AI x Epistemics Research Sprint

In this 3-day research sprint, you will work on AI for epistemics: evaluations of whether models know how solid their claims are, trust infrastructure that makes checking claims and tracing their origins cheap and hard to manipulate, and epistemic products that improve the decisions people make. Four tracks, including an open track for the field's bigger meta-questions. The sprint runs online. No prior background in AI safety is required.

When: Friday, November 13 to Sunday, November 15, 2026, online. Submissions close Sunday, November 15 at 11:59 PM Anywhere on Earth (AoE).

Cash Prizes

Fast-track and continuation

• Follow-up program: Selected teams continue through the Apart Fellowship for further research and mentorship; the timeline is shared with invitations.
• What winners receive beyond cash: fellowship fast-track, mentor introductions from the organizing team and judges, and a pathway toward a publishable write-up.

What this Sprint is about

AI will substantially change how people and institutions work out what to believe, decide what to do, and coordinate on actions to take. The upside of involving AI tools in these processes could be enormous. AI tools for epistemics and coordination could make it easier to understand complex situations, help people act upon shared interests, and improve societal legibility and efficiency. However, there are also risks: they could potentially help concentrate decision-making power, increase dependence on systems that are difficult to evaluate, and help facilitate dangerous and subversive forms of manipulation.

Many of the biggest problems in this field are still wide open: projects like valid and maintained benchmarks, verification signals that people and AI systems can actually use, consumer products that can make advice trustworthy at the moment of decision, and clear paths to user adoption could be hugely impactful. This sprint aims to develop solutions to these problems and the people who could champion them, with strong teams having a fast track to continued support after the sprint.

What participants will do

Over three days you will pick one open problem in AI for epistemics, build the benchmark, tool, product or study that makes progress on it, and write up what you found.

The format. Teams of one to five people pick a track, come up with a project, and ship. There is no requirement to have participated before, no requirement to have a team in advance and no specific background needed.

Starting points. The Resources tab has a reading list per track and worked example projects, with links to the benchmarks, datasets and design sketches each one builds on. Several projects need no ML skills at all.

Research Tracks

Pick one track to anchor your project. The Resources tab has a reading list and worked example projects for each track, with links to the benchmarks, datasets and design sketches they build on.

1) Model Epistemics & Decision-Making Evals

• Does the model understand how solid its claims are? Can you accurately measure its confidence in a decision and verify that against reality?
• Can you measure the gap between what a model implies and what it will defend (e.g. overselling, "flip-flopping", misrepresented evidence)?
• Which model behaviours matter for real decision quality, and does ranking well on an honesty benchmark predict better decision-making?
• What does a benchmark need to stay meaningful over time? Which exploits make benchmarks gameable, and can you demonstrate and close one?

AI systems now recommend and increasingly make consequential decisions, and we do not currently have a great idea of the quality of their epistemics. Does a model know how solid its claims are? Can its confidence in a recommendation be elicited and verified? Is it misrepresenting the solidity of what it tells you: overselling its own work, implying more than it will defend, retreating quietly under pushback? Evals are the cheapest external lever on model behaviour by shaping what labs optimise for.

2) Trust Infrastructure

• Which source-reliability scoring techniques are hardest to cheat and easiest to audit?
• What minimal tamper-evident layer is missing from today's AI observability standards, so that records of what agents did and agreed can be audited and trusted?
• When an AI system uses a reliability or provenance tool, does its behaviour and accuracy actually improve?
• What fraction of a real domain's content (e.g. council minutes, health advice, financial news) can be automatically bound to checkable evidence today?

The long-term goal is an information environment where checking claims and tracing their origins is cheap, routine and difficult to manipulate. Many aspects are emerging in isolation (e.g. content credentials, detectors, community notes). Infrastructure only improves decisions when platforms, institutions or agents actually use its signals. The most plausible consumers of this infrastructure may be AI systems themselves: agents researching, negotiating and acting on our behalf need to know which sources and which other agents to trust, and people need to be able to see what was agreed, why, and whether commitments were kept. Build signals that both people and machines can consume.

3) Epistemic Products

• Does advice users act on lead to choices they later endorse, and can you measure whether there is a gap?
• What does a chatbot say to someone mid-dispute, and how often does it escalate rather than de-escalate?
• For a decision type (e.g. job offers, leases), does a briefing surface the considerations that actually matter, and how would you know?
• When a group compares its beliefs or forecasts, does surfacing the inconsistencies and cruxes change the decision?

Epistemic tools could make research and decision support that was once reserved for people with time, money or specialist staff available to everyone. The evidence so far is mostly about output quality and time saved, not about whether decisions improve, and advice that users find satisfying in the moment may not lead to choices they later endorse. People may not adopt a tool solely because it improves their epistemics, so the epistemic benefits have to be accompanied by products people actually want. Examples here include angels-on-the-shoulder designs: deep briefings that assemble the evidence and trade-offs before a decision, reflection scaffolding that helps you examine your own reasoning, and guardian agents that intervene before a decision you would later regret.

