Compliance Methodology
Bayesian weight-of-evidence model
Each event is a piece of evidence about compliance with one of the 14 clauses. An analyst assigns a Likelihood Ratio (LR) to every event, and the weight of evidence is derived deterministically:
- LR < 1 → positive weight → evidence favors compliance
- LR = 1 → zero weight → neutral
- LR > 1 → negative weight → evidence favors violation
LR to weight
| LR | Weight (W) | Direction |
|---|---|---|
| 0.01 | +2.00 | Strong compliance |
| 0.1 | +1.00 | Leans compliance |
| 1 | 0.00 | Neutral |
| 2 | -0.30 | Leans violation |
| 50 | -1.70 | Strong violation |
From evidence weight to compliance %
All evidence weights are summed into a cumulative weight (CW) — the headline signal of the trend chart over time.
For the 0–100 gauges, however, we use the mean weight per event (w̄) rather than the raw sum, because the sum saturates to 0% or 100% almost immediately as evidence accumulates. For example a mean of 0 is 50%, +1 is ~91%, and −1 is ~9%.
This percentage is computed live and independently for each clause, each party and the whole agreement, over that part's own set of evidence. So "overall agreement health" is the average direction of all evidence (not the average of the two parties); because each party's evidence set differs, it can be lower or higher than either party's figure.
Aggregation at four levels: clause, pillar, party, overall
An event can be linked to several clauses, so it matters exactly which evidence set each score is computed over:
- Clause score: every linked event contributes its full weight to that clause. An event that breaches several clauses is a complete piece of evidence for each of them.
- Pillar score: not the average of its clauses' scores; the weights and evidence counts of all clauses in the pillar are pooled and the mean is taken over the pool. An event linked to several clauses of the same pillar is therefore counted once per link — multi-clause events carry more weight within their pillar (as more central evidence for it).
- Party score (Iran / US): each event is counted exactly once, by attributed party, regardless of how many clauses it touches. "Both parties" events count toward both scores.
- Overall agreement health: the mean weight of all published evidence, each event exactly once — neither the average of the two parties nor the average of the clauses.
Two consequences of this design: (1) since the gauges use the mean, the index measures the average direction of the evidence, not accumulated certainty — many same-direction pieces of evidence do not raise confidence, only show the mean direction. (2) evidence must be recorded independently; logging several duplicate reports of one underlying incident multiplies that incident's effect accordingly.
Attribution of allies
Each ally's action is attributed to the treaty party it serves: