1Overview

The Regime Death Index (RDI) is a quantitative model for measuring the erosion of Iran’s Islamic regime. It combines two dimensions: the Military Capacity Index (MCI) and the Governance Fragility Index (GFI). Each variable is scored with an operational definition, weighted sub-indicators and a scoring rubric so that the calculation is transparent and reproducible.

MCI_raw = Σ (destruction_% × weight)                       8 weighted variables
MCI_adj = ISAF × MCI_raw                                   ISAF ~ Beta(80, 20), mean 0.80
GFI     = simple average of 6 governance erosion variables
RDI     = 0.35 × MCI_adj + 0.40 × GFI + 0.25 × (MCI_adj × GFI / 100)
Collapse threshold: 65

ISAF multiplies MCI only. It is the share of military destruction that is permanent, i.e. not reconstructible under wartime conditions. It does not scale GFI and it does not scale the whole bracket. (v3.2 printed RDI = ISAF × [ … ]; that formula was never what the published series was computed with. See §11.)

The interaction term MCI_adj × GFI / 100 models mutual reinforcement: military destruction and governance erosion amplify each other when they co-occur.

Maximum attainable RDI is 100 (MCI_raw = GFI = 100 with ISAF = 1); at ISAF = 0.80 the maximum is 88.

2Measurement Units and Scoring Rules

Unit. All variables are measured on a 0–100 scale. MCI variables — “destruction %”: 0 = intact capacity, 100 = total destruction. GFI variables — “erosion level”: 0 = healthy governance, 100 = complete collapse.

General score bands

Raw data → score

Each variable comprises weighted sub-indicators. Each sub-indicator is scored 0–100 against an objective metric (destroyed facilities / known facilities, CENTCOM assessment, satellite imagery, …). Variable score = weighted average of its sub-indicators.

Reproducibility rule

Two analysts scoring the same variable from the same cumulative evidence dossier, using the anchors in §4, arrive at scores within ±5 units of each other. The rule’s input is the dossier plus the anchors, not “raw data” in the abstract; a score without a dossier behind it is not covered by the rule.

Tested 10–11 September 2026 on Day 77 (see §12 and the test packet): two independent blind raters given an identical dossier agreed within ±5 on five of six GFI variables; the sixth (social base) differed by 15 and was flagged as thinly evidenced by both. The rule holds between raters using the rubric.

The rule does not hold between the rubric and the published Day 1–100 GFI series, which was scored before the anchors existed. See §2a.

Change rule

A variable changes only on documented evidence: source + date + a change of at least ±2 units. Daily reports update only the variables that changed and the events that are new.

2a. Two GFI lines

The published series (Day 1 – Day 100) carries GFI scores produced by cumulative analytical judgement, day by day, before the §4 anchors were written. When those scores were tested against the anchors at Day 77, the two society-facing variables (social base, religious legitimacy) agreed within 3–11 units, and the four state-apparatus variables (central command, IRGC cohesion, war economy, information control) sat 30–38 units above what the anchors support. The pattern is consistent: the series scored those four by the damage inflicted on the apparatus — strikes, commanders killed, revenue lost — while the operational definitions in §4, and therefore the anchors, score loss of function: whether the apparatus still decides, obeys, finances and broadcasts. Both are defensible measurements; they are not the same measurement.

The series is therefore not restated. It is kept as published, labelled GFI-D (damage-weighted, pre-rubric), and a second line, GFI-R (rubric-scored), is published beside it for anchor days: Day 30, Day 46, Day 77 and Day 100. Each GFI-R point is produced by a blind rater from a cumulative dossier through that day, by the procedure in the test packet. RDI is reported for both lines at those days, with the same MCI, ISAF and formula; only the six GFI inputs differ. Readers see two lines and the reason for the gap.

The four GFI-R points, scored 12 September 2026 by four blind raters from a single dated master dossier cut at each day (master_dossier.py; packets GFI-R_packet_day*.md):

The lines diverge over time: GFI-D rose through the ceasefire, GFI-R did not. Under the rubric the regime’s governance apparatus was damaged by Day 30 and did not lose further function afterwards; under the series it kept eroding. Readers are given both, and the reason.

From any scoring after the publication of this document, GFI is scored by the §4 anchors from a cumulative dossier — the GFI-R method — and GFI-D is not continued.

3MCI Variables — Operational Definitions and Sub-indicators

σ is the standard deviation used in the Monte Carlo (0–100 scale). The values below are those used in the daily code from Day 46 onward; the v3.2 table listed different σ for eight variables and is superseded.

