hydro-look

An open pipeline for the state of Ecuador's hydro fleet: reservoir levels and inflows, per-plant generation, and the national energy balance — collected daily from the public CELEC and CENACE endpoints, versioned in this repository, and published as a static site.

Not an official source. Every number here is a copy of what CELEC or CENACE published, with the response that produced it archived alongside it.

Status

Phase	State
0 · Reconnaissance and fixtures	done — scripts/recon/RECON_REPORT.md
1 · Full ingest (CELEC ORDS + CENACE SMEC + Información Operativa)	code complete; levels backfilled to 2014-09-20, Mazar's historian inflow to 2010-02-10 (6,062 days — its true first reading, found by a walk that stopped on twelve empty months rather than on a --from), CELEC Sur energy to 2015-11-01, hourly energy for Coca Codo Sinclair (2016-06-24→), Agoyán (2017-01-01→) and Manduriacu (2017-07-31→). What remains is the acceptance criterion three consecutive green scheduled daily runs. The crons moved to odd minutes on 2026-09-23 (12:47 and 16:53 UTC), because GitHub delays busy ones, and freshness.yml now reports a slot that never fires. Before the move the count had not started: the 2026-09-22 12:15 UTC slot had still not fired at 13:07, with nothing queued, although daily.yml had been on main since 01:26. GitHub's scheduler is the reason, and this repository's only scheduled run to date started 5 h 22 min after its slot, so the close date follows the first firing that actually happens rather than the calendar. Closed 2026-09-24: the third consecutive green scheduled run, and every scheduled run since (21 to 2026-10-03) has been green
2 · CENACE history and reconciliation	SMEC backfilled 2016-05-01 → 2026-09-20 (3,780 days, 0.40% missing) and reconciled against the ORDS per-plant energy. The fifteen missing days were re-asked on 2026-09-22 and none recovered, so 0.40% is this source's floor. Outstanding: the Información Operativa cross-check needs ≥ 20 snapshot days and has 1
3 · Additional reservoir levels	done 2026-09-22 — 21,541 historian rows: daily level and inflow for Coca Codo Sinclair (2016-03-07→), Agoyán (2016-07-05→) and Manduriacu (2017-08-01→). The Mazar control month matches repDiaHid12m on all 31 days to 0.0000 m, and the caudal semantics are now settled on 4,281 days rather than assumed: mridCaud is inflow, not turbined flow (data/crosschecks/caudal-semantics.md)
4 · Covariates and quality	reference tables (plants, thresholds, rationing_episodes) and the npm run check gates done, public/api/status.json published; ONI 1950-01 → 2026-07 ingested, and ERA5 1990-01-01 → present at all eight basin rows (backfilled at the seven verified centroids 2026-09-24). Catchments delineated 2026-09-23: npm run catchments routes flow over the Copernicus GLO-90 DEM above each of the seven dams and checks every catchment against INAMHI's per-scheme polygons (geonode:hidroelectricasshape) — six agree at 99–100% IoU, and Mazar at 91%, where INAMHI leaves out a north-bank lobe beside the dam. basins.csv now carries a verified centroid per catchment (data/reports/catchments.md). Done 2026-09-24. See PLAN.md §2.4
5 · Modelling v1	done 2026-09-22 — public/api/forecast.json carries p10/p50/p90 Mazar level at 7/14/30/60/90 days and days-to-threshold under three named analogue years, with the whole censored crossing distribution beside them. The shipped model is a water balance closed around the operator: the reservoir's area-elevation curve and its turbine's m³/s-per-MW are fitted from this repository's own readings, and release is a rule curve read back off the level every simulated day. Over 105 monthly origins from 2018-01 it is 24.5% better than persistence at 60 days and 34.8% at 90, and indistinguishable from it under a month, which the document says rather than hides. §7's climatological-drift rung loses at every horizon and the open-loop water balance §7 specified loses by 69% at 90 days; both are kept in data/reports/backtest.md as recorded negatives. The crisis check is the unflattering one: of Mazar's two 2024 spells below 2115, the P50 called neither in advance, though the ensemble's dry tail put the October crossing 8.5 days out against an actual 7
