This page is for a reader who has two minutes and no context. Every other page in this project answers one question; this page says which one, and what to look at first. Everything is offline, needs no keys, and every number is reproducible from the repository.
A data centre can be built in 2–3 years, but connecting it to the grid takes 7–10 years. On stage the utility was asked what a 98%-availability contract would change, and answered "we're trying". GRIDRUSH answers that question with a computed plan: which edge sites to connect, at what MW, under which availability tier and with what flexibility — verified over all 8,760 hours of 2025 and a real AC power flow.
Click order on the stage dashboard: site map → three-strategy card → MW by month → worst day → Replay loop. Nothing else is clicked.
The one page shown on stage. Six panels, left to right, top to bottom:
| Panel | What it shows | What to look at |
|---|---|---|
| KPI strip | four headline numbers of the winning plan | 4.30× net value vs the best honest strategy · 85.4 MW online at month 12 · $279.9M modelled compute margin · 0 MWh grid violations |
| 1 · Site portfolio | the 32 sites the agent energized, on a map of substations, shaded by availability tier | dark amber = 99.99% commitment, light amber = 99%; the 8 refused sites are grey. Hover any dot for its MW, SLA, month, curtailment hours |
| 2 · Three strategies | the comparison that carries the pitch: all-firm, all-non-firm, and the agent | 0 MW / $65.1M → 173 MW / −$449.5M → 85.4 MW / +$279.9M. The middle row is the punchline: speed without reliability is loss-making |
| 3 · Energized MW by month | running MW for months 1→36, by connection type | ramp reaches full power by month ~12; this is why the flexibility premium exists |
| 4 · Worst day | a real 2025 day: demand, free headroom, required curtailment for the naive plan and for GRIDRUSH | at 18:00 the naive plan must shed 28.4 MW, GRIDRUSH 1.0 MW; at the 20:00 peak (43,860 MW) it is 18.2 vs 0.0 |
| 5 · Agent learning | score of the best plan at every iteration | fast search plateaus at 179; the LLM-written policy jumps to 280; 207 iterations. Press Replay loop to re-run the animation |
| Assumptions | one line per modelling parameter, tagged REAL / SYNTHETIC / SPEECH | this is the honesty panel — point at it if a judge asks "is this real?" |
Four operators on their own 2025 data with the same policy, no tuning: CAISO 4.30×, ERCOT 3.41×, PJM 4.07×, MISO 3.88× — zero violations everywhere. Below them, two "reality check" cards: physics inside the optimiser (2.39×) and real PG&E substations (1.22×). This is the page to open for "show me it generalises" or "is any of it real".
One measured AC power-flow study on a real 20 kV medium-voltage feeder (pandapower /
mv_oberrhein: 179 buses, 2×25 MVA transformers, 37.1 MW served).
| Number | Meaning |
|---|---|
| 0.9756 p.u. | base-case minimum voltage — the feeder is healthy before we add load |
| 8.2 MW | spare capacity before the first thermal violation (a transformer hits 100.4%) |
| 31.8 MW | the five largest sites of the agent's plan — they do not fit on one feeder (71 violations) |
| 2.84–6.60 MW | per-bus headroom: what a single bus can take with zero violations |
The conclusion on the page is the useful part: this is a siting problem, not a megawatt problem — 0.5–10 MW edge sites are the right unit, and they must be distributed.
The same dashboard layout, but the loop ran with a measured admission limit of 6.53 MW of new load per feeder as a hard constraint inside the portfolio builder — and the final plan is verified with eight real AC power-flow runs, all clean.
Why 6.53 and not 8.2? They answer different questions. 8.2 MW is the first thermal violation when the top-5 profile is scaled uniformly; 6.53 MW is the conservative limit we enforce: the same total spread over the five best-headroom buses in proportion to their headroom, with every bus at or above 0.95785 p.u. and every branch at or below 100%.
What to look at: the headline drops from 4.30× to 2.39× ($279.9M → $155.5M, 85.4 → 44.7 MW). That −44.5% is the honest cost of real physics — say the number before a judge asks.
The site model rebuilt on the public PG&E GRIP Data Portal (ArcGIS REST, no key): eight real substations with real coordinates, facility ratings and 2026 loadings — e.g. SAN RAMON 302.2/136.9 MW, STOCKDALE 327.1/275.2, CLAYTON 193.1/124.9.
Result: 1.22×, 397.7 MW at month 12, 11,533 contract-compliant MW-months, zero violations.
Why so much lower — and why this is the most valuable finding: at the substation aggregate level, real PG&E substations are large and are not the bottleneck, so flexibility buys little. The constraint lives at the circuit level: in a sample of 2,000 circuits, 124 had a positive 2026 deficiency (538 rows are a "no data" sentinel, so the honest share is 124 of 1,462 circuits with data = 8.5%). That is exactly the data GRIDRUSH asks for.
The identical dashboard on four operators' 2025 data. Use one of these to answer "does it work outside California?" — the ratios and the three-strategy comparison are on each page, and nothing was retuned per region.
Deck: ten slides — problem, what the agent decides, three strategies, why it wins, worst day, the learning loop, two reality checks, honest limitations, the ask. Use it as a leave-behind or to brief a judge in 90 seconds.
Video: the three-minute walkthrough of the same story, for readers who prefer watching.
The archive contains everything: the simulator, the RSI loop and its logs, all result JSONs, the dashboards, the deck, the one-pager and the written material. Offline, no keys, no network. The repository adds history and the full source of every page.
| Number | Meaning |
|---|---|
| 4.30× | net value of the agent's plan vs the best of three honest baseline strategies, over 36 months |
| 3.82× | the same comparison in physical units: 2,490.6 vs 651.3 contract-compliant MW-months (contains no prices) |
| 85.4 MW | compute online at month 12, all of it inside its own SLA (the firm-only plan has 0) |
| +$279.9M | modelled compute margin over 36 months — a margin metric on a synthetic portfolio, not revenue |
| −$449.5M | what the "sign everything non-firm at max" strategy loses after 1.15M MWh of SLA breach |
| 28.4 MW → 1.0 MW | required curtailment at 18:00 on the worst day: naive portfolio vs GRIDRUSH |
| 179 → 280 | learning curve: plateau of the fast search, then the LLM-written policy's jump |
| 207 | iterations: 200 parametric + 7 LLM-written policy generations (one failed, kept in the log) |
| 2.39× | the same agent with the measured AC-power-flow limit inside the loop — the honest, physics-checked number |
| 1.22× | the same agent on eight real PG&E substations — where the constraint moves to the circuit |
| 6.53 MW | measured admission limit per MV feeder (voltage ≥ 0.95785 p.u., loading ≤ 100%) |
1. "Zero modelled violations" is structural for flexible service: flexible load is clipped to headroom by construction. The constraint that actually binds is each site's SLA, enforced with zero breach energy. Violations would matter for a firm-heavy portfolio.
2. The ratio is invariant to the price level (4.30× at ⅓×, 1× and 3× margin), and the physical ratio 3.82× contains no prices at all. Flattening the SLA premium and doubling the penalty still gives 4.04×. What is assumption-dependent is the absolute dollar level, not the comparison.
3. The "all firm" baseline is honest but weak (651 MW-months at 200 MW of nameplate) because firm capacity is capped by the substation's tightest hour. That is precisely why the real-PG&E result (1.22×) matters.