# TIER A SOURCE LEDGER — seed extracted from RESEARCH-PASS v1

Governing methodology: `docs/METHODOLOGY-v0.1.md`. Verified values, dates, definitions,
horizons, and URLs are copied from `docs/RESEARCH-PASS-v1.md` and
`docs/critique-input-verified-table.md`. Scoring columns on the right are **v0.1
site judgments**, not source claims.

This ledger does **not** itself publish a headline index number. It records
what is on the record and how v0.1 treats each row. The computed snapshot is
`docs/SNAPSHOT-v1.md` / `snapshots/v1.json`.

Utterance rule (v0.1): only a **primary numeric utterance** may score. Paraphrases,
third-party conversions, and UNVERIFIED percentages remain in the ledger, flagged,
and are excluded from every statistic.

---

## Scoring panel (v0.1 site judgments)

| ID | Source | Date | Verified value (exact) | Definition | Horizon | Canonical URL | Utterance | Family | Gauge map | Horizon class | Mixture | Weight class | Weight | Notes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| A-CARL-2022 | Joe Carlsmith, *Is Power-Seeking AI an Existential Risk?* | Report Apr 2021; arXiv 2022-06-16; May 2022 author update | Existential catastrophe from misaligned power-seeking AI: **~5%** (main-text product of premises 65%·80%·40%·65%·40%·95%); May 2022 author note raises overall to **>10%**. No single unconditional P(misaligned power-seeking AI) stated. | Existential catastrophe via permanent human disempowerment from misaligned APS systems | By ~2070 | https://arxiv.org/pdf/2206.13353.pdf | Primary numeric (point ~5%; later bound >10%) | FAM-CARL | **G1** (existential catastrophe / disempowerment). Not G2 (not literal extinction). Ambiguity flag: definition is disempowerment-shaped, not ≥10% dead. | H-century (2070 ∈ [2070, 2100]) | **In G1-2100**. Point used: **5%** (the only exact product). Companion: **>10%** later overall, shown beside, not substituted as a point. | Structured analysis | 2.0 | Task-comparability: six-premise decomposition; resolution is author credence, not a tournament question. Both figures stay in the ledger. |
| A-AII-2023 | AI Impacts 2023 Expert Survey (~2,778 ML researchers) | Fielded Oct 2023; published 2024-01-04; arXiv:2401.02843 | Median **5%** “extremely bad (e.g. human extinction)” from HLMI; median **5%** extinction-or-permanent-severe-disempowerment; **10%** on that outcome via human inability to control advanced AI; median **5%** within 100 years. | Extinction or similarly permanent and severe disempowerment | Open-ended / one variant 100 years | https://blog.aiimpacts.org/p/2023-ai-survey-of-2778-six-things | Primary numeric (survey medians) | FAM-AII | **G1** with ambiguity flag (extinction *or* disempowerment). Not scored as G2. | H-century (100-year variant exists) | **Out of mixture** (prior wave of FAM-AII; 2024 supersedes). Remains in ledger and in substitute-wave sensitivity. | Survey (superseded wave) | 0.0 in mixture (family cap) | Same instrument family as A-AII-2024. Combined family weight would otherwise double-count one survey program. |
| A-XPT-SF | XPT superforecasters (Karger et al., FRI) | Tournament 2022; PDF 2023-07-10 (v 2023-08-08) | Median **0.38%** AI extinction by 2100; **2.13%** AI catastrophe (≥10% dead in 5 years) | Extinction = population <5,000; catastrophe = ≥10% dead in 5 years, AI-caused | By 2100 | https://forecastingresearch.org/pdf/existential-risk-persuasion-tournament.pdf | Primary numeric (tournament medians) | FAM-XPT | Catastrophe **2.13% → G1** (native G1 match). Extinction **0.38% → G2** (native G2 match). | H-century (2100) | **In G1-2100** at 2.13%. **In G2-2100** at 0.38%. | Superforecaster panel (same tournament as A-XPT-DE) | 1.5 (family split of 3.0) | Closest published operational match to G1/G2 as defined here. Calibration is cross-task, not AI-specific — weight is not a license to dominate FAM-AII. |
| A-XPT-DE | XPT domain experts (Karger et al., FRI) | Same | Median **3%** AI extinction by 2100; **12%** AI catastrophe | Same FRI definitions | By 2100 | Same PDF | Primary numeric (tournament medians) | FAM-XPT | Catastrophe **12% → G1**. Extinction **3% → G2**. | H-century (2100) | **In G1-2100** at 12%. **In G2-2100** at 3%. | Domain-expert tournament panel | 1.5 (family split of 3.0) | Same questions as A-XPT-SF, different people. Do not conflate with “all-expert” total-x-risk headlines (6%/20%) from the same report. |
