Why This Fits AI for Science & Safety

draft first cut, 2026-09-13.

Foresight’s “AI for Science & Safety Nodes” call doesn’t treat AI-for- science and AI-safety as separate asks — it frames them as one transition: use advancing AI capability to unlock scientific breakthroughs while keeping that transition decentralized rather than concentrating compute and power in a few hands. It names three funding layers, and breadcrumbs has a concrete answer for each.

Local compute — community-owned, not rented

co/core is a member-owned cooperative pooling compute people already own to run open models for each other, instead of renting from large clouds — today specifically Apple Silicon Macs (its confidential-computing story is built on Apple’s Secure Enclave), worth naming as the current reality rather than a ceiling: the cooperative model generalizes to other hardware as the network grows. It’s OpenAI-API-compatible, so classification and judging work in breadcrumbs’ sensemaking layer can route through it by default rather than a corporate API. Every job produces a cryptographically signed receipt — the same attestation instinct breadcrumbs already applies to claims and evidence, applied here to compute provenance instead. This is genuinely convergent, not just adjacent: co/core’s own AppView exposes receipts, jobs, and providers as signed public records over XRPC — the same data-plane shape as everything else in this stack, so a compute receipt could plausibly become part of an Evidence node’s own provenance rather than a separate system to bridge.

Decentralized alignment, by architecture

Every claim, evidence node, and attestation lives in its own author’s personal data repository — there is no central database to capture, and federation means any community can run its own trust view over the same shared graph. The anti-concentration property this RFP asks for isn’t a safety story bolted onto breadcrumbs after the fact; it’s what the data plane already is by construction.

Accountability for AI-authored contributions

The part institutional peer review has no answer for yet: as AI systems increasingly participate in research — not hypothetically, but already documented (autonomous editing agents quietly turning read-only web access into write access on public infrastructure, research agents now running multiple research-days per human day while the monitoring meant to catch problems weakens) — the attestation layer has to identify and hold accountable what made a claim, not only verify humans. Breadcrumbs’ Claim/Evidence structure is exactly the audit trail that failure mode needs: an agent’s claim is a record, individually attributable, versioned, and subject to the same tiered verification as a human’s — not exempt from it by default.

Supercollaboration, not headcount

The core diagnosis breadcrumbs is built against — first-author pipelines strangling grassroots collaboration — is the same problem this RFP names when it asks for smart allocation of compute and workflow automation instead of large teams. The build plan (one person, forking existing production infrastructure rather than starting cold) is already shaped that way, not scaled up to it after funding arrives.

AI-first science, concretely

A modular, continuously-published document model — decomposing a paper into reusable, individually-verifiable claims — is the actual mechanism for unlocking breakthroughs at the pace AI-accelerated research already moves, without losing the ability to check any one step of it.

The accountability point is the one specific to this RFP, not a generic “open science is good” pitch — and it’s grounded in something already happening, not a hypothetical risk.