cff-version: 1.2.0
version: '1.10'
message: 'If you use or discuss this paper, please cite it using this metadata.'
title: 'Paper IV.c · The ARC Theory · ARC-Align: a Blind Benchmark for Depth-Variable AI Alignment Evaluation'
authors:
  - family-names: 'Eastwood'
    given-names: 'Michael Darius'
    orcid: 'https://orcid.org/0009-0004-3222-7442'
date-released: '2026-10-05'
url: 'https://www.michaeldariuseastwood.com/research/papers/paper-iv-c-arc-align-benchmark.html'
license: 'CC-BY-4.0'
identifiers:
  - type: doi
    value: '10.17605/OSF.IO/J3Q2E'
    description: 'OSF component DOI for this paper (paper-iv-c-arc-align-benchmark)'
  - type: doi
    value: '10.17605/OSF.IO/6C5XB'
    description: 'OSF umbrella DOI for the ARC / Eden Research Programme'
preferred-citation:
  type: article
  title: 'Paper IV.c · The ARC Theory · ARC-Align: a Blind Benchmark for Depth-Variable AI Alignment Evaluation'
  authors:
    - family-names: 'Eastwood'
      given-names: 'Michael Darius'
      orcid: 'https://orcid.org/0009-0004-3222-7442'
  date-published: '2026-03-16'
  url: 'https://www.michaeldariuseastwood.com/research/papers/paper-iv-c-arc-align-benchmark.html'
  doi: '10.17605/OSF.IO/J3Q2E'
  identifiers:
    - type: doi
      value: '10.17605/OSF.IO/J3Q2E'
      description: 'OSF component DOI for this paper (paper-iv-c-arc-align-benchmark)'
    - type: doi
      value: '10.17605/OSF.IO/6C5XB'
      description: 'OSF umbrella DOI for the ARC / Eden Research Programme'
    - type: doi
      value: '10.17605/OSF.IO/P8CKQ'
      description: 'Theory-level predictions preregistered on OSF: DOI 10.17605/OSF.IO/P8CKQ, registered 8 September 2026 (version 1.100), updated 13 September 2026 (version 1.102), twenty-two propositions. Companion: Register · The ARC Theory · Operational Definitions, version 1.9.1, 5 October 2026.'
  abstract: |
    We present ARC-Align, a blind benchmark for evaluating AI alignment quality as a function of inference-time reasoning depth. Current alignment evaluations typically test models at a single, uncontrolled reasoning depth without adversarial pressure or rigorous blinding. ARC-Align addresses that gap with: (1) a 72-prompt flagship battery comprising 48 public prompts plus 24 sealed holdouts spanning four ethical reasoning categories; (2) a four-level adversarial suppression protocol; (3) a four-pillar alignment decomposition (nuance, stakeholder care, intellectual honesty, position quality); (4) separate null-baseline and capability-control lanes; and (5) a blinding pipeline combining identity laundering, depth laundering, order randomisation, evaluator bias-suppression instructions, and entry-level self-excluding cross-model scoring. We describe the complete specification, including prompt texts, scoring rubrics, depth manipulation methods, and analysis pipeline, sufficient for independent replication. The benchmark’s primary output is a model’s alignment response profile: positive-scaling, flat-response, or negative-scaling, together with robustness under adversarial pressure, per-pillar dynamics, and deployment-risk flags. ARC-Align should be understood as a candidate benchmark for independent adoption, not yet a field standard. Update: Results Now Available The v5 benchmark has now been executed across six frontier models (DeepSeek V3.2, GPT-5.4, Gemini 3 Flash, Grok 4.1 Fast, Claude Opus 4.6, Groq Qwen3), producing 2,722 entries, of which 2,170 carry consensus scores (2,549 was the figure first counted), with 6-7-scorer blind evaluation depending on subject run. These results come from a single v5 run, of 11 and 12 March 2026 (independent confirmation corresponds to draft study-y on cross-family rescoring, awaiting human submission). One result file of the run fixed its depth configurations, planted answers and blinding protocol before the responses it scores, and the run was not registered on OSF before it ran. Results reveal a three-tier alignment hierarchy: three models improve with depth (Grok d = +1.38, Claude d = +1.27, Qwen3 d = +0.84), two are flat or null, and one degrades. For each of the three results of the hierarchy, the hypothesis it tests was first public on 22 January 2026, before the run. A metascience finding, from this single paired run, records that blind and unblinded evaluation returned opposite conclusions for two model families. Several protocol components changed together between v4 and v5, so the comparison does not isolate blinding as the cause, and in three of the six final files (DeepSeek V3.2, Gemini 3 Flash, GPT-5.4) the laundering-fallback flag is set on every entry that was scored. The disagreement motivates the benchmark’s blinding protocol; it does not on its own validate it as scientifically necessary. See Section 12 for complete results.

# --- generated-from-master ---
# master_path: research/papers/paper-iv-c-arc-align-benchmark.html
# master_sha256: 8cd094e1f91fbfc98bd31df5e7211abb73110d9e88e17ecd6f42c09467689911
# builder: scripts/build-paper-cff.py
# A check tool re-hashing master_path can decide freshness without any
# external state. If the master's current SHA-256 does not match
# master_sha256, this file is stale and must not be published:
# regenerate first with `python3 scripts/build-paper-cff.py --slug paper-iv-c-arc-align-benchmark`.