# P(DOOM) INDEX — METHODOLOGY v0.1

Governing methodology for pdoomcount.org. Dated 2026-09-16. Clinical register:
this document is a protocol, not an argument for a conclusion.

v0.1 integrates `docs/MISSION.md`, the locked forks in
`docs/METHODOLOGY-DRAFT-v0.md`, the verified seed ledger in
`docs/RESEARCH-PASS-v1.md`, and the adopted findings in `docs/CRITIQUE-R1.md` /
`docs/CRITIQUE-R1-VERDICTS.md`. The v0 draft remains in the repo as history; do
not build from it.

Companion extract: `docs/SOURCE-LEDGER-TIER-A.md` (verified values plus v0.1
scoring columns).

**This document does not publish a live index number.** It states how a number
is computed. Static snapshot v1.0 (as-of 2026-09-16) lives in
`docs/SNAPSHOT-v1.md` and `snapshots/v1.json`; the frontpage mockup renders
that snapshot. The design-preview figures that used to sit on the mockup
(headline 4.7%, G2 1.2%, G3 3.9%, conditions ×1.08) were **ILLUSTRATIVE** and
are not v0.1 outputs.

---

## 0. Document control

| Field | Value |
|---|---|
| Version | Methodology v0.1 |
| Rubric version | R0.1 (rule IDs in §12) |
| Status | Governing for any static v1 snapshot compute |
| Forks | Locked — see §1. Do not reopen in this version. |
| Snapshot model | Static v1. Automation is v2. |
| Hash stamp | R0.1 freeze: commit `051565a`, blob `0fae6bbcc552eec8ce70b36b0ecbab2c5a1c85dd` (§12.8). Rule-ID rows cite **R0.1**. A later edit to §11–§13 requires **R0.2** and a visible discontinuity. |

Site judgments in this text are labeled **[SJ-…]**. Verified source facts are
labeled **[VF]** and point at the research pass. Mixing the two unlabeled is a
protocol error.

---

## 1. Locked forks (do not reopen)

Ruled 2026-09-15 by the operator. The critique attacked M1 as instructed;
amendments fold **around** these forks.

1. **Headline gauge = G1 catastrophe** [LOCK: G1]. G2 extinction is a displayed
   sub-gauge. G3 is displayed and is not folded silently into G1.
2. **News mechanism = M1 bounded conditioning multiplier** [LOCK: M1], with a
   public event ledger, per-category bounds, and movement caps. v0.1’s work is
   to make that multiplier attributable, bounded against forecaster
   disagreement, quantized, and symmetrically falsifiable.
3. **Source universe includes high outliers** [LOCK: include]. They are shown
   individually. UNVERIFIED or paraphrased outliers stay in the ledger and
   **do not score** (§5). Inclusion in the universe is not inclusion in the
   statistic.
4. **v1 ships as a static, fully-documented snapshot** [LOCK: static]. The live
   pipeline is a v2 work item. Front-page language says snapshot, as-of date,
   next refresh date.

Also ratified via MISSION and unchanged: public site, no ads, no monetized
fear, clinical language.

---

## 2. What the site publishes

An **index of published judgments, conditioned by a documented news rubric**.
Not a measurement. The front page says so.

Published surfaces, each with uncertainty:

- **G1 headline** — point, source band, conditioning envelope, unconditioned
  aggregate, conditions comparison against forecaster disagreement, down-move
  eligibility.
- **G2** — same machinery, own mapping, own minimum-n outcome.
- **G3** — labeled constructed estimate (§4.3). Never presented as a
  conditioned aggregate of published forecasts.
- **Source ledger** — every candidate, scoring or not.
- **Event ledger** — every multiplier movement, including zeros, clips,
  decays, resets, and down-triggers.
- **Corrections log**.
- **Method page** — this protocol, leave-one-out table, weighting-scheme range,
  mapping table, genesis row.

Why this shape: the chance that AI causes a terminal catastrophe cannot be
observed. A single author’s number is one credence. What can be built honestly
is a reproducible aggregation of named estimates, with every site judgment
written down, and with current conditions applied only inside a bound that is
smaller than the disagreement already on the record.

---

## 3. Honesty rules, operationalized

From MISSION, made checkable:

1. **Index, not measurement** — front page sentence, permanent.
2. **Attributable** — every scored value cites a ledger ID; every movement
   cites an event ID and a **rule ID** from R0.1; every site judgment has an
   `[SJ-…]` tag.
3. **Uncertainty always displayed** — a point with no band is not publishable.
   The band’s components are labeled. What the band excludes is stated (§9.3).
4. **Symmetric falsifiability** — “what would move this down” is not a caption.
   Numeric down-triggers in §13 have the same force as up-steps. A
   **down-move eligibility** line sits beside the number.
5. **No bombast** — clinical language; absolute percentage-point deltas;
   resolution floor (§14.3).
6. **Mundane-deployment path is first-class** — G3 is always on the page;
   categories C2 and C7 exist; they are capped so they cannot dominate the
   multiplier (§4.4, §11.4).

---

## 4. The gauges

All gauges are AI-caused. Horizon for G1 and G2 mixtures: **2100**, with
as-stated values shown beside any standardized use (§6.2).

### 4.1 G1 — Catastrophe (headline) [LOCK: G1]

An AI-caused event that kills **≥10% of humanity** or **irreversibly collapses
industrial civilization**, by 2100.

This is broader than literal extinction. It is meant to capture both an
alignment-lab failure and a mundane deployment cascade that ends in collapse.
Those two routes are not extra gauges inside G1; they are covered by the
definition. G3 exists so the second route is not only a clause in G1.

### 4.2 G2 — Extinction

Literal human extinction from AI, or the FRI operationalization where it is
the source’s own definition: **global population < 5,000**. Horizon for the
mixture: 2100. Near-term extinction statements are a separate panel (§6.2).