| A-ORD-2020 | Toby Ord, *The Precipice* (2020) | 2020 | **~10%** (“about one in ten”) existential risk from unaligned AI | Existential risk from unaligned AI | Next 100 years | Primary = the book. Wikipedia cites it; this pass did not re-OCR the page. | Primary numeric (book) | FAM-ORD | **G1** (existential risk includes unrecoverable collapse, not only extinction). Not G2. | H-century (100 years) | **In G1-2100** at 10%. | Structured analysis | 2.0 | Book-length reasoning. Value carried as 10% with a tilde inherited from the source. |
| A-HINT-2024 | Geoffrey Hinton | 2024-12-27 (Guardian/BBC) | **“10% to 20%”** chance AI leads to human extinction (raised from prior ~10%) | Human extinction | ~30 years | https://www.theguardian.com/technology/2024/dec/27/godfather-of-ai-raises-odds-of-the-technology-wiping-out-humanity-over-next-30-years | Primary numeric (interval) | FAM-HINT | **G2** (literal extinction). Not mapped into G1 at face value. | H-near (30 years) | **Out of 2100 mixtures.** Shown on the near-term panel as-stated (10–20%). Prior ~10% recorded as movement, not a second scoring row. | Named statement | 1.0 (near-term panel only) | The p(doom)-wiki “50%” line is a different framing (good-vs-bad outcomes) and is not this figure. Midpoint 15% is **not** scored as if he said 15%. |
| A-BENG-2023 | Yoshua Bengio | 2023-07-15 (ABC interview) | **“around, like, 20 per cent”** catastrophic | Catastrophic AI outcome / p(doom) | Tied to path to dangerous AI; no fixed year | https://www.abc.net.au/news/2023-07-15/whats-your-pdoom-ai-researchers-worry-catastrophe/102591340 | Primary numeric (colloquial point) | FAM-BENG | **G1** with ambiguity flag (not FRI catastrophe; not a horizon). | H-unspecified | **Out of 2100 mixtures** (no calendar horizon). Ledger + named-statement class. Used in the with/without named-statement sensitivity only if a horizon rule is later version-bumped. | Named statement | 1.0 (ledger / named class; 0.0 in 2100 mixture) | Qualifies as a primary numeric utterance; fails horizon standardization for the 2100 gauges. |
| A-LECUN-2023 | Yann LeCun | Clip attributed 2023-12-18 | Verified phrasing: **“Below the chances of an asteroid hitting the earth.”** The **<0.01%** figure is a third-party conversion, not his stated number. | p(doom) | Not specified | https://x.com/liron/status/1736555643384025428 | **Paraphrase / conversion — does not score** | FAM-LECUN | No gauge map for scoring. | — | **Excluded from every statistic.** Ledger row retained, flagged. | — | 0.0 | LOCK include-outliers applies to *his row in the ledger*, not to a converted percentage. |
| A-YUD | Eliezer Yudkowsky | — | **UNVERIFIED** as a primary numeric quote (secondary press cites >95% / >99%; TIME 2023 essay argues near-certain ruin without a clean %). | Existential / civilization-ending AGI | Not pinned | https://web.archive.org/web/20230330001659/https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/ | **UNVERIFIED — does not score** | FAM-YUD | No gauge map for scoring. High-outlier flag. | — | **Excluded from every statistic.** Ledger row retained as an UNVERIFIED high outlier with this caveat. | — | 0.0 | LOCK include-outliers: the row is shown, not dropped. It does not enter the median. |
| A-HUB-2026 | Evan Hubinger | 2026-09-09 (X, reported by BBC) | **“>10% within the next decade”** that AI could “kill all humans” | Human extinction | 10 years from Sep 2026 | https://x.com/EvanHub/status/2097497037956891126 | Primary numeric (lower bound) | FAM-HUB | **G2** (kill all humans). Not mapped into G1 at face value. | H-near (10 years) | **Out of 2100 mixtures.** Near-term panel as-stated (**>10%**, not rewritten as 10%). | Named statement | 1.0 (near-term panel only) | Bound, not a point. Lab affiliation (Anthropic) is a COI label, not a disqualification. |
| A-AII-2024 | AI Impacts 2024 Expert Survey (1,580 researchers) | Fielded Dec 2024; paper Sept 2026 | Pooled median **10.0%** (mean 18.2%) extinction-or-severe-disempowerment; HLMI extremely-bad median **5.0%**. | Extinction or similarly permanent and severe disempowerment | Open-ended / one variant 100 years | https://aiimpacts.org/wp-content/uploads/2026/09/ESPAI2024.pdf | Primary numeric (survey medians) | FAM-AII | **G1** with ambiguity flag. Headline scored value: pooled median **10.0%** on extinction-or-severe-disempowerment. HLMI-extremely-bad 5.0% is a companion, not a second mixture row. Not scored as G2. | H-century (100-year variant exists) | **In G1-2100** at 10.0%. | Survey (current wave) | 3.0 | First time the series median reaches 10% on the unconditional extinction/disempowerment question. Family cap: A-AII-2023 does not add weight. |

