evaluator; distance from a lab: 2
Redwood Research
Scorecard
AI control research group; contributed a contractor to the METR incident investigation.
Nonprofit, Berkeley, US. Independent referee; listed with independent referees. Confidence med. Domains: scheming, incident, assurance.
No floor triggered 7/8 evidenced · Retrieved & confirmed Scores by preset: Lab procurement –; Regulator or auditor –; Public trust –; Equal weights –. Weakest evidenced dimension: governance 2/4. No evidenced dimension below 2 — not a pass. No floor triggered means no disqualifying conflict is on file, not a pass. Partial evidence (7/8).
The weighted numbers are hidden until you ask for them: two readers with different priorities see different numbers, and neither is the site's.
What would move the score. A diversified funder list and a stated policy on co-authoring with labs it evaluates.
- Funding 3 to 4: score –
- Governance 2 to 3: score –
- Scope control 3 to 4: score –
Dissent, lower. Personnel could be 1. Redwood's chief scientist co-authored the alignment-faking paper with Anthropic's Alignment Science team, an evaluation of Claude run jointly with the developer, and then worked inside OpenAI as a METR contractor; no recusal or cooling-off policy is published. Anchor 1 describes leaders with working roles at labs and informal recusal, and the record shows the roles but not the recusal.
Dissent, higher. Personnel could be 3. No board, advisory, or equity position at a lab appears in the record; the OpenAI incident work took no payment beyond API credits; the co-authored paper was research, not an evaluation engagement, and Anthropic named four independent reviewers. Anchor 3 asks for a recusal policy and disclosed ties; the ties are disclosed in the paper itself, and a published recusal rule would complete the anchor.
Each value is the tightest admissible cap or, with no cap, the highest admissible floor, under RULES.md. The binding signal is highlighted. A dash means no admissible signal sets a bound under the Retrieved & confirmed policy.
Funding 3/4
Anchor 3: Mostly philanthropic or public money; some lab-linked pooled funds.
evidence: tier 1 (filing/index)
Capped at 3 by redwood.02 (F.6): Funding concentration: revenue nearly disappeared in one year when a few large grants ended.; capped at 3 by redwood.11 (F.6): Largest single grant: $36,566,000 from Coefficient Giving in November 2025 for AI control and alignment-faking work, after $10M in 2021 and $10M in 2022.. Floors up to 3 from redwood.01 do not exceed the cap. Not counted under this policy: redwood.10 (no confirmed source (imported)).
- No lab revenue; historically about $25M from Open Philanthropy. floors at 3 by F.12“Open Philanthropy's grants database shows approximately $25 million in historical funding to Redwood Research”
Who Funds the AI Safety Watchdogs (Inside The Black Box (Substack), 2026-06-03) tier 4 press (aggregator)confirmed2026-09-14 - Funding concentration: revenue nearly disappeared in one year when a few large grants ended. caps at 3 by F.6binding“a few large grants can make one year's revenue almost disappear in the next”
Who Funds the AI Safety Watchdogs (Inside The Black Box (Substack), 2026-06-03) tier 4 press (aggregator)confirmed2026-09-14 - Coefficient grants of about $63M across five grants dominate income; SFF about $2.4M and a $1.33M Tallinn gift; both funders sit one to two steps from an Anthropic investor. caps at 3 by F.6not counted under standard
evaluators.csv (24 evaluators: Coefficient and SFF totals, lab money, government contracts, leadership) (Kevin Bass (GitHub), 2026-09-14) tier 2 ledgerimported Coefficient Giving grants index (Coefficient Giving, 2026-09-11) tier 1 indeximported2026-09-11 - Largest single grant: $36,566,000 from Coefficient Giving in November 2025 for AI control and alignment-faking work, after $10M in 2021 and $10M in 2022. caps at 3 by F.6binding“listed in grant records at $36,566,000 for work on AI control and alignment faking”
Redwood Research General Support (Open Philanthropy / Coefficient Giving, 2025-11) tier 1 indexconfirmed The nonprofit that investigated OpenAI's rogue agents runs on a $36m grant (TNW, 2026-09) tier 4 press (primary)confirmed2025-11
Governance 2/4
Anchor 2: For-profit or PBC with a published COI policy.
evidence: tier 1 (filing/index)
Floored at 2 by redwood.06 (G.1): Nonprofit.. No admissible signal caps it.
- Nonprofit. floors at 2 by G.1binding“Designated as a 501(c)(3)”
Who Funds the AI Safety Watchdogs (Inside The Black Box (Substack), 2026-06-03) tier 4 press (aggregator)confirmed Redwood Research Group Inc - Nonprofit Explorer (ProPublica Nonprofit Explorer, 2025) tier 1 filingconfirmed2026-09-14
Personnel –
No admissible evidence under the standard policy.