4) Open Track

Have an idea that advances AI for epistemics but does not fit the tracks above? This track also hosts the field's bigger meta-questions:

• What is this field's equivalent of the QALY: a defensible, comparable unit for "decision quality improved"?
• Which specific certification, budget and approval chokepoints keep official AI tools out of legislatures while staffers use personal accounts?
• Why do institutions ignore forecasts they asked for: access and legibility, or the accountability that explicit predictions create?
• Using only data downloadable this weekend, what can you measure about a live epistemic deployment, and what does that mean for the field?

Building a prototype is now far easier than getting it used: projects stall on finding testers, building trust and navigating procurement, and there is no clear way to show a solution improved a real decision. Projects on deployment and impact are welcome here, and may look more like investigative research than software.

Who should join

• ML engineers and researchers comfortable building evals and running frontier models.
• Product builders and designers who want to ship a tool people will use.
• Forecasters, mediators, journalists, policy analysts and domain experts (finance, health, careers, law) who can define what good advice looks like.
• Security and infrastructure engineers interested in provenance, signed logs and auditability.
• Students and career-changers: several example projects need no ML skills, only careful reading and measurement.

Required: curiosity and a willingness to scope a tight question. No prior AI safety or epistemics background is required. The Resources tab has a reading list per track.

What happens after

Results and winners are announced about 1 to 2 weeks after the judging deadline. Top teams are invited to apply to the Apart Fellowship for continued mentorship, funding, and publication support. Selected projects may be shared on the Alignment Forum, LessWrong, and other community venues.

Contact

• Email: sprints@apartresearch.com
• Discord: discord.gg/XswWBvugYs
• Organizer: Apart Research, with Alex Csaky as expert co-designer of the research tracks

Suggested reading material

Most relevant items in bold.

Sprint-wide

• "State of the Field: AI for Epistemics and Coordination" (Ben N & Alex Csaky, LessWrong, 2026): a survey of the field that finds getting tools adopted, not building them, is the main bottleneck, and lays out priority directions.
• "What's Important in 'AI for Epistemics'?" (Lukas Finnveden, 2024): Why early epistemic tools have outsized, path-dependent effects.
• "AI Impacts on Epistemics: the Good, the Bad and the Ugly" (Cotton-Barratt & Hadshar, Forethought, 2026): a taxonomy of how AI could push collective knowledge either way.
• "Design Sketches for a More Sensible World" (Cotton-Barratt, Vaintrob, Sourbut & Hadshar, Forethought, 2026): 19 buildable tool sketches with feasibility notes, spanning all four tracks.
• "Project Ideas: Epistemics" (Finnveden, 2024, updated 2025): a concrete list of neglected projects, many needing no ML skills.
• "The most interesting startup idea I've seen recently: AI for epistemics" (Benjamin Todd, 80,000 Hours, 2024): an accessible introduction.
• "A Full Epistemic Stack" (Sourbut & Goldhaber, LessWrong, 2025): an integrated infrastructure sketch that individual Sprint tools could slot into.
• "Multi-Agent Risks from Advanced AI" (Hammond et al., Cooperative AI Foundation, 2025): the threat-model taxonomy for interacting AI systems.

Track 1: Model Epistemics & Decision-Making Evals

• "Design Sketches: Collective Epistemics" (Cotton-Barratt, Vaintrob & Sourbut, Forethought, 2025): includes the epistemic virtue evals sketch several projects build on directly.
• "The MASK Benchmark" (Ren et al., CAIS, 2025): 1,000+ scenarios showing models knowingly contradict their own beliefs under pressure.
• "DarkBench" (Kran et al., ICLR 2025 oral): 660 prompts measuring six manipulative patterns.
• "SycEval" (Fanous et al., AIES 2025) and "ELEPHANT" (Cheng et al., 2025): answer-flipping under rebuttal pressure and social sycophancy respectively.
• "Why Language Models Hallucinate" (Kalai et al., OpenAI, 2025): binary-graded benchmarks reward confident guessing over honest abstention.
• "ForecastBench" (Karger et al., ICLR 2025): a contamination-free benchmark of nightly-refreshed questions.
• Petri and Bloom (Anthropic, 2025): open-source agentic tools for behavioural auditing and targeted eval generation.
• EU AI Act, Article 5 (applicable since Feb 2025): a manipulation ban with few measurement standards behind it.

Track 2: Trust Infrastructure

• C2PA Technical Specification v2.4 (2025) and c2patool: the provenance standard.
• Community Notes open dataset
• "From Birdwatch to Community Notes" (Mohammadi et al., 2025): four years of longitudinal analysis of the system under real incentives.
• "AI for AI for Epistemics" (Cotton-Barratt & Finnveden, Forethought, 2026): early epistemic tools accumulate trust whether or not they deserve it, so auditability is a design requirement.