6 · Site and JSON API	built and deployed 2026-09-22 — public/api/latest.json joins status.json, forecast.json and adequacy.json, and the Spanish page is a Next.js static export rendered from data/curated at build time: reservoir gauges against their declared bands, Mazar's forecast fan over six months of recorded cota, a year of inflow against its own climatology, six months of the national mix, the adequacy tile, feed freshness, downloads and the method notes. It needs no JavaScript to read — the only scripts are the Next runtime and the analytics beacon — and every chart is inline SVG from tested pure functions. The adequacy tile is now built and backed by a model (Phase 6c). The 6b narrative panel is live (first ok narrative 2026-09-23; seven consecutive daily narratives by 2026-09-29, the phase's acceptance criterion). Since 2026-09-23 there is a page per reservoir (/embalses/<site>/), a daily permalink (/dia/<date>/), embeddable cards, an Atom feed, a documented public contract with JSON Schemas (/datos/, public/api/schema/), and a scorecard that reads the published forecasts back. The site is deployed on Vercel, connected to this repository
6c · Energy adequacy (§7 target 3)	done 2026-09-22 — public/api/adequacy.json publishes expected deficit in GWh/day at 7/14/30/60/90 days and a four-level risk tier, from one identity: unsuppressed demand − hydro − thermal − imports − other. Demand is served load, not demanda_distribucion (which misses 3–11 GWh/day of losses and unregulated consumers — more than the whole Colombian interconnection), and it is fitted excluding the rationing episodes, because measured load during a cut is the load that was allowed. Over 99 monthly origins the net requirement beats a trailing 28-day mean by 12.4% at 7 days and 11.4% at 90; band coverage is 60–67% against a nominal 80% and is published as such. §7's inflow→hydro link is a recorded negative: r = 0.47 on 30-day means, GWh per m³/s drifting 83% across the record, and five rungs built on it all losing to persistence. The check that makes it publishable is the 2024 episode, where measured suppression (20.3 GWh/day) and computed deficit (17.0) come from different sides of the identity and agree within 3.3; the two short episodes do not agree, and the page says so. Applied to every month of the record, the tiers flagged 3 of 99 origins and all 3 preceded cuts, while 6 of the 9 origins that preceded cuts went unflagged — it does not cry wolf and it misses most of the wolves. Building it also found that the SMEC completeness gate is one-sided — see data/reports/adequacy.md
7 · Extensions	XM's side of the Colombian interconnection is ingested (xm_exchange_daily, xm_system_daily, 2016-05-01 →); an export-availability model over it was built and backtested and is a recorded negative (data/reports/adequacy.md). The candidate list in ENHANCEMENTS.md was implemented 2026-09-23 — raw archive as write-once NDJSON day files, per-source failure isolation and alerting, the public data contract, the site's new pages and accessibility, the scorecard, adequacy v2 with its import sensitivity, inflow forecasts for four plants, and the tooling upgrades; its head says item by item what is done. The ERA5 backfill at the seven centroids ran 2026-09-24, and on 2026-10-03 the seven-day Mazar point moved to M4-gbm-direct, the design that wins on the verified centroid (PLAN.md Phase 4). On 2026-10-13 M3's analogue years began starting from today's river (scaleToState in water-balance.ts): each year's inflow is multiplied by today's trailing-week inflow over its own, fading with a 15-day half-life, because a drawdown at 25 m³/s was being forecast as a refill from the 70–80 m³/s most Octobers began with. MAE falls at 7–60 days (by 0.84 m, 14%, at 30 days, with the paired interval below zero), ties at 90 (+0.10 m, interval across zero), the 14–30-day optimism halves, and the P50 now calls the October 2024 crossing a week out; data/reports/backtest.md reruns the comparison daily. A season-aware release rule was tried the same day and is a recorded negative. Against that M3, M4-gbm-direct's seven-day lead shrinks to 0.21 m with an interval across zero and a band no nearer 80%, so seven days falls back to M3 by the ladder's own rule (data/reports/m4-backtest.json, refitted the same day). Streamflow return periods from INAMHI–GEOGLOWS, 2026-09-25: npm run geoglows matches each dam to its GEOGLOWS v2 river (all eight, Amaluza included, within ±1.4% of the delineated catchment) and publishes the 2- to 100-year flood flows INAMHI's Hydroviewer draws — reproduced from the model and checked against the portal within 0.1 m³/s — beside the same fit on CELEC's record, in latest.json and on each reservoir page. GEOGLOWS' forecast and INAMHI's corrected hydropower forecasts were both backtested against CELEC's inflow and neither beats persistence on its own, so neither is published as it stands; used inside this repository's inflow forecast instead, the analogue + climatology ensemble they were tested against became the published rung (6 → 11 of 12 plant-horizons), and GEOGLOWS' forecast joined it as a third member at Agoyán, Manduriacu and Minas San Francisco (all 12 now publish; data/reports/geoglows-experiment.md) (data/reports/return-periods.md, , ; §2.8 and Phase 7)