3.1 Missile Production — weight 18 %, σ 3

Percentage destruction of ballistic and cruise missile production, assembly and maintenance capacity relative to pre-war level. Excludes launchers (3.2) and existing stockpiles.

3.2 TEL Launchers — weight 15 %, σ 5

Percentage of mobile ballistic-missile Transporter-Erector-Launchers destroyed or neutralised relative to pre-war inventory. Direct calculation: destroyed / total.

3.3 Air Defense — weight 12 %, σ 4

Percentage loss of air-defense capacity (SAM systems, radars, fighters). Key indicator: do US/Israeli forces have complete aerial dominance?

3.4 Drone Capability — weight 10 %, σ 5

Percentage destruction of military-drone production, maintenance and operational capacity. Shahed production is decentralised and harder to destroy than ballistic missiles; the 10 % weight reflects that replaceability.

3.5 Naval Forces — weight 12 %, σ 5

Percentage destruction of naval capacity: conventional fleet, asymmetric capability, military ports.

3.6 Proxy Network — weight 15 %, σ 7

Percentage destruction of Iran’s capacity to use proxy forces (Hezbollah, Houthis, PMF, …) for extraterritorial operations.

3.7 Territorial Control — weight 8 %, σ 4

Percentage loss of the regime’s physical control over Iranian territory (occupied zones, lost airspace). Without a ground invasion this variable stays low; the 8 % weight reflects that.

(Sub-indicators for 3.7 were absent in v3.2 and are added here; weights are analytical and are subject to the §11 note on weights.)

3.8 Internal Security — weight 10 %, σ 6

Percentage destruction of the regime’s domestic repression capacity (FARAJA, Basij, IRGC security units).

4GFI Variables — Governance Fragility

GFI variables are more qualitative than MCI. Their erosion level is derived from qualitative indicators and news evidence. All six are equally weighted; GFI = simple average.

Scoring anchors for GFI (new in v4.1)

Each GFI variable is scored against five anchors. The analyst places the evidence at the nearest anchor, then adjusts by at most ±10 within the band with a stated reason.

5Iranian Society Adjustment Factor (ISAF)

ISAF models structural features of Iran that govern how much military destruction is permanent. It multiplies MCI_raw only.

ISAF_raw = 1 + (0.08 + 0.06 + 0.04) − (0.18 + 0.12 + 0.05) = 0.83

Value in the model and its history

The v3.2 document’s “effective 0.86 after calibration to historical RDI trend” is withdrawn: no such calibration was ever applied in the code, and calibrating a scaling factor against the index it scales is circular.

Update rule

ISAF is revised only for endogenous change — evidence about reconstruction capacity, institutional structure or society — never for external strikes, which enter through the MCI variables. Every revision is announced in the report of the day it takes effect, with the reason, and the pre-revision RDI is shown beside the post-revision RDI on that day.

Why ISAF is the dominant uncertainty

At Day 77 inputs, one standard deviation of ISAF (0.04) moves RDI by 3.2 units. The largest single variable (IRGC cohesion, σ 6) moves it by 1.1. ISAF therefore appears in the tornado chart as a driver in its own right (§6).

6Monte Carlo v4.1

10,000 iterations, fixed seed (20260910), Gaussian copula over the 14 variables, Beta marginals, ISAF sampled independently.

Algorithm

1. Correlation matrix. The 22 authored pairs (§6, table) populate a 14×14 matrix. That matrix is not positive semi-definite (smallest eigenvalue −0.32) and cannot be Cholesky-decomposed as authored. It is repaired with Higham’s nearest-correlation algorithm and an eigenvalue floor of 10⁻⁴. The repaired matrix — not the authored one — is the canonical matrix and is published as rdi_correlation_canonical.csv.
2. Correlated normals. Z ~ N(0, I); X = L·Z with L·Lᵀ = R_canonical.
3. Copula transform. U = Φ(X), giving correlated uniforms.
4. Beta marginals. For each variable, μ = score/100, σ as in §3–§4, converted to Beta(α, β) by moment matching, with σ capped at 0.99·√(μ(1−μ)) and α, β ≥ 1.01 (identical to the daily code).
5. ISAF. One draw per iteration from Beta(80, 20).
6. RDI per iteration via §1; outputs: mean, median, 90 % interval, sd, P(RDI ≥ 65), distance to threshold = 65 − median.

