Design: California Reservoirs, Droughts & Population — post + interactive model

Date: 2026-05-28 Status: Approved (design); ready for implementation planning Author: Jacob (with Claude)

1. Summary

A blog post in Jacob’s essayistic voice paired with an interactive, data-grounded water-balance model of California’s surface reservoirs (1950–2024). The reader drags levers — population, per-capita use, new reservoirs, a drier climate — and watches the statewide reservoir level recompute over time, along with a count of low-reservoir events avoided.

The post is exploratory but lands a thesis the model earns:

We argue endlessly about building more dams, but concrete is the weakest lever — reservoirs move water in time, they don’t create it. The scarcity gap is set by climate (which we don’t control) and by demand (dominated by agriculture, which stopped tracking population decades ago). Restraining demand does several times more than any reservoir we could realistically build. The popular answer is the least effective; the effective answers are the politically hardest.

This thesis is a synthesis between the two framings we started with (“supply can’t win” vs. “population / we stopped building”). The spike confirmed: supply genuinely can’t win, and population is a bigger lever than concrete (validating Jacob’s pushback) — but the true demand giant is agriculture, and climate dominates everything.

2. Spike findings (already run; these justify the thesis)

A prototype model (prototype.mjs, in the gitignored brainstorm scratch dir) ran the real series through an annual water balance and tested each lever. Results are robust across inflow calibrations (k = 2.0 / 2.2 / 2.5); the lever ranking is the load-bearing output, not the absolute MAF figures.

Share of the cumulative “supply gap” each lever closes:

Low-reservoir events avoided (of 25 baseline events; storage < 33% of capacity):

The “cut ag” rows are diagnostic lever tests, not shipped controls — ag is not a user knob (see §3). They’re included to show how large the latent agricultural lever is relative to the controls readers can touch.

Key qualitative insight for the prose: in severe multi-year droughts (1977, 1989–92, 2014–15) the reservoirs drain to deadpool no matter which lever you pull — the water never fell. Lower demand helps in moderate dry years and speeds recovery. Wet-year spill barely changes as capacity is added, proving the binding constraint is inflow, not storage.

3. The model

Annual water balance over water years 1950–2024, in million acre-feet (MAF). Conceptually a single statewide surface-water bucket.

For each year t:

pop      = popOverride ?? CA_POP(t)
gpcd     = min(GPCD(t), gpcdCap ?? Infinity)
urban    = pop * gpcd * 365 / 325,851e6        // gallons -> MAF
ag       = AG(t)                                // real historical series; fixed (not a knob)
demand   = ag + urban

inflow   = RUNOFF(t) * K_INFLOW * climateMult   // Sac 4-river runoff scaled to statewide proxy
capacity = CAP(t) + addedCapacity               // physical buildout + toggled/slider reservoirs

avail       = storage_prev + inflow
delivered   = min(demand, max(0, avail - DEADPOOL))
shortfall   = demand - delivered                // "supply gap" (met in reality by groundwater)
storage     = avail - delivered
spill       = max(0, storage - capacity); storage = min(storage, capacity)

lowEvent    = storage < LOW_FRAC * CAP(t)        // 33% of *physical* capacity

Constants (prototype values; tune during build): K_INFLOW ≈ 2.2, DEADPOOL = 4.0, INIT_STORAGE = 20.0, LOW_FRAC = 0.33.

Calibration / honesty guardrail. The baseline run (all levers at “actual”) must reproduce the historical pattern: storage drains to deadpool in the real droughts and spills in the real wet years. We plot recorded statewide storage (CDEC, available from ~1970) as a faint reference so readers can see the model tracks reality before trusting the counterfactuals. Absolute MAF figures are illustrative; the lever ranking is the claim.

Levers (controls)

1. Population scenario — override CA population (presets: actual / freeze at 1970 / freeze at 1950 / a slider). Scales urban demand only.
2. Per-capita use (conservation) — cap or scale GPCD. Separate from headcount.
3. New reservoirs — toggles for real projects, each adding capacity: Sites (+1.5 MAF), Temperance Flat (+1.26), Shasta raise (+0.63), B.F. Sisk raise (+0.13), Del Puerto (+0.08), Los Vaqueros expansion (+0.115), Pacheco (+0.13), Auburn (+2.3, never built). Plus a “go further” slider for hypothetical extra MAF.
4. Drier future (climate) — scalar multiplying inflow (e.g. −10% / −20%).

Ag demand is a fixed real series (decision: not a user knob), but the prose/model notes that ag is the dominant share and the largest latent lever.

Outputs

• Primary chart: statewide reservoir storage over time — scenario line vs. actual baseline, with capacity ceiling, the low-event threshold line, drought shading, and markers on low-reservoir years. (The chart built in the brainstorm session is the reference design.)
• Secondary: unmet-demand-per-year bars (the supply gap).
• Headline metrics: low-reservoir events (and events avoided vs. baseline), years in shortfall, % demand met, water spilled.