Where the data comes from

Source	What it gives	History
CELEC ORDS repDiaHid12m	daily level, inflow and declared band for Mazar, Amaluza, Minas San Francisco, Delsitanisagua — 365 days per request	2014-09-20
CELEC ORDS repDiaEner12m	daily energy per CELEC Sur plant, 182 days per request	2020-09-20
CELEC ORDS repDia* (one day each)	level/inflow, power, turbined flow, units online, spill, plant factor, day-ahead plan, SNI total	per day, 2016 →
CELEC ORDS {code}EnerDia	24 hourly values per plant-day, for all seven dashboard plants	2019-06 →
CENACE SMEC ResultadoInforme1.do	the closed day's national balance in kWh by generation type, imports, exports, distribution demand	2016-05-01 →
CELEC ORDS pointValuesMesH24	daily level and inflow per mrid, a month per request — the only route to Coca Codo Sinclair, Agoyán and Manduriacu, and the deepest route of all for Mazar	2010-02-10 → (Mazar inflow); 2016-03-07 → for the three
CENACE Información Operativa	live production, demand by distribution utility, last validated day	snapshot
INAMHI–GEOGLOWS return periods (inamhi.geoglows.org)	2- to 100-year flood flows at the river each dam stands on, as INAMHI's Hydroviewer draws them: a Gumbel fit on GEOGLOWS v2's daily simulation from 1980, computed here from the public geoglows-v2 bucket by npm run geoglows and checked against the portal's own charts (all eight rivers within 0.1 m³/s). CC BY 4.0, like the rest of the model; the store's own 1940 fits, carried as daily and hourly, are CC BY-NC-SA 4.0	reference, store revision 2026-06-10
GEOGLOWS v2 river forecast (geoglows-v2-forecasts)	Each 00 UTC issue's high-resolution member, as daily means at leads 1–10, for the rivers of Agoyán, Manduriacu and Minas San Francisco: the third member of the inflow forecast at those plants (data/curated/geoglows_forecasts, npm run geoglows:daily). Reported in status.json as geoglows_forecast. CC BY 4.0, with ECMWF's copyright, source and disclaimer, which attribution carries ()

Six things worth knowing before using any of it:

• repDiaNivQIng answers with the previous day's numbers. Asked for date D it returns rows stamped D whose level and inflow are D−1's, identical to what repDiaHid12m publishes for D−1 — measured on 113 consecutive days and on captures from 2016, 2019, 2022, 2024 and 2026. Rows from it are therefore stored under the day they describe, not the day they are stamped. repDiaPotQTurb is shifted the same way: it publishes Molino's turbined flow, digit for digit, as repDiaNivQIng's Sopladora inflow for the same requested date (116 of 116 days, and every capture), and both are midnight snapshots rather than daily figures. The other per-day reports are measured and not shifted; DATA_DATE_OFFSET_DAYS in src/lib/registry.ts has the evidence for each.
• The historian's flow mrids are inflow, and carry more precision than the reports. mridCaud (30538 for Mazar) is q_ingresado, agreeing with repDiaHid12m on 4,281 days at r = 1.0000; the only difference is that the report rounds to whole m³/s and the historian does not, so the report is exactly round(historian) on every one of those days. Turbined flow, the other candidate, correlates at r = −0.06.
• A zero inflow means opposite things on the two routes, and neither is "the river stopped". The historian publishes decimals, so its 0.00 is the service saying nothing — 82 of them sit below every one of those series' own non-zero floors (84.00 m³/s for Coca Codo Sinclair, 35.00 Agoyán, 10.40 Manduriacu) and cluster like a fault, including the eighteen publishing days before Mazar's first real reading. Those are dropped. The 12-month reports publish whole m³/s, so their 0 is round(x) for any x below 0.5: on 2024-11-08, at the worst of the rationing drought, repDiaHid12m published 0 for Mazar and the historian published 0.142 for the same day. Those are kept. An inflow above 10,000 m³/s is dropped on both routes — the historian published 23,221.10 for Mazar on 2013-11-27, between neighbours of 34.31 and 0.00.
• The SMEC completeness gate is one-sided, and 57 balance days are not days. parse/smec.ts rejects a page served before its metering arrived by requiring distribution demand to be at least 20% of generation, which catches a page whose demand has not landed and passes one whose generation has not: 2018-01-05 carries 4.19 GWh of national generation against 62.72 GWh of demand and sails through, because 62.72/4.19 is 15. Served load is distribution demand plus transmission losses plus unregulated demand, both positive, so a day whose generation plus net imports falls below its own distribution demand is a page caught mid-render — 55 days, about 1.5% of the record; the live count is in , which regenerates daily. The same ratio catches the opposite fault: on 2025-07-17 and -18 CENACE published at 153.62 and 113.19 GWh against a fortnight's median of 2.2, flagged in its own column at +6,284%, lifting national generation to 247 GWh on a 91 GWh day. All 57 are counted by and excluded by ; none is deleted.

Running it

When something breaks, or an issue opens, start with RUNBOOK.md.

npm ci
npm test            # every parser against the recorded Phase 0 responses; no network
npm run typecheck
npm run lint

npm run ingest -- daily                    # the last three days from every source (--days)
npm run ingest -- daily --dry-run          # parse and validate, write nothing
npm run ingest -- daily --out /tmp/batch   # fetch and stage only, as CI does; touches no data
npm run ingest -- apply --in /tmp/batch    # merge a staged batch into data/ (idempotent)
npm run ingest -- backfill --source ords-levels --from 2014-09-20
npm run ingest -- backfill --source smec --from 2016-05-01 --max-requests 10000
npm run ingest -- latest                   # live tiles only
npm run ingest -- smec-earliest            # binary search for SMEC's oldest report
npm run ingest -- covariates               # recent ERA5, 16-day forecast, full ONI series
npm run ingest -- covariates --dry-run     # validate without writing, including with --out
npm run ingest -- covariates --from 1990-01-01 --max-requests 100  # resumable climatology

npm run ingest -- backfill --source ords-historian --from 2016-01-01   # a month per request

npm run forecast                           # backtest the ladder, then write forecast.json and the report
npm run forecast -- --dry-run              # compute and print; touch no file
npm run forecast -- --no-variant           # skip the ENSO comparison (about a third of the runtime)

npm run adequacy                           # backtest, then write adequacy.json and the report
npm run adequacy -- --dry-run              # compute and print; touch no file
npm run adequacy -- --ceilings             # print the demonstrated thermal and import ceilings and stop

npm run check                              # shape, ranges, reference integrity, raw_refs, quarantine
npm run check -- --freshness               # also fail when a feed has stopped arriving
npm run check -- --out public/api/status.json
npx tsx scripts/tls-expiry.ts --live       # certificate expiry and pin drift, as tls-expiry.yml runs it

npm run geoglows                           # INAMHI's return periods at each dam's river, beside CELEC's own fit
npm run geoglows:forecast -- --step 3      # backtest GEOGLOWS' archived forecasts against CELEC inflow
npm run inamhi:hydropower -- --earliest    # backtest INAMHI's hydropower forecasts (runner only: the host refuses the sandbox)
npm run geoglows:experiment                # try both forecasts as members of the inflow forecast
npm run geoglows:daily                     # fetch GEOGLOWS' forecast for the inflow ensemble (the daily model step runs it)
npm run publish:api                        # write public/api/latest.json from the committed tables
npm run publish:api -- --dry-run           # build it and print a summary; touch no file
npm run publish:api -- --restamp           # also rewrite the contract block on the other documents

npm run export:bulk                        # one <table>.csv.gz per curated table, into public/api/bulk/
npm run dev                                # the site, against whatever is in data/ right now
npm run build                              # export:bulk, then the static export, into out/