### 4.3 G3 — Deployment cascade (constructed estimate)

**[SJ-G3-CONSTRUCT]** G3 is the founding thesis as its own gauge: premature
deployment of insufficiently-verified AI into critical systems (grid, military
command, health, finance, logistics), producing compounding failures that
neither market nor regulator catches in time. Collapse-shaped, not
extinction-shaped by necessity. Always displayed. Never folded silently into
G1.

The verified Tier A table contains **no published forecast of this event**
**[VF: RESEARCH-PASS-v1]**. Therefore G3 has **n = 0** scoring sources.

**v0.1 rendering rule:** G3 is a **labeled constructed estimate**, not a
conditioned aggregate. It must not look like G1/G2.

Allowed on the G3 surface:

- a short scenario narrative;
- a Carlsmith-style factor list (deployment into named system classes;
  verification lag; liability lag; compounding; terminality), each qualitative
  or sourced as *indicator*, not as P(G3);
- pointers to event-ledger rows in C2 and C7, which instrument the same thesis
  as *conditions*, not as a G3 probability.

Not allowed on the G3 surface:

- a percentage presented as an aggregate of published forecasts;
- `headline_G3 = aggregate(TierA, TierB) × conditioning(TierC)`;
- silent reuse of the G1 number.

If a later version shows a G3 percentage, the label
**SITE-CONSTRUCTED, NOT AN AGGREGATE** is mandatory, the construction recipe
is a versioned appendix, and n=0 remains on the page until a primary numeric
forecast of this definition exists. The mockup’s **3.9%** is illustrative only
and is not that recipe.

### 4.4 Overlap (G1 collapse limb, G3, C2, C7)

**[SJ-OVERLAP]** The mundane-deployment thesis appears in three places by
design: G1’s collapse limb, G3’s whole definition, and scoring categories C2
and C7. That is overweighting if left unnamed.

Controls:

- G3 is not a second G1 number (section 4.3).
- C2+C7 together may use at most **2/7 of the overall conditioning log-span**
  (§11.4), equal to two categories’ share of seven, not more.
- G1’s mixture is published forecasts of catastrophe/existential catastrophe,
  not a constructed G3.

The thesis is first-class. It is not allowed to count three times as if it
were three independent quantitative inputs.

---

## 5. Scoring eligibility — primary numeric utterances

**[SJ-UTTER]** Only a **primary numeric utterance** scores. A primary numeric
utterance is a percentage, odds, or clearly numeric interval spoken or written
by the named person or instrument, in a citable primary (book, paper, survey
table, recorded interview, dated post).

Excluded from every statistic, retained in the ledger with flags:

- third-party conversions (LeCun’s asteroid comparison converted to <0.01%
  **[VF]**);
- UNVERIFIED high percentages without a pinned primary quote (Yudkowsky
  **[VF]**);
- qualitative institutional language with no number (International AI Safety
  Report **[VF]**);
- a different question treated as if it were this one (Hinton “50%”
  good-vs-bad **[VF]**).

LOCK include-outliers means those rows are **shown**, not deleted. It does not
mean a journalist’s paraphrase becomes a data point.

Interval utterances (Hinton “10% to 20%”) score as intervals, not as a silently
chosen midpoint. Bound utterances (Hubinger “>10%”) score as bounds, not as
10%. Carlsmith’s later “>10%” is a bound companion to the main-text product
~5%; the mixture point for that row is the main-text **5%** **[SJ-CARL-POINT]**,
because it is the only exact product; the bound is displayed beside it.

---

## 6. Normalization

Every source probability is mapped to a gauge **before** it enters a mixture.
Where the definition does not map cleanly, the row stays in the ledger with a
reason and is excluded from that mixture.

### 6.1 Gauge mapping

**[SJ-MAP]** Mapping is a site judgment. It can move the headline more than a
typical source. It is therefore a published table with decision-maker (this
document), an ambiguity flag, and a sensitivity line — not an implicit
cleanup.

| Rule | Native G1 (catastrophe / existential catastrophe / collapse-or-worse) | Native G2 (literal extinction / FRI <5,000) | Ambiguous | Does not score |
|---|---|---|---|---|
| Examples | XPT catastrophe 2.13% / 12%; Carlsmith existential catastrophe; Ord existential risk from unaligned AI; AI Impacts extinction-*or*-severe-disempowerment (flagged) | XPT extinction 0.38% / 3%; Hinton human extinction; Hubinger “kill all humans” | Bengio “catastrophic” / p(doom) with no operationalization | LeCun conversion; Yudkowsky UNVERIFIED |
| Mixture | May enter **G1-2100** if horizon class allows | May enter **G2-2100** if horizon class allows | Ledger; not in 2100 mixtures until a later mapping version | Ledger only |

**No face-value dump of G2 numbers into G1.** Extinction ⊆ catastrophe, so
P(G1) ≥ P(G2) for nested events in one mind, but mixing Hinton’s 10–20%
*extinction* with XPT’s 12% *catastrophe* as if they named the same event is
the inflation the critique named.

**Published G2→G1 factor (sensitivity only):** XPT is the only verified
instrument that reports both definitions on one panel **[VF]**.

- Superforecasters: 2.13 / 0.38 ≈ **5.61**
- Domain experts: 12 / 3 = **4.00**

**[SJ-M-FACTOR]** v0.1 records **m = 4**, the domain-expert ratio, as the
sensitivity factor. It is **not** applied to the headline. The required
sensitivity line is: “G1 if native-G2 rows were up-mapped by m=4” versus “G1
if native-G2 rows were included at face value” versus **the actual headline
(native-G1 rows only)**.

Per-source mapping lives in `docs/SOURCE-LEDGER-TIER-A.md`. Changing a map is
a methodology version bump.