---

## Omitted under the actual-numbers-only rule

| Candidate | Why omitted from this ledger as a scored row |
|---|---|
| International AI Safety Report 2025/2026 | Loss-of-control and extinction discussed qualitatively only. No institutional numeric probability. |
| Metaculus community questions / prediction markets | **UNVERIFIED** this research pass (live pulls failed). Tier B slot exists in the methodology; not in the v1 snapshot mixture until verified at build time. |
| p(doom) wiki Hinton “50%” | Different question (good-vs-bad outcomes), not the extinction interval above. |

---

## Instrument families (weight caps)

| Family | Members | Combined-weight cap | v0.1 treatment |
|---|---|---|---|
| FAM-AII | A-AII-2023, A-AII-2024 | 3.0 | Current wave only in the mixture. |
| FAM-XPT | A-XPT-SF, A-XPT-DE | 3.0 | Both panels in; 1.5 + 1.5. |
| FAM-CARL | A-CARL-2022 | 2.0 | Single structured analysis. |
| FAM-ORD | A-ORD-2020 | 2.0 | Single structured analysis. |
| FAM-HINT | A-HINT-2024 | 1.0 | Named; near-term panel. |
| FAM-BENG | A-BENG-2023 | 1.0 | Named; no 2100 horizon. |
| FAM-HUB | A-HUB-2026 | 1.0 | Named; near-term panel. |
| FAM-LECUN | A-LECUN-2023 | 0.0 | Paraphrase exclusion. |
| FAM-YUD | A-YUD | 0.0 | UNVERIFIED exclusion. |

A source that later appears in Tier B (e.g. a Metaculus question that is itself an aggregate of some of these people) must be flagged **two-tier** and does not receive independent weight in both tiers.

---

## Conflicts of interest (source-side)

Labeled in both directions. Affiliation is a label, not a filter.

| ID | COI note |
|---|---|
| A-CARL-2022 | Written while at Open Philanthropy; x-risk grantmaker context. |
| A-AII-2023 / A-AII-2024 | Broad ML-researcher sample; includes lab-affiliated respondents. Survey, not a lab line. |
| A-XPT-SF / A-XPT-DE | Forecasting Research Institute tournament; Superforecasting vs domain-expert incentives differ by construction. |
| A-ORD-2020 | Academic / existential-risk research; not a frontier-lab employee. |
| A-HINT-2024 | Former Google Brain; public critic of unconstrained deployment. |
| A-BENG-2023 | Academic; founder of an AI-safety nonprofit (stated in later public work; the 2023 quote is the scored utterance). |
| A-LECUN-2023 | Chief AI Scientist, Meta. Incentive to downplay existential claims. |
| A-YUD | Independent; long-standing advocacy for shutdown-style policy. Incentive to upweight ruin. |
| A-HUB-2026 | Alignment researcher at a frontier lab (Anthropic). Incentive cuts both ways (lab success vs. warning). |

Site-side COI (operator thesis; product incentive to appear to move) is a separate ledger row in the methodology, not a source row.

---

## G3 note

No row in this table is a published forecast of **G3 deployment cascade** as defined in the methodology (premature deployment into critical systems producing compounding failures). G3 therefore has **n = 0** published numeric sources and is not computed as a conditioned aggregate.