Authored correlation pairs and their canonical values

The repair also introduces implied correlations of |ρ| ≤ 0.07 between pairs that were authored as zero (largest: IRGC cohesion ↔︎ religious legitimacy 0.07, central command ↔︎ social base 0.06). They are consequences of the authored pairs, not new assumptions. The full 14×14 canonical matrix is the CSV.

Effect of restoring the copula (Day 77 inputs)

The mean does not move. The spread does: published P(≥ 65) figures for Days 46–100 are understated by roughly a quarter of their value.

Tornado sensitivity

For each of the 14 variables, RDI is computed at μ−σ and at μ+σ, all else at μ, ISAF at its mean. ISAF is added as a fifteenth driver at mean ± 0.04. Drivers are ranked by range. At Day 77: ISAF 3.16; IRGC cohesion 1.09; religious legitimacy 1.09; proxy network 0.92; information control 0.91; social base 0.91; TEL launchers 0.66; all others below 0.55.

7Collapse Threshold Model

The 65 threshold is an analytical estimate: when roughly two-thirds of the regime’s permanent military capacity and governance capacity is gone, reconstruction becomes practically impossible. Collapse may occur above or below it.

Necessary conditions (all three must be active)

Definition change, 12 September 2026. Until this version the first element read “zero oil revenue”. It was never measurable against a Tier-1 source and, on independent check, was false at every anchor day: Iranian exports averaged 1.85 mb/d in March 2026, above the pre-war trend, because the IRGC controlled Hormuz and passed its own tankers; loadings fell to 567 kb/d only after the US blockade of 14 April (Day 46), and floating-storage sales to China continued through June. Under the new wording the condition first activates on Day 74 (12 May 2026), four weeks after the blockade began. The published Day 30–73 reports stated “3 of 3 necessary conditions active”; the record is restated as 2 of 3 for those days and 3 of 3 from Day 74. The Day 77 and Day 100 threshold statements are unchanged.

Accelerators (0 to 4 active)

Shocks (sudden triggers)

Shock impacts are applied as a stated addition to the point RDI on the day the shock is confirmed, are shown separately from the variable-driven RDI, and are not carried inside the Monte Carlo.

83×3 Security Capacity Framework

Since v3.2, the regime’s security capacity is analysed on three axes, each in three levels.

Cells are updated with documented evidence (satellite imagery, IDF/CENTCOM statements, HRANA/NCRI reports). From Day 70 the daily report carries the nine cells inside a single “Erosion” section that reports changes only (§9).

9Daily Report Structure

Reports up to Day 69 used the 13-section v1.0 layout. From Day 70 the reports use layout v2.0 (ten sections): the three standing sections on FARAJA, IRGC/Basij and supply chain — which had repeated unchanged content since Day 39 — are replaced by one “Erosion” section that reports only what changed. The section list below is reconstructed from the report-building sessions and must be confirmed against the Day 70+ master prompt before publication: strategic landscape and watchpoint; key events (strict 24-hour window); cumulative casualties; variable changes; Monte Carlo (histogram, tornado, 12-week scenarios); threshold model; erosion; time trend; limitations; strategic conclusion (72-hour forecast).

Reports are produced in Persian and English (HTML and PDF) from a locked master prompt with a three-step workflow: briefing and initial search; deep erosion search; build.

10Sources and Transparency

Military: CENTCOM, IDF Spokesperson, JINSA, Critical Threats (AEI), FDD, ISW OSINT: OSINT613, Bellingcat, Maxar/Planet satellite imagery Media: CNN, Al Jazeera, NPR, Washington Post, Reuters, Axios, AP Persian-language: Iran International FA, Vahid Online, Mamlekate Human rights: HRANA, NCRI, Amnesty International, IHRNGO Regime official (inverse reading): Fars, Tasnim, IRNA Iranian officials on the record (economic erosion, from Day 77): government spokesperson, deputy ministers, chamber-of-commerce and industry-association officials, cited by name and date.

Evidence tiers used for scoring: Tier 1 — official confirmation (CENTCOM/IDF) or satellite imagery; Tier 2 — two or more independent OSINT or think-tank sources; Tier 3 — single report or inverse reading of regime media. A score resting only on Tier 3 evidence is flagged in the report.