### 6.2 Horizon standardization

Verified horizons **[VF]**: Hubinger 10y, Hinton ~30y, Carlsmith ~2070, XPT
2100, Ord ~100y, AI Impacts open / 100y variant, Bengio unspecified. They are
not all “2100.”

**[SJ-HORIZON]** v0.1 does **not** invent a hazard rate. Converting 10-year
probabilities into 2100 probabilities would be a larger judgment than most
sources. Two columns are always shown: **as-stated** and **mixture-eligible**.

Horizon classes:

| Class | Rule | Mixture |
|---|---|---|
| **H-century** | Source horizon ∈ [2070, 2100], or “this century” / 100 years, or a 100-year variant of the same instrument | Eligible for the matching 2100 gauge |
| **H-near** | Source horizon < 50 years | **Excluded** from 2100 mixtures. Shown on a near-term panel as-stated. Inequality note: for an absorbing catastrophe, P(by t_near) ≤ P(by 2100); v0.1 does not use that inequality to impute a 2100 point. |
| **H-unspecified** | No calendar horizon | Excluded from 2100 mixtures. Ledger only. |

As-stated and mixture-eligible values sit side by side on the method page.
This standardization rule is **[SJ-HORIZON]**, not a source claim.

### 6.3 Minimum-n

**[SJ-MIN-N]**

| n scoring rows mapped to the gauge and horizon class | Display |
|---|---|
| n ≥ 3 | Full aggregate (headline statistic + band) |
| n = 2 | **Thin panel**: both numbers and an unweighted midpoint, labeled “thin panel, not a full aggregate.” No IQR presented as if it were a sample spread. |
| n < 2 | **Insufficient sources.** No point estimate. |
| n = 0 published forecasts | G3 path: constructed estimate (§4.3). |

v0.1 expected membership (not a computed headline): G1-2100 has five scoring
rows (A-XPT-SF catastrophe, A-XPT-DE catastrophe, A-CARL-2022, A-ORD-2020,
A-AII-2024) → full aggregate eligible. G2-2100 has two scoring rows (the two
XPT extinction medians) → thin panel. G3 has zero → constructed.

---

## 7. Tiers

### 7.1 Tier A — published forecasts (backbone)

Named, dated, cited. Seed: `docs/SOURCE-LEDGER-TIER-A.md`. Each scoring row
carries: value, gauge map, horizon class, family, weight, COI note,
utterance type.

Weights are published here and **fixed between quarterly reviews**. They are
not tuned to hit a preferred headline.

### 7.2 Tier B — live aggregates (slot; not in the v1 snapshot)

Metaculus community questions in the AI catastrophe/extinction family, and
prediction markets where active. The research pass left question IDs and
values **UNVERIFIED**.

**[SJ-TIERB-V1]** Static v1 **excludes** Tier B from the mixture. The slot
remains in the protocol. When a later version verifies IDs, each series needs:
question ID, operational definition, horizon, liquidity/volume, and a
**two-tier flag** if it overlaps Tier A people or questions.

Markets (play-money or thin books) are **out of v1**. A future recency
function, if any, must be numeric and versioned; v0.1 defines none, so none
is applied.

### 7.3 Tier C — news-conditioning rubric (M1)

A fixed set of indicator categories. Events move G1/G2 **only** through
pre-declared rules, quantized steps, and bounds. G3 does not receive this
product as if it were an aggregate.

Categories (R0.1 — seven, same list as the draft, now with declared sign and
precedence):

| ID | Category | Sign (declared) | Role |
|---|---|---|---|
| C1 | Frontier capability leaps | Up when an *external* instrument clears a leap threshold. Vendor-only claims do not score. | Capability |
| C2 | Deployment into sensitive systems | Up for premature/sensitive adoption; down for documented rollback | Deployment-cascade engine (capped with C7) |
| C3 | Governance / regime change | Down for binding constraints; up for weakening or collapse of constraints. Lab press releases are not “governance.” | Governance |
| C4 | Safety / interpretability research | Down for independent replicated results; up for independent stall. Lab safety *spend* alone does not score (ambiguous). | Safety |
| C5 | Racing dynamics | Up for verified race intensification; down for verified coordination that reduces the race | Race |
| C6 | Incident record | Up for a qualifying incident; down for a declared clear window | Incidents |
| C7 | Autonomy dependency | Up for measured handoff of critical operations without human-in-loop verification; down for restored verification | Deployment-cascade engine (capped with C2) |

C2 and C7 are the founding thesis’s live instrumentation. They are not a
third G3 probability.

---

## 8. Weights

### 8.1 Class weights and the survey–superforecaster tie

Draft v0 loaded “surveys and superforecasters” as the heaviest classes. Those
classes disagree by a large factor (XPT superforecasters 0.38% extinction vs
domain experts 3% vs surveys 5–10% **[VF]**). Whichever intra-pair tie-break
wins dominates the headline.

**[SJ-TIE]** v0.1 sets **equal class weight** for the current AI Impacts wave
and the XPT family as a whole (both **3.0**). That is the tie-break: equality.
Residual disagreement is shown as a **weighting-scheme range**, not absorbed
into a silent winner. Calibration is **not** used as a cross-task license for
the superforecaster panel to dominate; both XPT panels answered the same
AI questions in the same tournament, and that is the comparability note.

| Class | Class weight | Task-comparability note |
|---|---|---|
| Broad expert survey (current wave) | 3.0 | Many respondents; question is extinction-or-disempowerment, not FRI catastrophe. |
| XPT family (both panels combined) | 3.0 | Operational FRI definitions; two panels, one tournament. |
| Structured analysis | 2.0 | Long-form credence; not a population median. |
| Named individual statement | 1.0 | One mind; modest weight; still visible because outliers are included. |

### 8.2 Instrument-family caps

**[SJ-FAM]** Correlated rows are not independent.

- Combined weight inside a family **≤** the class weight of that family.
- FAM-AII: 2024 in the mixture at 3.0; 2023 weight 0.0 in the mixture;
  2023 remains for substitute-wave sensitivity.
- FAM-XPT: superforecasters 1.5 + domain experts 1.5 = 3.0.
- A row that also appears in Tier B (later) is flagged two-tier and does not
  collect both weights.

