• Shortfall % (social): How far below the social foundation threshold California falls. A bill that reduces shortfall is moving California toward the safe space.
• Overshoot % (ecological): How far beyond the planetary boundary California has gone. A bill that reduces overshoot is pulling California back within ecological limits.

Our primary output for each bill is the projected change in shortfall or overshoot percentage points (p.p.) for every affected indicator. A negative change (e.g., -2.8 p.p.) means improvement — California moves closer to the safe and just space.

Phase 1: Triage & Category Detection

A fast AI model analyzes each bill's title, summary, and subject tags to determine:

• Which of the 21 doughnut categories the bill substantively impacts (most bills impact 2-5)
• How complex the analysis will be (simple / moderate / complex)
• Whether the bill has zero doughnut impact (procedural, commemorative bills)

Phase 2a: Standard Analysis (Simple/Moderate Bills)

For each relevant category, an AI model analyzes the bill using the full bill text, CalDEC indicator definitions with current California values and desirable targets, and data source citations. The model estimates how the bill would change each indicator's value and computes the resulting shortfall/overshoot change.

Phase 2b: Research-Augmented Analysis (Complex Bills)

For complex bills (omnibus, cross-cutting), the system first searches for relevant academic papers and policy research, then feeds this evidence into a more capable AI model for deeper analysis grounded in empirical findings.

Step 1 — Per-Bill Predictions

For each bill, the AI predicts raw indicator changes (e.g., “reduce CO2 emissions by 0.3 tonnes/capita”). These are converted to shortfall/overshoot percentage point changes using the indicator's current California value and target threshold.

Step 2 — GBD Multiplicative Compounding

When multiple bills affect the same indicator, their changes are combined using the GBD (Global Burden of Disease) multiplicative model — the same approach used by WHO/IHME in the Lancet for 20+ years. Each bill reduces (or increases) the remaining shortfall rather than the original, so improvements naturally compound and can never exceed full closure.

This prevents the absurd results you'd get from simple addition: if 500 bills each reduce a 100% shortfall by 1 percentage point, simple addition says -500 p.p. (impossible!). GBD compounding gives ~99.3% reduction — nearly full closure, but never exceeding it.

Step 2b — Policy Mechanism Interaction

Not all bills affecting the same indicator should be combined equally. Bills are tagged with 30 standardized policy mechanism types, and these mechanisms fall into three interaction tiers:

• Regulatory Standards (max-takes-all): Emission limits, price caps, labor standards — only the strictest binds. A $1,500/month rent cap makes a $1,800 cap redundant. Mechanisms: emission standards, chemical restrictions, price regulation, labor standards, carbon pricing.
• Population-Constrained Programs (diminishing): Job training, housing assistance, healthcare coverage — each additional bill at 50% weight. They compete for the same people and resources. Mechanisms: workforce development, housing assistance, healthcare coverage, education programs, public health programs, and 9 others.
• Independent Capacity (additive): Infrastructure, conservation, transparency — all count fully. Each builds something new or opens a distinct channel. Mechanisms: infrastructure investment, clean energy deployment, ecosystem restoration, conservation designation, and 6 others.

Across different mechanisms, the standard GBD compound applies — different mechanisms are independent channels that don't substitute for each other.

Worked Example: Reducing Unemployment

Three bills target U-6 unemployment (current shortfall: 50%):

• Bill A (workforce_development): -3 p.p.
• Bill B (workforce_development): -2 p.p. → discounted to -1 p.p. (same mechanism, 50% weight)
• Bill C (infrastructure_investment): -4 p.p. (full weight, different mechanism)

Within workforce_development: compound [-3, -1] → -3.94 p.p.
Within infrastructure_investment: -4 p.p.
Across mechanisms: compound [-3.94, -4] → -7.63 p.p.

vs. flat GBD (no mechanism awareness): compound [-3, -2, -4] → -8.65 p.p.

Why This Approach

The mechanism interaction model is grounded in policy research:

• GBD Comparative Risk Assessment (WHO/IHME, Lancet 2024) — foundation for multiplicative compounding
• Perino, Ritz & van Benthem, “Overlapping Climate Policies,” Economic Journal 2024 — regulatory substitution evidence
• Hepburn et al., “Policy Complementarity and the Paradox of Carbon Pricing,” Oxford Review of Economic Policy 2023 — different-mechanism complementarity
• Wang, “Diminishing Returns of Policy Pilots,” Review of Policy Research 2024 — diminishing returns in overlapping programs
• CBO scoring methodology — interaction discounts (~50%) for overlapping provisions
• Milgrom & Roberts (1990) supermodularity framework — complementarity when addressing different binding constraints

Existing Policy Context

Each bill is scored for its marginal impact — the incremental change it would produce beyond what California's existing laws and programs already achieve. When the AI estimates that a workforce development bill reduces unemployment shortfall by 0.3 percentage points, that estimate already accounts for the existence of programs like ETP, CalWORKs, and other current workforce initiatives.

The mechanism interaction tiers (max-takes-all, diminishing, additive) handle only the interaction between new bills in the current session. They do not double-count interactions with existing policy, because the per-bill estimates are already marginal over the current baseline. The shortfall/overshoot percentages themselves reflect the state of California with all existing policy already in effect.

Uncertainty Ranges

Each bill's per-indicator impact comes with low, mid, and high estimates reflecting the inherent uncertainty in policy impact prediction. We propagate these ranges through the aggregation in two steps:

1. Triple compounding: Run the full mechanism-aware compounding three times — once with all low estimates, once with all mids, and once with all highs. This produces a “naive” range representing the all-pessimistic and all-optimistic scenarios.
2. √N shrinkage: The naive range assumes all bills simultaneously hit their worst or best case, which is statistically implausible. By the central limit theorem, when N independent estimates are combined, the aggregate uncertainty narrows by a factor of 1/√N. We apply this shrinkage: if 100 bills affect an indicator, the displayed range is 1/10th as wide as the naive range. With 4 bills, it's half as wide. With 1 bill, the full individual range is shown.

The resulting range (e.g., “Close 40-60% of the gap”) reflects remaining statistical uncertainty after this narrowing. The true outcome is likely somewhere within these bounds.

Where the range is narrow (less than 5 percentage points), we display only the mid estimate to keep the interface clean.

Step 3 — Gap Closure %

The compounded change is expressed as gap closure percentage — “what percentage of the gap between where CA is and where it needs to be does this legislation close?” This metric is intuitive and bounded (0-100%+), unlike raw percentage points which are hard to interpret without context.

Worked Example: GHG Emissions

Say California's CO2 footprint is 19.6 tonnes/capita with a target of ≤0.985 tonnes/capita, giving a shortfall of ~1890%.

• Bill A predicts -15 p.p. shortfall change (reduces shortfall from 1890% to 1875%)
• Bill B predicts -20 p.p. shortfall change
• Simple sum: -35 p.p. GBD compound: -34.8 p.p. (nearly the same for small changes)

Now imagine 364 bills each averaging -3 p.p. change against 1890% shortfall. Simple sum: -1092 p.p. GBD compound: -1037 p.p. — roughly 55% gap closure. The compounded result reflects reality: each bill works on the remaining gap, so cumulative impact has diminishing marginal returns.

On the dashboard this shows as: “Climate Change — Close ~55% of the gap (364 bills)” with a projected value of ~9.5 tonnes/capita, down from 19.6.