The site is a build-time render of files already in this repository: next build reads data/curated and public/api, writes out/, and the page needs no JavaScript to read: every number, chart and data table is in the HTML, and the only scripts it ships are Next's runtime and the Vercel Analytics beacon. There is no request path and no server, which is decision 6 made literal — Vercel holds no credentials, runs no ingestion and queries nothing. A number changes on the site when a number changes in this repository, and not otherwise. out/ is servable by anything static, so npx serve out is a faithful preview (without vercel.json's headers).

The pages:

Path	What it answers
/	the day's questions in order — the mix, what changed since yesterday, the day's reading, the eight reservoirs on one scale, Mazar's forecast with its skill beside each horizon, inflow, the national balance, adequacy, feed freshness and the method notes
/embalses/mazar/	Mazar's full record, its analogue years, the 2024 crisis check and its two declared floors
/embalses/<site>/	the other seven reservoirs from one template (generateStaticParams over latest.json): level, bands or their absence, slopes, the whole record and a year of inflow against its climatology
/dia/ and /dia/<fecha>/	what was published about each day, rebuilt from forecast_runs, adequacy_runs and narrative_snapshots, with the observed level beside each forecast whose date has passed
/datos/	the public data: each document's fields and units (read from the JSON Schemas), update times (read from daily.yml), the bulk tables and the stability promise, with schema.org Dataset markup
/embed/<site>/	a fixed 400 × 260 card per reservoir for an <iframe>, without the site's chrome
/feed.xml	the daily reading as Atom, one entry per day the validator passed

Every page has Open Graph and Twitter metadata with a canonical URL, built by pageMetadata in src/lib/site/meta.ts, one <h1>, and its breadcrumb trail as schema.org BreadcrumbList; the preview image is drawn at build time from the day's headline and Mazar's band by next/og, with no network, and every page links to it. robots.txt, sitemap.xml, icon.svg (the masthead's mark) and a Spanish 404 come from src/app/. The three typefaces ("Páramo": Instrument Serif, Geist and Geist Mono over a teal-and-terracotta palette, light and dark from the reader's system setting) are self-hosted from public/fonts/, the Latin subset of each with its SIL Open Font License beside it. The site's absolute URLs come from HYDRO_LOOK_SITE_URL, default https://hydro-look.vercel.app.

Every headline on the page is a rule over the day's numbers in src/lib/site/story.ts, so a sentence cannot outlive the number it describes. Each section of the home page is a component under src/app/components/sections/; every colour is a class in globals.css, not an inline style. Tables are real <table>s with header cells, and every chart carries its plotted values in a collapsed <details> table beneath it. The inflow, mix and cross-section drawings are one scalable SVG each whose labels CSS resizes for a phone; the forecast fan and the twelve-year record change shape on a phone and are still drawn twice. Printing the page gives the numbers on white with each link's address beside it.

The public API

public/api/*.json are the stable public URLs a third party can fetch — with CORS open and a five-minute cache, set in vercel.json — and since schema version 1 each says what can be relied on (src/lib/publish/contract.ts):