Discount on parties’ claims: a quantitative claim made by a belligerent about its own performance or its adversary’s losses (an IDF launcher count, an IRGC “downed aircraft” claim, a government damage total) is entered into the dossier at 0.75 of the stated figure unless a Tier-1 source independent of that party confirms it. The daily reports applied this factor from the start; it is recorded here for the first time. It applies to figures used in MCI sub-indicator arithmetic. It is not a score adjustment and is never applied to a GFI anchor placement.

11Methodological Limitations

1. Correlations are judgement, not data. The 22 ρ values are analytical. The matrix they form was not positive semi-definite and had to be repaired; the repair moved four pairs by 0.06–0.11 (§6). A Gaussian copula carries no tail dependence; a t-copula (df = 4) was tested at Day 77 and changed P(≥ 65) by under one point, so the Gaussian form is retained.
2. ISAF dominates. One sd of ISAF outweighs any single variable threefold. The index is, in practice, as certain as ISAF is.
3. Threshold 65 is an analytical estimate.
4. Partial double-counting between missile production and TELs, and between IRGC cohesion and internal security; the correlation matrix models it but does not remove it.
5. Weights rest on analytical judgement. AHP or Delphi validation is still open.
6. The GFI-R line has four points and no daily resolution. Between anchor days the rubric line is undefined; nothing in it should be read as a daily trend. Two of its six variables (IRGC cohesion, social base) carried a thin-evidence flag in every run and moved by up to 15 between raters.
7. “Oil revenue zero” was an inference, not a measurement. The March reports’ line is refuted by Tier-1 tracking (§7). It affected two things: the war-economy variable in GFI-D, which is not restated (§2a), and the third necessary condition of the threshold model, which is redefined and its record corrected (§7). The GFI-R raters had already weighted Kpler; no GFI-R value changes.
8. Days 46–100 were simulated without the copula. Point RDI values are unaffected; interval widths and P(≥ 65) for those days are understated (§6) and are not restated retroactively.
9. Persian-source gap. Web search does not fully cover Telegram-based sources.
10. v3.2 did not describe the model that produced the series. It printed ISAF on the whole bracket at 0.86, a copula that was not running, and σ values that were not in use. This document corrects all three.
11. Rater B’s IRGC-cohesion evidence ends around Day 46. The cumulative dossier for Days 47–76 is thin on that variable, and the Day 77 GFI-R score for it (40) carries a thin-evidence flag. A fuller dossier may move it either way.

12Regression Case — Day 77 (15 May 2026)

Inputs: missile production 93, TEL launchers 73, air defense 90, drones 73, naval 80, proxy network 51, territorial control 37, internal security 63; central command 96, IRGC cohesion 75, war economy 98, information control 78, social base 71, religious legitimacy 58.

Expected: MCI_raw 72.30; MCI_adj 57.84; GFI 79.33; point RDI 63.45. Monte Carlo v4.1 (seed 20260910): mean 63.18, median 63.25, sd 2.58, 90 % interval [58.9, 67.3], P(≥ 65) 24.3 %.

GFI-R at Day 77 (blind rater, cumulative dossier, 11 September 2026): central command 60, IRGC cohesion 45, war economy 65, information control 40, social base 60, religious legitimacy 55. GFI-R 54.17; point RDI 49.74; Monte Carlo mean 49.73, sd 2.4, 90 % interval [45.8, 53.6], P(≥ 65) 0.0 %; distance 15.3.

GFI-R at Days 30, 46 and 100 (same procedure, 12 September 2026): Day 30 — 60, 50, 65, 45, 55, 50 (GFI-R 54.17, point RDI 49.39 with MCI 94, 76, 92, 85, 76, 48, 25, 52); Day 46 — 65, 55, 45, 45, 55, 50 (52.50, RDI 48.37 with MCI 93, 73, 90, 73, 78, 49, 35, 58); Day 100 — 60, 40, 60, 50, 40, 50 (50.00, RDI 47.47 with Day 77 MCI). Day 77 from the master dossier — 60, 40, 60, 55, 50, 45 (51.67, RDI 48.38); the Part 3 result above (54.17 / 49.74) is the first run and is superseded by this one.

All sets are regression cases. Any change to rdi_model.py that moves the point RDI, or moves the Monte Carlo figures by more than sampling noise without a documented model change, is a defect.

13Version History

Theoretical foundations

Paul Collier (state-collapse threshold), Barbara Walter (governance indicators and civil-war prediction), Jack Goldstone (interactive state-instability formulation), Homa Katouzian and Ervand Abrahamian (Iran-specific calibration), Roger Nelsen (copula theory).