4. Interactive piece — layout & architecture

Mirrors the existing lab/comp-viz/ precedent exactly.

lab/water-viz/
  index.html        full-page version (Layout B: controls in a left sidebar)
  embed.html        in-post iframe (Layout A: controls on top, charts stacked)
  styles.css
  embed.css
  src/
    data.js         real series as JSON consts, each with an inline source citation
    model.js        pure water-balance fn: (series, levers) -> {trace, metrics}
    charts.js       canvas/SVG time-series + bar chart
    main.js         full-page wiring
    embed-main.js   embed wiring

• Vanilla JS ES modules, <script type="module">, no framework — same as comp-viz.
• The post embeds via <iframe src="/lab/water-viz/embed.html" ...> plus an “open the full version” link, matching the comp-viz post pattern.
• model.js is a pure function (no DOM) so it’s trivially testable and the prototype’s logic ports directly.

5. Post structure (Jacob’s voice)

Roman-numeral sections, concrete → abstract, measured claims, honest caveats, no neat bow. Rough arc:

• I. The recurring image: California reservoirs low again; the reflexive “build more dams.”
• II. What reservoirs actually do — store wet-year water for dry years; they don’t create water. Introduce the three forces: climate, demand, supply.
• III. The interactive model (embed). Invite the reader to try the obvious fix — build every proposed reservoir — and discover it avoids ~0 drought events.
• IV. Why: you can’t fill what won’t fill; proposed capacity is ~2% of annual use; spill barely moves. Population is a bigger lever than concrete — but agriculture (flat, ~80%, decoupled from population) is the real giant, and climate swamps all of it.
• V. The uncomfortable conclusion + honest limitations.

Target rubric: LLM score < 30, Jacob score > 65 (per AGENTS.md).

6. Data sources & provenance

All series gathered in the research pass; stored in data.js with per-series citations.

• Runoff (engine input): Sacramento 4-river index (“SacWYsum”, MAF), annual 1950–2024. Source: CDEC WSIHIST — https://cdec.water.ca.gov/reportapp/javareports?name=WSIHIST . Range 5.12 (1977) – 37.82 (2017), mean ~17.65.
• Reservoir capacity buildout: major-reservoir completion years + capacities (CDEC, Wikipedia “List of largest reservoirs of California”, USBR). Scaled so the stepped total reaches ~43 MAF statewide (PPIC’s “major reservoirs” figure; ~50 MAF for all ~1,500).
• Storage (validation reference): CDEC (statewide aggregate only from ~1970); long-term avg ~23.5 MAF (USGS Dettinger & Anderson 2015). Pre-1970 statewide storage does not exist.
• Water use by sector: Pacific Institute (Cooley 2020) & PPIC (2019), from DWR Bulletin 160 / California Water Plan. Ag ~25–37 MAF (peak ~1980, flat-declining since); urban ~3.1 (1960) → 9.3 peak (2007) → 7.0 (2015). Developed split ≈ 80% ag / 20% urban.
• Per-capita (GPCD): 177 (1960) → 231 (1990–2007 plateau) → 146 (2015). PPIC / Pacific Institute.
• Population: CA 10.586M (1950) → 39.538M (2020), flat after (US Census + CA DOF); US 151M → 331M (Census).
• Proposed reservoirs: California Water Commission / WSIP, project authorities, PPIC. PPIC framing: all proposed new storage ≈ 3.3 MAF capacity but only ~0.76 MAF/yr new supply (~2% of use). Off-stream projects (Sites, Sisk, Los Vaqueros, Pacheco, Del Puerto) fill only from wet-year high flows.
• Ag/population linkage (for prose, not a modeled channel): CA supplies ~75%+ of US fruits/nuts; ~$23.6B exports (2022); shift to permanent crops (almond acreage 327K→1.24M, 1980→2020) that can’t be fallowed. But ag water volume decoupled from population/acreage — flat-to-declining 40 years while output value and population rose. So we model ag as a fixed real series and discuss the linkage qualitatively.

Data fallbacks: where only benchmark years exist (capacity, ag, GPCD, population), linearly interpolate to annual and label as interpolated in data.js comments. Runoff is true annual.

7. Limitations (the post’s honesty section)

• Single statewide bucket: ignores geography (the Delta, that NorCal water can’t freely move south), the CVP/SWP split, and Colorado River imports.
• Groundwater is out-of-model: the “supply gap” is in reality met by groundwater overdraft — itself the slow-motion crisis (subsidence, SGMA). The gap is best read as “stress / reliance on overdraft,” not literal taps running dry.
• Snowpack (a huge natural reservoir, ~17 MAF avg April 1) is folded into runoff timing, not modeled separately.
• Absolute MAF outputs depend on the K_INFLOW calibration; the robust result is the lever ranking.
• Ag demand is a smoothed series; real ag use is weather-responsive.

8. Out of scope (YAGNI)

• No monthly/daily resolution — annual water years only.
• No separate CVP/SWP or regional modeling.
• No deeper-droughts or snowpack-shift levers (considered, cut to keep scope sane).
• No backend/data API — all series baked into data.js as static consts.