglish. These procedures are team-authored planning guidance, not MPPCB findings, official orders or engineered treatment plans. Each substantive reply links its observations or guidance. Numeric claims must be present in the cited source. Current conditions and permissions still require local confirmation. Replies do not replace an officer or technical review.

Recent messages are sent to the configured AI provider only when a question is submitted. The conversation stays in browser memory, is not saved by NeerSetu, and is cleared by starting a new conversation or reloading. Do not share personal details or secrets. Provider retention follows the provider's own policy. If AI is unavailable or a reply fails checks, an explicit error is shown rather than a canned answer presented as AI.

• Current (2026) water quality — the latest reading is from 2025.
• Whether a site is authorised, open or banned for immersion in 2026 — status is “unverified” everywhere.
• Whether immersion caused a change — before/after differences have many possible causes, and many readings were already high before immersion.
• Health effects, fish kills or toxicity — not measured.
• Units — the reports print none; standard units are assumed.
• Coliform — not measured, so Class B compliance cannot be established.

Back to the ranking