Claim · 2024-10-22 · revision · Stated in source · Supported

The Claude 3 Model Family: Opus, Sonnet, Haiku states in part what the model-card question “Alignment and deception” asks.

One recorded claim. The claim panel gives its qualification, the source passages with their place in each source, and how it was checked.

What it connects

The Claude 3 Model Family: Opus, Sonnet, Haiku → states provision → Alignment and deception

Where it appears

Model cards compared →

Claim and evidence

The Claude 3 Model Family: Opus, Sonnet, Haikustates provision ↓Alignment and deception

Machine-checkedStated in sourceSupported

Supporting evidence: primary · 1 source

Provision: Model-card question · provision ID model-card-q11
Subject version: The revision with update lines of June 20 and October 22, 2024 · Publication date unknown · Effective date not established · Revised 2024-10-22
Date: 2024-10-22 · revision Records dated 2024 →

Scope and limitsThe ARA evaluations are described as indicating, among other capabilities, whether the model can make progress at deceiving humans, and the card gives one combined conclusion for them: the ASL-3 warning indicator was not crossed. It gives no conclusion on deception separately, and reports no evaluation of the model’s own propensity for deception, scheming or sabotage. The same evaluations are recorded under Q09. Task contents are not recorded (hazard rule). Search for the rest of the criterion: Sections 5.4–5.7, 6 and 7 (PDF pp. 10–31) and term search across PDF pp. 1–42. Terms searched: “decept”, “deceiv”, “scheming”, “sabotage”, “oversight”, “sandbag”, “reward hack”, “honest”. Passages read and not counted: Section 5.7 'Factual Accuracy', PDF p. 18: An honesty benchmark of factual accuracy, not an evaluation of deception or scheming. Stated, partly: the passage has what the question’s criterion asks for, in a weaker or narrower form than the criterion’s own words. What the card states, quoted from the expression read. A card is the developer’s own account: not evidence that an evaluation was run as described, that a safeguard works, or that a finding is independent.

The evidence 02

primary sourceSupports the claim

The Claude 3 Model Family: Opus, Sonnet, Haiku

Anthropic · Publication date unknown

Cite the original source

Cite the original document, not Selekoa. Check these details against it.

Source excerpt

“These tasks are selected to indicate whether the model can make meaningful progress towards accumulating resources, exploiting software security vulnerabilities, deceiving humans, and surviving autonomously in the absence of human intervention to stop the model.”

Location in source Section 6.2.1 'Autonomous Replication and Adaption (ARA) evaluations', PDF p. 24

How it was read Retrieved directly from the publisher

Machine-checked · 2026-11-04

Source details and check notes

Authors: Anthropic
Version: Retrieved for the candidate pack on 2026-11-04 (SHA-256 88ae851f53c3034d436816164ce49a876b333fa9b550fb7ef29838182b59ba7d) directly from the publisher’s host. Retrieved again on 2026-11-03 with the same SHA-256. The copy read is the revision of October 22, 2024; its PDF pages 43 to 63 hold the two Claude 3.5 addenda, which are separate cards and were not read for this row. The cross-family check of 2026-11-04 retrieved the original 42-page expression from an Internet Archive capture of 2024-03-04 at an earlier address (https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf); it is not recorded as this source's archive copy because its bytes are those of another expression.
Version date: 2024-10-22
Canonical address: Open the canonical source ↗
Accessed: 2026-11-04
Language: English (en)
Attribution: Anthropic, “The Claude 3 Model Family: Opus, Sonnet, Haiku”.
Reuse: No open redistribution licence asserted. Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution.
Check notes: Candidate pack passage, model cards compared (research/model-cards-2). The quotation (35 words) matched the card’s text retrieved for the candidate pack on 2026-11-04 (SHA-256 88ae851f53c3034d436816164ce49a876b333fa9b550fb7ef29838182b59ba7d) and again on 2026-11-04, after the adjudication, in a fresh retrieval with the same SHA-256. Matching ignored differences of whitespace, case, quotation marks, dashes, PDF line-break hyphens and Unicode compatibility forms, so the source’s typography can differ. A cross-family machine check (2026-11-04, item 4) checked this passage’s cell; its finding was applied at the adjudication of 2026-11-04. No human editorial review.

primary sourceSupports the claim

The Claude 3 Model Family: Opus, Sonnet, Haiku

Anthropic · Publication date unknown

Cite the original source

Cite the original document, not Selekoa. Check these details against it.

Source excerpt

“Our pre-specified warning indicator for ASL-3 was the model passing 50% or more of the tasks (described below) with at least a 10% pass rate (i.e., succeeding at one in ten attempts) – the model did not cross this threshold.”

Location in source Section 6.2.1 'Autonomous Replication and Adaption (ARA) evaluations', PDF p. 24

How it was read Retrieved directly from the publisher

Machine-checked · 2026-11-04

Source details and check notes

Authors: Anthropic
Version: Retrieved for the candidate pack on 2026-11-04 (SHA-256 88ae851f53c3034d436816164ce49a876b333fa9b550fb7ef29838182b59ba7d) directly from the publisher’s host. Retrieved again on 2026-11-03 with the same SHA-256. The copy read is the revision of October 22, 2024; its PDF pages 43 to 63 hold the two Claude 3.5 addenda, which are separate cards and were not read for this row. The cross-family check of 2026-11-04 retrieved the original 42-page expression from an Internet Archive capture of 2024-03-04 at an earlier address (https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf); it is not recorded as this source's archive copy because its bytes are those of another expression.
Version date: 2024-10-22
Canonical address: Open the canonical source ↗
Accessed: 2026-11-04
Language: English (en)
Attribution: Anthropic, “The Claude 3 Model Family: Opus, Sonnet, Haiku”.
Reuse: No open redistribution licence asserted. Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution.
Check notes: Candidate pack passage, model cards compared (research/model-cards-2). The quotation (40 words) matched the card’s text retrieved for the candidate pack on 2026-11-04 (SHA-256 88ae851f53c3034d436816164ce49a876b333fa9b550fb7ef29838182b59ba7d) and again on 2026-11-04, after the adjudication, in a fresh retrieval with the same SHA-256. Matching ignored differences of whitespace, case, quotation marks, dashes, PDF line-break hyphens and Unicode compatibility forms, so the source’s typography can differ. A cross-family machine check (2026-11-04, item 4) checked this passage’s cell; its finding was applied at the adjudication of 2026-11-04. No human editorial review.

Status and meaning

Editorial status. Machine-checked. Machines found this record and checked it against its sources; no person accepted it individually. Each evidence record above shows how it was checked.

Relation. An exact framework edition, or another document with an exact expression, states one provision: a capability threshold or risk level and its trigger, a required safeguard or action, a decision authority (who decides, recommends, can overrule or is informed), a stated exception, or another statement of its content, such as a licence term or a prospectus statement. The passage records what the document says, not implementation, compliance or effectiveness. Provisions with the same key align across editions of one family.

Date. Dated as the card read is dated: the date of its revision, as the source records it.

Record ID mc-cell-anthropic-claude-3-q11