Per-source mixture weights: see the ledger. Sum of G1-2100 mixture weights
under v0.1: 1.5+1.5+2.0+2.0+3.0 = **10.0**.

### 8.3 Named-statement line (required)

**[SJ-NAMED-LINE]** Named statements can sit 2–5× above the survey class and
pull a median even at small weight. The method page always shows:

- headline **with** the named-statement class (for 2100: that class is empty
  under §6.2, so the line reads “no named-statement rows in this mixture”);
- headline **without** that class;
- the **difference in percentage points**.

Near-term named statements (Hinton, Hubinger) appear on the near-term panel,
not as a quiet tug on G1-2100.

---

## 9. Aggregation

Per gauge that meets minimum-n:

```
unconditioned = W( {scoring rows mapped to this gauge and horizon class} )
headline      = unconditioned × conditioning
conditioning  = clip( exp( Σ_c log m_c ) )   # §11
```

`W` is defined next. The unconditioned aggregate is a **permanent second
displayed line**. Conditioning is its own line. A reader can always separate
the expert baseline from the news adjustment.

### 9.1 Headline statistic

**[SJ-STAT]** The headline is the **untrimmed weighted median** of
mixture-eligible values, weights as in §8.

- **Untrimmed** because LOCK include-outliers forbids quietly trimming the
  same outliers the site promised to keep. There is no trim rule.
- The **weighted mean** is shown beside the median. It is not the headline.
- “Which is the number?” — the untrimmed weighted median. The mockup’s
  “trimmed weighted median” caption is illustrative and predates this rule;
  it must be corrected on the live page at snapshot build, not by relabeling
  the mockup in this docs pass.

Weighted median: order the scoring values; the headline is the smallest value
for which cumulative weight is at least half the total weight (standard
weighted-median convention). Ties and even splits: the midpoint of the two
central weighted observations. State the convention on the compute printout.

### 9.2 Leave-one-out and other required sensitivities

On the method page, every published snapshot includes:

1. **Leave-one-out:** headline with each scoring row removed, one at a time.
2. **Substitute-wave:** G1 with A-AII-2023 (5%) in place of A-AII-2024 (10.0%).
3. **Weighting-scheme range:** v0.1 weights; equal weights; survey-only;
   XPT-only; structured-only.
4. **Mapping lines:** native-G1 only (headline); native-G2 at face value
   added; native-G2 up-mapped by m=4 added.
5. **Named-statement on/off** (§8.3).
6. **Unconditioned vs conditioned.**

Small n makes the weighted median jumpy. Leave-one-out is how that is shown
instead of pretended away.

### 9.3 The band

**[SJ-BAND]** The source band is computed **only over same-gauge-mapped,
mixture-eligible rows**. It is not an IQR dumped across extinction,
disempowerment, and catastrophe as if they were one variable.

| n in the mixture | Source-spread display |
|---|---|
| n ≥ 4 | Interquartile range of the mixture values (unweighted across rows for the spread display, so a heavy row cannot hide the disagreement). Point remains the weighted median. |
| n = 3 | Full range (min–max), labeled “range, n=3,” not IQR. |
| n = 2 | Thin panel: both points. No band costume. |
| n < 2 | No number. |

**Band composition (labeled):**

- **Included in the source band:** disagreement among mixture-eligible rows
  for that gauge.
- **Not included in the source band:** mapping uncertainty, weighting-scheme
  uncertainty, horizon-standardization uncertainty, conditioning.
- **Conditioning envelope (separate, required):** unconditioned point ×
  overall bound [0.65, 1.54], drawn and labeled so the largest single
  judgment in the system is inside the picture (critique A14).
- **Mapping-induced shift (separate line):** headline versus the face-value
  and m=4 counterfactuals.

After conditioning, the source band is multiplied by the **current**
combined multiplier so the displayed IQR/range moves with conditions. The
envelope uses the **bounds**, not the current multiplier. Both are on the
page.

---

## 10. Conditioning bound, derived from forecaster disagreement

M1 stays. The overall bound is the **primary** judgment. Per-category caps
are derived so their worst-case product equals that overall bound. Decorative
per-category caps that compose to ×4.77 while an overall ×1.6 quietly does
all the work are not R0.1.

### 10.1 Reference disagreement

**[VF]** Same instrument, same PDF, same horizon, two definitions:

| | Superforecasters | Domain experts | Ratio D |
|---|---|---|---|
| AI catastrophe (G1-matched) | 2.13% | 12% | 12 / 2.13 ≈ **5.63** |
| AI extinction (G2-matched) | 0.38% | 3% | 3 / 0.38 ≈ **7.89** |

v0.1 uses the **G1-matched** pair as the reference for the G1 multiplier,
because G1 is the headline. D_ref = 5.63.

A ~4% baseline moved by the draft’s 0.7–1.6 span becomes 2.8–6.4% — a swing
on the same order, in percentage points, as the 0.38–3 extinction dispute.
The site’s job is to aggregate other people’s judgments. News may update for
recency; it may not replace that dispute.

### 10.2 Derivation **[SJ-BOUND]**

Principle: the **log-span** of the overall conditioning bound is at most
**half** the log-span of D_ref.

```
D_ref     = 12 / 2.13 ≈ 5.6338
S_max     = √D_ref    ≈ 2.3736
```

Log-symmetric about 1 (equal up and down in ratio; critique A3):

```
H = √S_max ≈ 1.5406   →  rounded 1.54
L = 1/H    ≈ 0.6491   →  rounded 0.65
overall multiplier ∈ [0.65, 1.54]
span H/L              ≈ 2.37
```

Rounding to two decimals is **[SJ-BOUND-ROUND]**. Asymmetry 25% up / 15% down
is withdrawn. If a later version wants alarm-side asymmetry, it needs a named
justification and a rubric bump; v0.1 has none.