• schema_version is the format version. It changes only when a field is removed, renamed or retyped; an added field does not change it. Each document has a JSON Schema under public/api/schema/, and tests/publish.test.ts validates every committed document against its schema, with a small in-repo validator that refuses keywords it does not enforce.
• data_date has one meaning everywhere: the Ecuadorian day the readings describe. In the model documents it equals origin_date. as_of (the day the job ran) stays in latest.json and status.json as a deprecated alias until schema version 2, and each document lists its deprecations in a deprecated block.
• Codes are English slugs with Spanish labels beside them: feeds carry id and label_es, band declarations report_endpoint / dashboard_chart_title / operational_minimum with declaration_es, adequacy tiers tier_code (comfortable, watch, tight, deficit) and tier_label_es beside the model's own Spanish word. adequacy.json's current.narrative_tier_field names the tier the daily narrative is written about (the worst), and the page says so.
• license, attribution and absolute see_also close every document.
• previous and delta_1d in latest.json give each reservoir's level and inflow the day before, and the national balance's, so "since yesterday" needs no CSV.
• inflow.return_periods in latest.json places each reservoir's inflow against two sets of 2- to 100-year flood flows kept side by side: geoglows.inamhi (what INAMHI's Hydroviewer draws for the GEOGLOWS v2 river matched to the dam's catchment; the store's own 1940 fits ride along as daily and hourly) and measured (the same Gumbel fit on CELEC's daily record, from five complete years). They agree at Mazar (1.03×) and Coca Codo Sinclair (0.97×) and differ by 12–139% at the other five dams; geoglows.simulated_vs_measured says how closely the model follows the measured river at all. Sopladora has neither: its inflow is Molino's turbine discharge. data/reports/return-periods.md has the comparison, and PLAN.md §2.8 and Phase 7 the method.

The scripts owned by the models write their documents through the same stamp (publicJson in forecast.ts and adequacy.ts, withContract in narrative.ts and check.ts), and npm run publish:api -- --restamp rewrites the stamp on whatever is already published.

npm run check runs in CI on every push and again after every ingest. The split is deliberate: shape and range checks are a function of the files alone, so they hold for as long as the commit does, while freshness is a function of the clock and would turn every pull request red as the data aged. A feed that has never produced a row is reported rather than failed, so a source that has not had its first run yet does not block the gate.

The sources are unreachable from most sandboxes; ingestion runs in GitHub Actions (daily.yml at 12:47, 16:53, 20:47 and 00:47 UTC, backfill.yml by dispatch). Every backfill is resumable — it skips days already in the store — so a long history is filled by dispatching the same command until it reports no new rows. The schedules sit on odd minutes because GitHub delays scheduled runs on busy ones; the first two runs for the old 12:15 and 16:30 slots started hours late.

Each ingest workflow runs as three jobs. fetch asks the sources and stages a batch (rows as JSON, raw responses as they will be archived) with no write token and no secret, and uploads it as an artifact. apply downloads it and runs .github/scripts/apply-and-push.sh, which resets to the branch tip, applies, regenerates status.json and latest.json, commits and pushes — retrying from a fresh tip if another run pushed first — and it is the only job in the ingest concurrency group, so a backfill's hours of fetching never queue anyone. A run that found nothing commits nothing: when only generated_at stamps changed, the attempt is discarded. report opens one issue per failure mode (ingest-failure, model-failure, covariates-failure, backfill-failure) and closes it when the same workflow next goes green. Every run writes a step summary: rows added and updated per source, requests sent per host (retries included), errors, and for the narrative the call's cost, latency and the spend to date, which status.json also carries. freshness.yml checks main a few hours after each daily slot and keeps one stale-feed issue for a feed that has stopped or a slot that never fired; tls-expiry.yml checks weekly for a certificate within 30 days of expiry or a pinned host whose certificate changed.

Replaying a batch. A batch that failed to land is not lost: download the run's ingest-batch-<run id> artifact (backfill-batch-…, covariates-batch-… and xm-batch-… for the other workflows) and apply it by hand on an up-to-date checkout:

gh run download <run id> -n ingest-batch-<run id> -D /tmp/batch
npm run ingest -- apply --in /tmp/batch     # upsert rows and raw responses into data/
npm run check -- --freshness --out public/api/status.json
npm run publish:api
git add data public && git commit -m "Replay batch from run <run id>"

apply is an upsert, so replaying onto newer data is safe, and replaying twice changes nothing. A batch staged before the raw archive moved to day files applies too: its bundles are refiled and its rows' raw_refs rewritten on the way in.

Partial failure. Each source's errors stay its own: a pinned host whose certificate check fails, a 5xx that outlives its retries, or a 200 that parses to fewer rows than the endpoint's floor in registry.ts is recorded on the batch and fails the run at the end, while every other source is still asked and written. Retries back off with jitter, honour a 429's Retry-After, and stop at one deadline per request. Rows that fail their contract are written to data/quarantine/<table>/<run>.csv with the reason instead of stopping the write, and npm run check fails while any quarantine file exists: fix the parser, re-apply the batch, and delete the file. Raw responses are written before the tables, and every file through a temp file and a rename, so an interrupted run cannot leave a row pointing at a response that was never archived — npm run check verifies that every raw_ref resolves.