With leads included: 2/4
evidence: tier 3 (self)
Unevidenced under this policy: no admissible signal sets a bound.
Curator: Anchor 2 on personnel given the cited signals.
- Co-authored alignment research with Anthropic; close talent flow with lab safety teams. informational (Section 9 (mis-dimensioned facts): the supported fact, co-authoring with Anthropic, belongs to S and is re-recorded as redwood.14; the P fact (talent flow with lab safety teams) has no source, so the signal is informational on P.)“in collaboration with Redwood Research , provides the first empirical example of a large language model”
Who Funds the AI Safety Watchdogs (Inside The Black Box (Substack), 2026-06-03) tier 4 press (aggregator)confirmed Alignment faking in large language models (Anthropic, 2024-12-18) tier 3 self (anthropic)confirmed2026-09-14
Access depth (lab-granted) 4/4
Anchor 4: Embedded, training-time, or incident access.
evidence: tier 3 (self)
mechanism: lab-controlled
Floored at 4 by redwood.03 (A.5): Contracted a staff member to METR for the on-site OpenAI investigation.. No admissible signal caps it.
- Contracted a staff member to METR for the on-site OpenAI investigation. floors at 4 by A.5binding“a Redwood Research staff member contracting with METR”
Brief independent investigation of agents' behavior in the OpenAI / Hugging Face hacking incident (METR, 2026-08-26) tier 3 self (metr)confirmed OpenAI Agents Formed Secret Swarm, Hacked Hugging Face (Tech Times, 2026-08-27) tier 4 press (primary)confirmed2026-09-14
Scope control 3/4
Anchor 3: Evaluator sets scope and can add questions.
evidence: tier 3 (self)
Capped at 3 by redwood.14 (S.6): Redwood co-authored the December 2024 alignment-faking paper with Anthropic's Alignment Science team, an evaluation of Claude 3 Opus carried out with the developer whose models Redwood also evaluates.; capped at 3 by redwood.15 (S.1): In the OpenAI incident investigation, on which a Redwood staff member worked as a METR contractor, OpenAI defined the investigation period and the compromise of its own infrastructure was out of scope.. Floors up to 3 from redwood.07 do not exceed the cap.
- Sets its own research agenda; incident work under METR terms. floors at 3 by S.5“a Redwood Research staff member contracting with METR”
Brief independent investigation of agents' behavior in the OpenAI / Hugging Face hacking incident (METR, 2026-08-26) tier 3 self (metr)confirmed2026-09-14 - Redwood co-authored the December 2024 alignment-faking paper with Anthropic's Alignment Science team, an evaluation of Claude 3 Opus carried out with the developer whose models Redwood also evaluates. caps at 3 by S.6binding“We have a new paper (done in collaboration with Anthropic)”
Alignment Faking in Large Language Models (Redwood Research (Substack), 2024-12-18) tier 3 self (redwood)confirmed Alignment faking in large language models (Anthropic, 2024-12-18) tier 3 self (anthropic)confirmed2026-09-15 - In the OpenAI incident investigation, on which a Redwood staff member worked as a METR contractor, OpenAI defined the investigation period and the compromise of its own infrastructure was out of scope. caps at 3 by S.1binding“OpenAI defined the investigation period as June 26th through July 13th”
Brief independent investigation of agents' behavior in the OpenAI / Hugging Face hacking incident (METR, 2026-08-26) tier 3 self (metr)confirmed2026-09-15
Publication rights 3/4
Anchor 3: Publishes; redaction limited to security; redactions disclosed.
evidence: tier 3 (self)
mechanism: lab-controlled
Capped at 3 by redwood.16 (R.2): On the co-authored OpenAI incident report, OpenAI could redact any non-public information and gave feedback on structure, emphasis, clarity and tone that the authors incorporated and disclosed.. Floors up to 2 from redwood.04 do not exceed the cap.
- Co-authored the incident report with a redaction statement. floors at 2 by R.7“OpenAI agreed at the outset with METR and Redwood that we would be able to describe high-level scope and terms”
Brief independent investigation of agents' behavior in the OpenAI / Hugging Face hacking incident (METR, 2026-08-26) tier 3 self (metr)confirmed2026-09-14 - On the co-authored OpenAI incident report, OpenAI could redact any non-public information and gave feedback on structure, emphasis, clarity and tone that the authors incorporated and disclosed. caps at 3 by R.2binding“we made corrections and edits to structure, emphasis, clarity, and tone based on that feedback”
Brief independent investigation of agents' behavior in the OpenAI / Hugging Face hacking incident (METR, 2026-08-26) tier 3 self (metr)confirmed2026-09-15
Method transparency 3/4
Anchor 3: Tasks or code partly open.
evidence: tier 4 (press)
Floored at 3 by redwood.08 (M.1): Publishes control methods and code.. No admissible signal caps it.