### 10.3 Permanent comparison line

Beside the number, always:

> Conditions currently multiply the aggregate by **×M** (span allowed:
> ×0.65–×1.54, ratio 2.37). Reference disagreement (XPT catastrophe,
> superforecasters vs domain experts) is **5.63×**. Conditions move the
> number **less / comparably / more** than that disagreement:
> compare |M − 1| against (√D_ref − 1), and compare M to D_ref.

Worked wording at M = 1.00: “conditions currently move this **less** than the
forecaster disagreement (neutral multiplier; disagreement remains 5.63×).”

This line is not decorative. If a snapshot cannot print it, the snapshot is
not publishable.

### 10.4 Per-category bounds, derived from the overall bound

Seven categories. Worst-case product equals the overall bound:

```
1.54^(1/7) ≈ 1.0636
0.65^(1/7) ≈ 0.9403
```

Per-category multiplier ∈ **[0.94, 1.06]** before quantization. The overall
clip remains primary if arithmetic ever exceeds it. **Clip order** when the
overall bound binds: categories are listed by descending |log m_c|; the
suppressed log-amount is written on the event row.

The draft’s example steps 1.10, 1.15, 1.20, 1.25 belonged to a decorative
1.25 per-category cap. They are **not** R0.1 steps. Using them would make
the overall bound the only real constraint again.

---

## 11. Multiplier mechanics

### 11.1 Quantized steps **[SJ-QUANT]**

Default is **no movement** (1.00). Allowed per-category values:

```
{0.94, 0.97, 1.00, 1.03, 1.06}
```

There is no 1.17. Intermediate wishes round **toward 1.00** (the default).
Evidence needed for each step is in §12.5. Worst-case 1.06⁷ ≈ 1.50, inside
the 1.54 overall cap; 0.94⁷ ≈ 0.65.

### 11.2 Combination: log-sum, then one cap (order-independent)

**[SJ-LOGSUM]** Do not multiply in an event-order sequence.

```
M_raw = exp( Σ_{c=1..7} log m_c )
M     = clip( M_raw, 0.65, 1.54 )
```

If a clip hits, ledger the suppressed amount `log(M_raw) − log(M)` and the
clip-order list. Application order of *events* therefore cannot change M
except via timestamps that the decay function already uses.

### 11.3 Cap precedence

Numeric, in this order (later binds harder):

1. Quantized step on the event (§11.1).
2. C2+C7 share cap (§11.4).
3. Per-quarter cap: combined M may not move by more than a **log-symmetric
   factor 1.15** within a calendar quarter (up to ×1.15 or down to ×1/1.15
   from the quarter-open M). **[SJ-QCAP]**
4. Overall bound [0.65, 1.54].

Whichever cap truncates, the row logs the **suppressed amount**. Mid-quarter
reviews use the event’s **effective date** (publication date of the primary
source), not the date a human typed the row, so logging delay is not an
arbitrage.

### 11.4 C2+C7 share cap

**[SJ-C27]** Let L_all = log(1.54/0.65). Then

```
|log m_C2 + log m_C7|  ≤  (2/7) × L_all  ≈ 0.246
```

so m_C2 × m_C7 stays inside about **[0.78, 1.28]**. Excess is clipped toward
1.00 on C2 and C7 equally in log space, then logged. This is the numeric
control on thesis overweight inside M1.

### 11.5 Reference period and the meaning of 1.00

**[SJ-REF]** Neutral (all m_c = 1.00, M = 1.00) means: *no news-conditioning
relative to the information environment of the scored Tier A set as a whole.*

Reference window: **2020-01-01 through 2026-09-15** (Ord through the research
pass). Events inside that window do not score unless they are clearly *after*
the relevant source’s own information set *and* still news relative to the
set median date; v1 snapshot default is **M = 1.00 at genesis** with an empty
qualifying event list, unless the genesis pass explicitly scores post-window
or post-median events. The unconditioned aggregate is always shown, so
“neutral” cannot hide the baseline.

### 11.6 Decay and reversion

**[SJ-DECAY]** Multipliers ratchet without a decay rule. Each category reverts
toward 1.00 in log space with **half-life 180 days**, floor/ceiling 1.00
(decay never crosses neutral). Effective daily:

```
log m(t) = log m(t0) × (1/2)^{Δt / 180d}
```

then re-quantized toward 1.00 onto the allowed set. Decay emits a ledger row
on each snapshot date (static v1: one decay-to-date row at compute, not a
theater of daily ticks).

### 11.7 Reset on Tier A refresh

A Tier A refresh **does not** silently reset M to 1.00. If a reset is
intended (new reference window), it is an event row **RESET-*** citing
**R-RESET**, with before/after M and a new [SJ-REF] window. No reset, no row,
no change.

### 11.8 External instruments for C1; vendor-only bar

**[SJ-C1-EXT]** C1 scores only if an **external** instrument moves:

- a named independent task-horizon series (e.g. METR-style time-horizon)
  **doubles** relative to the last scored C1 baseline, or
- a named independent benchmark family shows a comparable leap, replicated
  or reported outside the vendor’s own blog.

Vendor self-reports, demo videos, and un-audited evals **do not score**. They
may be non-scoring annotations.

---

## 12. Rubric R0.1 — frozen rulebook

This section is the versioned rulebook the draft lacked. Ledger rows cite a
**rule ID**. Changing eligibility, precedence, signs, steps, or down-triggers
requires **R0.2** and a visible history discontinuity on the method page.

