How much to trust this site

Scored against independent references

In-priority computation vs DWR's own per-right call analysis: 98.7% mean agreement, 99.7% median, 90.9% of days reproducing DWR's exact percent-of-day, over 47 series on 25 structures and 85,822 structure-days. Scored 2026-09-08, re-scored nightly.

DWR gage record vs USGS, day by day, on co-listed gages: 67 basins pass, 3 warn, 1 fail (5 without a co-listed gage). Named: clear_creek warn at 93.3%, boulder_creek warn at 94.8%, d5_wd70_roan_creek_basin warn at 88.4%, d6_wd54_slater_timberlake_creeks fail at 75.6%.

Cached records vs the HydroBase snapshot: 73 basins pass, 0 warn, 0 fail.

Stoppage explanations vs DWR's out-of-priority analysis: 16 of 19 checkable stoppages agree (84.2%); 1,224 stoppages have no DWR analysis to check against and are not counted either way.

Read from the freshness report generated 2026-09-09 06:34; a FAIL that is not in the discrepancy ledger turns /healthz red.

Is the data current right now

Yes as of 2026-09-09 06:34: 3 deployed basin caches, the oldest 0.7 days old. Drafts: present, 0 stale.

Statewide products: SNOTEL + runoff outlook (statewide) ok (newest 2026-09-07); US Drought Monitor ok; well permits (statewide) ok.

DWR API day: 1993 of 2000 requests remain.

Does the engine reproduce its own numbers

Basin	Run	Inputs	Output hash	Two runs agree	Matches the served analysis
clear_creek	2026-09-08 23:57 engine 0.5.0 @ 29fd647	6,916 files fc878ac16a524ffc…	8cb0102b8d7796d5…	yes	not comparable (why?)

Each receipt fingerprints every input file, runs the analysis twice, and hashes the output with the run timestamp removed. Anyone with the same HydroBase snapshot and commit can compare.

Corrections we have published

2026-09-08 issue — we wrote: “The mountains are bare -- all 118 snow stations read zero, which is exactly what late summer should look like.” Not all of them. Twenty of the 118 SNOTEL stations were reporting a trace of snow on the ground that morning, 0.3 inches of water at most. What was zero everywhere was the usual amount for the date -- the 1991-2020 median -- and our sentence described the median as if it were the ground. Source: NRCS AWDB readings for 2026-09-07 as cached by the site at 11:40 UTC on 2026-09-08 (web/var/snow/snow.json): 20 sites with SWE above zero, maximum 0.3 in, every site's median 0. · the issue

Where the sources disagree

16 entries in the discrepancy ledger: 4 accepted, 5 open, 7 resolved. Every place two sources disagree, or we disagreed with ourselves, with what each side said and what we did. Entries are never deleted.

• isf-tabulation-division-wd43-wd47 · 2026-09-09 · accepted
• usgs-crosscheck-fail-wd54 · 2026-08 · open
• usgs-crosscheck-warn-x3 · 2026-08 · open
• methods-page-typed-score · 2026-09-08 · resolved
• dwr-usgs-off-by-one-day · 2026-08 · accepted
• reach-direction-assumed · 2026-08 · open
• cur-decreed-use-token · 2026-08 · open
• mirror-lacks-stream-miles · 2026-08 · accepted
• never-summer-snotel-silent · 2026-09-08 · open
• netamount-empty-case-column · 2026-08 · accepted
• cutline-caller-counted-junior · 2026-09-08 · resolved
• minus-999-as-flow · 2026-08-30 · resolved
• future-dated-divrec-placeholders · 2026-08 · resolved
• snow-stations-zero-2026-09-08 · 2026-09-08 · resolved
• undefined-as-mainstem · 2026-09-08 · resolved
• anonymous-vs-keyed-quota · 2026-09-08 · resolved

Checked by hand against the agencies' own sites

Each month twenty published figures are drawn at random (the draw is fixed by the month, so it cannot be redrawn for a friendlier sample) and a person re-derives each from DWR's or NRCS's own page. So far: 0 of 20 drawn figures checked, 0 matched, 0 did not. A mismatch becomes a correction and a ledger entry.

• 2026-09: 0 of 20 checked, 0 matched, 0 mismatched

Who has reviewed this

Nobody outside the project yet. The methods have not been reviewed by an outside water-law or engineering reviewer; that is the sentence to weigh before relying on anything here for a filing. Reviews are published here with the reviewer's name, what they read, and their disagreements verbatim.

Reviewer packets ready to send (the exact code and tests behind one claim, extracted from the source): call-applies, mass-balance, sentinels. Ask for one.

What the tests hold

2,440 test functions across 245 files, counted from the source as this page renders; 577 of them are named for an honesty rule (never, refuses, is not, invented, verified). The suite runs before every push, on the machine that pushes; this page counts what exists, it does not claim a pass.

When readers report a problem

1 reports received through Report a problem; 1 read by a person, 0 waiting. Of those read, 0 changed the site (a correction, a ledger entry, or a code change), 0 were checked and the site stood, 1 were questions, thanks, or noise, and 0 are read but not yet decided. A report that changed the site is also a correction or a ledger entry above.