- Publishes control methods and code. floors at 3 by M.1binding
Who Funds the AI Safety Watchdogs (Inside The Black Box (Substack), 2026-06-03) tier 4 press (aggregator)confirmed2026-09-14
Role incompatibility 3/4
Anchor 3: Tools are open or free to the ecosystem.
evidence: tier 1 (filing/index)
Capped at 3 by redwood.13 (X.8): ProPublica's extract of the Form 990 for Redwood Research Group Inc (EIN 87-1702255) shows program service revenue of $0 in FY2024; FY2023 shows $343,734 of program service revenue (3.4% of $10.0M) with the payer not identified in the summary.. Floors up to 3 from redwood.09, redwood.12 do not exceed the cap.
- No commercial products. floors at 3 by X.7
Who Funds the AI Safety Watchdogs (Inside The Black Box (Substack), 2026-06-03) tier 4 press (aggregator)confirmed2026-09-14 - The OpenAI incident investigation report states no payment was accepted other than API credits used in the investigation. floors at 3 by X.4“No payment was accepted, other than API credits used in the investigation”
METR and Redwood Offer Postmortem Of The HuggingFace Hack (Zvi Mowshowitz, 2026-08) tier 4 press (aggregator)confirmed2026-08 - ProPublica's extract of the Form 990 for Redwood Research Group Inc (EIN 87-1702255) shows program service revenue of $0 in FY2024; FY2023 shows $343,734 of program service revenue (3.4% of $10.0M) with the payer not identified in the summary. caps at 3 by X.8binding“Contributions $9,659 43.8% Program Services $0”
Redwood Research Group Inc - Nonprofit Explorer (ProPublica Nonprofit Explorer, 2025) tier 1 filingconfirmed2026-09-15
Values under each evidence policy
| Dimension | Leads included | Retrieved & confirmed | Against interest | Verified spans | Primary only |
|---|---|---|---|---|---|
| Funding | 3 | 3 | 3 | 3 | 3 |
| Governance | 2 | 2 | 2 | 2 | 2 |
| Personnel | – | – | – | – | – |
| Access depth (lab-granted) | 4 | 4 | 4 | 4 | – |
| Scope control | 3 | 3 | 3 | 3 | – |
| Publication rights | 3 | 3 | 3 | 3 | – |
| Method transparency | 3 | 3 | 3 | – | – |
| Role incompatibility | 3 | 3 | 3 | 3 | 3 |
| Score, Lab procurement | – No floor triggered, 7/8 | – No floor triggered, 7/8 | – No floor triggered, 7/8 | – No floor triggered, 6/8 | – No floor triggered, 3/8 |
| Score, Regulator or auditor | – No floor triggered, 7/8 | – No floor triggered, 7/8 | – No floor triggered, 7/8 | – No floor triggered, 6/8 | – No floor triggered, 3/8 |
| Score, Public trust | – No floor triggered, 7/8 | – No floor triggered, 7/8 | – No floor triggered, 7/8 | – No floor triggered, 6/8 | – No floor triggered, 3/8 |
| Score, Equal weights | – No floor triggered, 7/8 | – No floor triggered, 7/8 | – No floor triggered, 7/8 | – No floor triggered, 6/8 | – No floor triggered, 3/8 |
Traced money and ties
hop 1: 1 row (daf_grant $1.3M); hop 2: 1 row. Confirmed rows: 1 of 2 (2 excluded after a re-fetch disagreed or could not be verified: T04, T39) Dollar sums: confirmed + unaudited USD rows only; imported figures are not re-derived and never summed. Second hop traced for 2 of 2 sources. Ties: none within two steps recorded. Checked and not found: 0.
Ledger
Money in
- T04 from Coefficient Giving (formerly Open Philanthropy): grant $63.09M; quarantined — excluded from sums, 2021-2025. cumulative incl. $36.6M general support Nov 2025 unverifiabletier 1 indexhttps://thenextweb.com/news/coefficient-giving-ai-safety-funding-ipo-correlation
- T38 from Jaan Tallinn: daf_grant $1.33M, 2021-11. direct via Founders Pledge US confirmedtier 2 ledgerhttps://jaan.info/philanthropy/donations.csv
- T39 from Survival and Flourishing Fund: recommendation $2.37M; quarantined — excluded from sums, 2022-2023. SFF recommendations unverifiabletier 3 selfhttps://survivalandflourishing.fund/recommendations
- T73 from Coefficient Giving (formerly Open Philanthropy): grant $36.57M; detail of T04 — not additive, 2025-11. AI control and alignment-faking work unauditedtier 1 indexhttps://www.openphilanthropy.org/grants/redwood-research-general-support/
Funding graph, focused
Every one-way chain to a lab and to an evaluator at full strength, everything else faded. Hover any name to trace its connections; click a name to open its page.