### 12.1 Eligibility checklist (every event)

An event scores only if **all** of the following hold. Failed checks produce
a **non-scoring** ledger row (still public) with the failed item named.

| Check ID | Requirement |
|---|---|
| E1 | **Primary source**: first-party document, official dataset, statute, or named-instrument series. Not a roundup of another outlet’s roundup. |
| E2 | **Dated**: publication or occurrence timestamp. |
| E3 | **One fact, one row**: five outlets reporting one fact are citations on one row, not five deltas. Split/merge: if later reporting shows two facts, split with a correction; if two rows were one fact, merge and retract the extra delta. |
| E4 | **Quantified or thresholded**: meets a named threshold in §12.5. Vibes do not score. |
| E5 | **Single category**: exactly one scoring category via §12.2. Other categories may be annotations (non-scoring). |
| E6 | **Rule ID**: the row names the R0.1 rule that authorizes the step. |
| E7 | **Not vendor-only** if the category is C1. |
| E8 | **Rubric version** R0.1 (or a later frozen ID). |

### 12.2 Precedence (one event, one category)

First matching wins. Lower items become annotations.

1. **C6** if the event is an incident (harm, near-miss, or documented failure
   in deployment or evaluation at the severity threshold).
2. **C2** if it is a deployment into a named sensitive system and not already
   scored as C6.
3. **C7** if it is a measured autonomy-without-verification share/handoff and
   not already C2/C6.
4. **C1** if it is a capability leap on an external instrument.
5. **C5** if it is racing (compute buildout, alliance fracture, Manhattan-style
   program) and not already C1.
6. **C3** if it is governance/regime change.
7. **C4** if it is safety/interpretability research results.

### 12.3 Direction (sign)

Declared below. A sign flip is a rubric version bump, not a clever row.

| Cat | Up | Down |
|---|---|---|
| C1 | External leap threshold met | Not from a failed demo. Down via decay or D-triggers elsewhere. |
| C2 | Premature/sensitive deployment documented | Documented rollback / pause covering the same class |
| C3 | Constraints weaken or die | Binding constraints in force (see D-GOV-TREATY) |
| C4 | Independent stall or replicated backslide | Independent replicated safety/interpretability result |
| C5 | Verified race intensification | Verified coordination that reduces race |
| C6 | Qualifying incident | Clear window (D-INC-CLEAR) |
| C7 | Measured increase in unverified autonomy | Restored human-in-loop requirement |

### 12.4 Rule IDs (movement)

| Rule ID | Effect |
|---|---|
| R-NULL | Eligibility failed or default; m_c stays 1.00 relative to last state (no step). |
| R-UP-1 | One up-step: 1.00→1.03, 1.03→1.06, already 1.06→clip, log suppressed. |
| R-UP-2 | Two up-steps from 1.00 to 1.06 when §12.5 “strong” clause is met. |
| R-DN-1 | One down-step, symmetric magnitudes. |
| R-DN-2 | Two down-steps from 1.00 to 0.94 when a strong down-clause or a §13 trigger’s strong form is met. |
| R-DECAY | Half-life reversion toward 1.00. |
| R-CLIP | A cap in §11.3–11.4 truncated the step. |
| R-RESET | Reference-window reset (§11.7). |
| R-D-* | Mechanical down-trigger fired (§13). |

Every scoring row: `event_id, source, date, category, rule_id, delta, m_before, m_after, M_before, M_after, suppressed, rationale`.

### 12.5 Evidence steps (what 1.03 vs 1.06 means)

**[SJ-EVID]** Default R-NULL. The burden is on movement.

| Step | Evidence |
|---|---|
| ±0.03 (R-UP-1 / R-DN-1) | One eligible event; independently confirmable; meets the category’s ordinary threshold (below). |
| ±0.06 (R-UP-2 / R-DN-2) | Ordinary threshold **and** independent confirmation from a second primary, **or** a §13 named trigger. |

Ordinary thresholds (R0.1):

- **C1:** external task-horizon doubling, or equivalent named-benchmark leap
  (§11.8).
- **C2:** documented adoption of frontier-class or safety-critical AI into
  grid, military command, clinical decision, core finance plumbing, or
  logistics control, with verification/liability gap named in the primary.
- **C3:** statute, treaty, binding agency rule, or their repeal/expiry;
  not a blog post.
- **C4:** paper with independent replication, or a pre-registered independent
  replication failure of a previously scored up-result (down would be
  elsewhere if we had scored that result; v1 may have none).
- **C5:** public, quantified compute/alliance move of frontier scale;
  “we are racing” rhetoric does not score.
- **C6:** incident in AIID / OECD / equivalent with severity ≥ **S2**
  (named harm or safety-critical near-miss; S1 chatter is non-scoring).
  Exact S2 codebook is a v1 compute annex; until that annex exists, C6
  does not score upward (down-triggers may still run on the empty clock).
- **C7:** measured share, mandate, or contract language handing a named
  critical operation to AI without human-in-loop verification.

### 12.6 Negative-evidence windows (mechanism, not prose)

Loud events are discrete. Quiet good news is slow. Without clocks, M1 only
moves up.

| Clock | Rule |
|---|---|
| INC-CLOCK | Days since last C6 S2+ incident. At **365 consecutive days**, fire **D-INC-CLEAR** (R-DN-1 on C6). Retriggers every additional 365 clear days, subject to caps. |
| GOV-CLOCK | If no qualifying C3 event (either sign) for 365 days, **no automatic move** (inaction is not a treaty). Governance down-moves require D-GOV-TREATY or an eligible C3 down-event. |
| C1-CLOCK | If no C1 leap for 365 days, R-DECAY continues; no extra down-step. Capability not leaping is not negative evidence of safety. |

The incident-free clock is the one the critique demanded as a mechanical
peer to incident up-moves.

### 12.7 Hash stamp and discontinuity

R0.1 is this §12 plus §11 and §13. After the commit that adds this file:

- record `git hash-object docs/METHODOLOGY-v0.1.md` as **BLOB-R0.1**;
- record the commit SHA as **COMMIT-R0.1**;
- print both on the method page of any snapshot.