Weather and ENSO use their own covariates.yml workflow, scheduled daily at 17:23 UTC, and the same staged apply/write queue as the existing ingestion. A manual dispatch with from fills ERA5 history; without from it refreshes 30 recent available days and the 16-day forecast. ERA5 stops six days before today to allow for the publication delay. Incomplete basin-days are retried on a subsequent history run. The shared request budget includes NOAA; a very small budget may require another run to reach it.

data/reference/basins.csv lists a verified centroid for the catchment above each of the seven dams (paute_mazar, coca_ccs, pastaza_agoyan, guayllabamba_manduriacu, daule_marcel_laniado, jubones_msf, zamora_delsitanisagua), delineated by npm run catchments and checked against INAMHI, plus the provisional Phase 0 paute point, kept because M4 and the narrative were fitted on it. Each is a point sample at the centroid, not basin-average precipitation. Coordinates are preserved in each weather row, but the readers select by basin id, so a row's point should not be moved in place: add a new basin id instead.

Forecast issued_at records our collection time, not the upstream model initialization time; separate collections are retained for later backtests. ERA5 rows have an empty issued_at. ONI's month is the centre of a three-month mean (January means December–February), and the curated series contains NOAA's latest revisions. It is not point-in-time training data: account for publication lag and revisions before using ONI in historical backtests.

Layout

src/lib/parse/      one pure parser per endpoint: raw text in, typed rows out
src/lib/sources/    fetch + archive + parse for each upstream system
src/lib/store/      year-partitioned CSV upsert, the NDJSON raw archive, status and run summaries
src/lib/contracts/  zod table schemas; a drifted response writes nothing
src/lib/quality/    checks over what is on disk, which the row-by-row contracts cannot see
src/lib/features/   series assembly, the fitted reservoir curve, the national balance with its
                    unusable days taken out, ONI read at its true lag
src/lib/models/     the M0–M3 ladder, the rolling-origin backtest, the forecast, the adequacy
                    model, and the report each one writes
src/lib/publish/    the current-state document the site's tiles read, the public data contract
                    every document is stamped with, its schema validator and the bulk export
src/lib/chart/      scales, ticks and SVG path geometry; pure and unit-tested
src/lib/site/       what the page reads at build time, and Spanish formatting
src/app/            the Next.js App Router page and its server-rendered SVG charts
scripts/ingest.ts   the CLI
scripts/check.ts    the quality gates and the public status document
scripts/forecast.ts the backtest and the published forecast
scripts/adequacy.ts the adequacy backtest, adequacy.json and its report
data/curated/       the tables, CSV, partitioned by year
data/raw/           every response as fetched, NDJSON, one file per source-endpoint-day
data/quarantine/    rows an ingest could not store, with the reason; empty unless something broke
data/reference/     plants, thresholds, rationing episodes, adequacy assumptions, basins,
                    mrids, TLS pins
data/reports/       backtest.md and adequacy.md, regenerated with their models
public/api/         latest.json, status.json, forecast.json, adequacy.json and narrative.json —
                    the documents the site reads and the stable public URLs a third party can
                    fetch — and schema/, their JSON Schemas
public/fonts/       the three self-hosted typefaces and their licences
tests/fixtures/     the Phase 0 responses the parsers are tested against

data/raw/ keeps every response verbatim, so a parser fix is a reprocessing job rather than a re-scrape. Files are plain NDJSON, one record per request, under monthly directories: <source>/<YYYY>/<MM>/<endpoint>.<day>.ndjson, where the day is the one the data belongs to (a backfill of 2018 files under 2018/), or the fetch day for live endpoints; the Información Operativa snapshot, taken on every daily run, gets a file per run (InformacionOperativa.2026-09-22T143642Z.ndjson). A row's raw_ref is <file>#<request key> (two of them, space-separated, on XM exchange rows). Until 2026-09-23 the archive was one gzip bundle per source-month-endpoint, rewritten whole on every run; git cannot delta gzip output, so each run stored a new full copy of every bundle it touched. Uncompressed files written once let git's own compression and deltas do the work: replayed in both layouts, the six daily runs of 2026-09-23 grew the pack by under 15 KiB each in the new one against 60–116 KiB in the old, and a shallow clone of the tip packs to 18.6 MB against 19.6 MB. The price is the checkout: 255 MB on disk against 24 MB. scripts/migrate-raw.ts did the move; refs in the old ….ndjson.gz#key form still resolve, to the day file that holds the record.