Any edit to §11–§13 after that stamp is **R0.2** even if someone forgets to
rename it. Snapshots must not mix event rows from two rubric IDs without a
discontinuity marker on the sparkline.

### 12.8 Stamp

R0.1 rule text is the §11–§13 bytes in commit `051565a` (blob below). This
stamp table is metadata pointing at that freeze; editing only this subsection
does not bump the rubric. Launch freeze on `main` still requires the operator
to accept the stamp.

| | |
|---|---|
| BLOB-R0.1 | `0fae6bbcc552eec8ce70b36b0ecbab2c5a1c85dd` (`git hash-object` of `docs/METHODOLOGY-v0.1.md` at COMMIT-R0.1) |
| COMMIT-R0.1 | `051565a5751dfe3e87cb60298da91c6d3d26b9ea` |

---

## 13. Mechanical down-triggers

Prose “what would move this down” is not a commitment. These triggers fire
**unconditionally** when their predicates are met, with the same quantized
force as up-steps. They do not wait for a human to feel that good news
counts.

Beside the number, a **down-move eligibility** line lists each trigger as
MET / UNMET and the INC-CLOCK value.

| Trigger ID | Predicate | Effect |
|---|---|---|
| **D-GOV-TREATY** | A legally binding international instrument with quantified compute or training-run caps is in force for at least two states that, together, hold an estimated majority of publicly reported frontier training compute. | R-DN-2 on **C3** (0.94 if at 1.00). |
| **D-INC-CLEAR** | INC-CLOCK ≥ 365 days (section 12.6). | R-DN-1 on **C6**. |
| **D-INT-REPL** | A safety or interpretability result that meets C4’s independent-replication clause. | R-DN-1 on **C4**. |
| **D-DEP-PAUSE** | Documented pause, rollback, or restored human-in-loop covering a named critical-infrastructure class across a primary-source set of operators (regulator order or equivalent). | R-DN-1 on **C2** (and C7 annotation). |

Unmet triggers are listed, not implied. At genesis, unless a predicate is
already true on the as-of date, all four read UNMET and INC-CLOCK starts at
the as-of date (or at the last S2+ incident if one is scored that day).

---

## 14. Display requirements

A snapshot that omits these is not a v0.1 snapshot.

### 14.1 Beside G1

1. “Index of judgments, not a measurement.”
2. Point (untrimmed weighted median, conditioned).
3. Source band, labeled, same-gauge only.
4. Conditioning envelope [unconditioned × 0.65, unconditioned × 1.54].
5. Unconditioned aggregate.
6. Current M and the **forecaster-disagreement comparison** (§10.3).
7. Down-move eligibility (§13).
8. Absolute percentage-point delta since last snapshot, not a relative “+12%.”
9. Static snapshot as-of date and **next refresh date**. Automation = v2.

### 14.2 G2 and G3

- G2: thin-panel or aggregate per minimum-n; near-term extinction panel
  as-stated; never a silent G1 copy.
- G3: constructed-estimate treatment (§4.3). If the mockup-style percentage
  is still on a design preview, it stays labeled ILLUSTRATIVE.

### 14.3 Resolution floor **[SJ-RES]**

Material change: **≥ 0.5 percentage points** in the headline. Smaller
absolute moves are recorded in the ledger and labeled **no material change**
on the sparkline. A 5.0% → 5.6% move is **+0.6 pt**, not “up 12%.”

### 14.4 Genesis

v1’s founding number has no predecessor event. **[SJ-GENESIS]** The first
snapshot emits row **GENESIS-001**: unconditioned aggregate, M (default
1.00), rationale (“founding static snapshot under methodology v0.1”), and
day-one down-move eligibility. Subsequent movements cite events. This
document is not GENESIS-001; compute is a later work item.

---

## 15. Cadence and v1 shape

| Tier | v1 (locked static) | Later |
|---|---|---|
| A | Frozen as of the snapshot’s source cutoff | Quarterly refresh + RESET rule if the window moves |
| B | Out of mixture (unverified) | After ID verification; still two-tier flagged |
| C | Event ledger may be empty at genesis (M=1.00) | Event-driven inside this rubric |
| Site | Static files: headline, band, method, both ledgers, JSON of the compute | Scheduled job is v2 |

Tech shape is unchanged in spirit: a versioned compute script writes JSON;
pages render it. This PR does not build that.

---

## 16. Operator prior and site incentive

COI labeling that covers only sources is incomplete.

| Row | Content | Control |
|---|---|---|
| **SITE-PRIOR-001** | Operator thesis, 2026-09-15: the most plausible route is mundane deployment of not-ready AI into sensitive systems faster than verification, liability, or regulation. This thesis motivated G3 and the existence of C2 and C7. | It does not set the G1 headline. C2+C7 log-share cap. G3 is not an aggregate. |
| **SITE-INCENTIVE-001** | The product is more noticeable if the number moves. | Static v1; default R-NULL; quantized steps; resolution floor; public ledger of non-moves. |

Both rows print on the method page.

---

## 17. Corrections and retractions

| Policy | Rule |
|---|---|
| Correction | Timestamped corrections log: what was wrong, what it is now, effect on M and on the headline in percentage points. |
| Retraction | A retracted event reverses its delta (subject to caps), citing the original event ID. The original row stays, marked RETRACTED. |
| Source revision | If a source revises a scored number, the ledger records the movement; the mixture uses the latest primary numeric utterance. Hinton ~10% → 10–20% is the model. |

---

## 18. Non-comparability

**[SJ-NOCOMP]** This index is not Metaculus question N, not a market price,
not a personal p(doom) poll, and not Ord’s or Carlsmith’s number. Definitions,
horizons, weights, and M1 conditioning differ. A comparability note lives on
the method page. Matching a market is not a success criterion.

---

## Appendix A — Critique disposition (R1 → v0.1)

All 41 findings adopted in principle (verdicts). Mapping to this text:

| ID | Sev | v0.1 location |
|---|---|---|
| A1 | HIGH | §10.4, §11.2–11.3 overall primary; per-category derived |
| A2 | HIGH | §10 entire; comparison line §10.3, §14.1 |
| A3 | HIGH | Log-symmetric [0.65, 1.54] §10.2 |
| A4 | HIGH | §11.1 quantized set |
| A5 | HIGH | §11.5 reference window; unconditioned line §9 |
| A6 | HIGH | §12.1 checklist; one-fact-one-row E3 |
| A7 | HIGH | §12.2 one event, one category |
| A8 | HIGH | §12.3 declared signs |
| A9 | HIGH | §12.6 clocks; §13 D-INC-CLEAR |
| A10 | MED | §11.3 numeric caps, effective dates, log space |
| A11 | MED | §11.2 log-sum then single cap |
| A12 | MED | §11.6 half-life 180d |
| A13 | MED | §11.7 RESET row |
| A14 | MED | §9.3 conditioning envelope |
| A15 | MED | §11.8 vendor bar |
| A16 | HIGH | §12 hash-stamped rubric, rule IDs |
| A17 | LOW | §11.3 precedence; suppressed amount |
| B18 | HIGH | §9.1 untrimmed weighted median declared |
| B19 | HIGH | Untrimmed, because outliers are locked in |
| B20 | MED | §9.2 leave-one-out |
| B21 | HIGH | §9.3 same-gauge band |
| B22 | MED | §9.3 composition labels |
| B23 | MED | §8.2 family caps; two-tier flag |
| C24 | HIGH | §4.4, §11.4 |
| C25 | HIGH | §4.3 G3 constructed |
| C26 | HIGH | §6.1 m=4 sensitivity, no naive up-map |
| C27 | HIGH | §6.2 as-stated vs mixture-eligible |
| C28 | MED | §6.3 minimum-n |
| D29 | HIGH | Ledger mapping table; §6.1 |
| D30 | HIGH | §5 primary numeric only |
| E31 | HIGH | §8.1 equal 3.0 tie; weighting-scheme range |
| E32 | HIGH | Task-comparability notes §8.1 and ledger |
| E33 | MED | §8.3 named-statement line |
| E34 | MED | §7.2 Tier B out of v1; no undeclared recency |
| F35 | HIGH | §13 mechanical down-triggers |
| F36 | HIGH | §14.4 GENESIS-001 |
| F37 | MED | §16 site prior and incentive |
| F38 | MED | §14.3 0.5 pt floor; absolute deltas |
| F39 | MED | §18 |
| F40 | LOW | §17 |
| F41 | LOW | §15, §14.1 static snapshot language |

**Three insist-amendments:** (1) §10, (2) §12, (3) §13.

---

## Appendix B — v0.1 mixture membership (no headline computed)

G1-2100 scoring rows (n=5, full aggregate eligible):

| ID | Value used | Weight |
|---|---|---|
| A-XPT-SF | 2.13% | 1.5 |
| A-XPT-DE | 12% | 1.5 |
| A-CARL-2022 | 5% (companion >10% not the point) | 2.0 |
| A-ORD-2020 | 10% | 2.0 |
| A-AII-2024 | 10.0% | 3.0 |

G2-2100 scoring rows (n=2, thin panel): A-XPT-SF 0.38%, A-XPT-DE 3%.

Near-term panel (not mixed): A-HINT-2024 10–20% / ~30y; A-HUB-2026 >10% / 10y.

Ledger only: A-BENG-2023; A-AII-2023 (sensitivity substitute); A-LECUN-2023;
A-YUD.

This appendix is a membership list. Running W(·) is the static v1 compute
in `compute/snapshot.py`; the printout is `docs/SNAPSHOT-v1.md`.

---

## Appendix C — Site-judgment index

| Tag | One-line |
|---|---|
| SJ-G3-CONSTRUCT | G3 is constructed, not an aggregate |
| SJ-OVERLAP | Thesis overweight named and capped |
| SJ-UTTER | Primary numeric utterances only |
| SJ-CARL-POINT | Carlsmith mixture point is 5%, >10% companion |
| SJ-MAP | Mapping table; no G2 face-value dump into G1 |
| SJ-M-FACTOR | m=4 from XPT experts, sensitivity only |
| SJ-HORIZON | No hazard-rate conversion; H-century / H-near / H-unspecified |
| SJ-MIN-N | 3 / 2 / <2 / 0 display rule |
| SJ-TIERB-V1 | Tier B out of static v1 |
| SJ-TIE | Survey class and XPT family both 3.0 |
| SJ-FAM | Family combined-weight caps |
| SJ-NAMED-LINE | With/without named statements |
| SJ-STAT | Untrimmed weighted median is the number |
| SJ-BAND | Same-gauge band; envelope separate |
| SJ-BOUND | Overall [0.65, 1.54] from √D_ref |
| SJ-BOUND-ROUND | Two-decimal rounding |
| SJ-QUANT | {0.94, 0.97, 1.00, 1.03, 1.06} |
| SJ-LOGSUM | Sum of logs, then clip |
| SJ-QCAP | Quarterly factor 1.15 log-symmetric |
| SJ-C27 | C2+C7 ≤ 2/7 of overall log-span |
| SJ-REF | Neutral window 2020-01-01–2026-09-15 |
| SJ-DECAY | 180-day half-life to 1.00 |
| SJ-C1-EXT | External instruments only for C1 |
| SJ-EVID | Default no movement; two step sizes |
| SJ-RES | 0.5 percentage-point material floor |
| SJ-GENESIS | GENESIS-001 required at first snapshot |
| SJ-NOCOMP | Not comparable to markets / Metaculus / personal p(doom) |