data/reference/ is committed input rather than output, and every row says how far it can be trusted. thresholds.csv is derived from operating_bands.csv and mrids.csv, so it needs no outside confirmation — and it keeps all three declarations rather than picking one, because they disagree: Mazar's design floor is 2098 by the dashboard chart title and 2100 by both report endpoints. Its operational minimum is 2115 — the level below which CELEC Sur takes the plant offline for sediment risk, as it has told the press and as it ran the plant in 2024 — so that row leads, and gauges, the forecast's crisis check and the narrative all read Mazar against it. plants.csv carries research from press, so every row of it is marked unverified with an empty verified_on until someone checks it. rationing_episodes.csv was checked on 2026-09-24: each row names its sources and says how the dates were confirmed, including against SMEC demand. adequacy_assumptions.csv is different again: it is the one file here meant to be argued with. The four ceilings the deficit is computed against default to what the fleet has demonstrated in the committed record, and editing a row changes the published deficit on the next run — which is what §7 means by thermal availability and import limits being explicit, editable inputs.

Tables

Table	Grain
observations_daily	date × site × variable × source — levels, inflows, energy, power, spill, plan
national_balance_daily	date × concept, SMEC's seven columns in kWh
operativa_snapshots	fetched_at × block × metric, including demand by utility
operating_bands	one row per distinct declared band, with the span it was observed over
weather_daily	date × basin × coordinates × kind × collection vintage; mm and °C, nulls preserved
enso_monthly	centre month × source, ONI anomaly in °C
forecast_runs	one row per forecast, with the fitted curve and rule behind it
forecast_values	run × horizon — p10/p50/p90, and the raw ensemble bounds beside them
adequacy_runs	one row per adequacy run, carrying the ceilings it stood on
adequacy_values	run × horizon — demand, hydro, net requirement with its band, deficit, margin and tier

The same reading from two endpoints is kept as two rows with different source values rather than silently preferring one, so disagreements between CELEC's own reports stay visible. The models need one number per day, so src/lib/features/series.ts resolves them by a declared source order — which decides coverage rather than truth, since the levels cross-check found the report endpoints and the historian carrying the same series on 100% of 1,668 overlapping days.

forecast_runs records enough of each fit — the area-elevation coefficients, the flow-per-MW, the release stance, the crest — to reproduce a forecast after the history behind it has grown another year. forecast_values keeps the model's raw ensemble bounds next to the published band because they are not the same thing: the ensemble spans what the analogue inflow years do, and the published band is that widened by the model's own out-of-sample error.

adequacy_runs records the ceilings as columns rather than pointing at adequacy_assumptions.csv, for the same reason: the file is editable, so a run that only referenced it would have its past silently rewritten by a later edit.

TLS

No rejectUnauthorized: false anywhere by default. generacioncsr.celec.gob.ec has had a valid certificate since 2026-09-17; www.cenace.gob.ec omits its intermediate, which is committed under data/reference/tls/ and appended to the trust store; smec.cenace.gob.ec serves a self-signed certificate that expired in 2009 and only negotiates at SECLEVEL=1, so it is pinned by SHA-256 fingerprint in data/reference/tls_pins.json and a mismatch fails every SMEC request (the other sources carry on).

Language

Code, commits and developer docs in English; the published site is in Spanish.

Licence

MIT (LICENSE) covers the code. Energy data comes from CELEC EP and CENACE. Weather data is supplied by Open-Meteo using ERA5 and forecast models; its data is licensed under CC BY 4.0, and the free API is for non-commercial use. ONI comes from NOAA PSL / CPC.