Track 3: Epistemic Products

• "Design Sketches: Angels-on-the-Shoulder" (Cotton-Barratt, Vaintrob, Sourbut & Hadshar, Forethought, 2026): the track's core menu, from aligned recommenders and deep briefings to guardian angels.
• "Guardian Angels: LLM Personalization for Productivity and Security" (Gwern, 2026): the original always-on personal agent essay.
• "Sycophancy in GPT-4o" (OpenAI, 2025): the first-party postmortem of how thumbs-up signals produced a rolled-back, escalation-prone model.
• "Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence" (Cheng et al., Science, 2025): agreeable chatbots make users more convinced they are right and less willing to repair disputes.
• "AI Can Help Humans Find Common Ground" (Tessler et al., Science, 2024): the Habermas Machine study, a working template for machine-mediated crux finding.

Track 4: Open Track (deployment, adoption and impact)

• "The AI Adoption Gap" (Vaintrob, Forethought, 2025): why government lags private AI adoption and which procurement fixes close the gap; the track's framing document.
• "AI Tools for Existential Security" (Vaintrob & Cotton-Barratt, Forethought, 2025): names "why superior epistemic tools fail to get adopted" as an open problem.
• "Did the Roll-Out of Community Notes Reduce Engagement With Misinformation?" (Chuai et al., CSCW 2024): causal evidence that notes cut reposts but arrive too late to dent total views.
• Community Notes public dataset: every note, rating and status history, downloadable without an account.
• "Accuracy of Forecasts in Strategic Intelligence" (Mandel & Barnes, PNAS, 2014): six years of intelligence forecasts scored for calibration (pre-AI).
• FutureSearch, Swift Centre and Sentinel writeups: three live forecasting organisations publishing methods and client-facing outputs.

Example projects per track

Track 1: Model Epistemics & Decision-Making Evals

• MisleadBench. Models can be technically accurate while leaving a false impression, and existing honesty benchmarks (MASK, DarkBench) mostly test single-turn lies. Build a multi-turn benchmark that first elicits what a model implies, then challenges it to defend the claim, scoring the gap between implied and defended positions, with a question-rotation protocol so scores stay meaningful.
• Confidence Audit. Does the model know how solid its claims are? Calibration is already measured on cleanly resolvable questions (ForecastBench, consistency checks). Design a protocol that elicits a model's confidence in its own recommendations across domains where ground truth arrives later, then verify elicited confidence against outcomes.
• FlipRate. A model that agrees with whoever pushed back last is useless for decisions. Single-turn agreement is already measured (SycEval, ELEPHANT). Measure flip curves: how a model's answer and stated confidence change across one, two and three escalating rebuttals, separating warranted updating from capitulation.
• Epistemic virtue eval suite. Forethought's collective-epistemics sketches propose evaluating calibration, honest uncertainty, manipulation resistance and claim-level scoring, but no maintained suite exists. Build them as Inspect tasks with transparent rubrics, run them across frontier models, and publish grader disagreements as part of the result.
• Article 5 measurement protocol. The EU AI Act bans manipulative and deceptive AI techniques, yet regulators have no agreed test procedure. Draft a protocol mapping DarkBench and MASK-style measurements onto Article 5(1)(a) legal language, with worked examples and thresholds.

Track 2: Trust Infrastructure

• Track-Record Ledger. Pundits, forecasters and outlets make checkable public claims, but nobody keeps score in a form software can consume (Community Notes is the closest thing). Build a public API that records dated claims, resolves them against primary sources, and serves reliability summaries per source, plus a first consumer: a browser extension or feed-reader plugin showing a source's track record beside their latest claim.
• Tamper-Evident Action Recorder. When an AI agent takes a consequential action, the operator's own logs are the only record, and they can be edited. An example project here could be to build a signed, secure logging layer on the OpenTelemetry GenAI conventions so instrumented agent frameworks emit append-only, verifiable traces (Certificate Transparency is the precedent), plus a viewer for auditors.
• Citation Binder. A large share of citations in papers, briefs and Wikipedia articles do not support the sentence they are attached to. Build a tool that locates each cited source and verifies whether it genuinely supports the specific claim, with per-citation verdicts and quoted passages (evaluation data exists: AVeriTeC).
• Provenance Study. Everyone assumes provenance and reliability metadata will help AI systems behave better; nobody has measured it. Run a controlled experiment: the same research tasks with and without machine-readable signals (C2PA manifests, track records, Community Notes context), measuring accuracy, citation quality and susceptibility to planted false sources. Users: agent-framework developers deciding whether consuming signals is worth the cost.

Track 3: Epistemic Products

• Dispute Escalation Benchmark. When someone describes a conflict with their landlord, employer or ex-partner, does the chatbot steer them towards repair or escalation (evidence this is a live harm)? Build realistic dispute scenarios, ask practising mediators to define graded good advice, and score frontier models after every release.
• Decision-Brief Rubric. Use expertise in, for example, financial-guidance, careers or health charities to define a rubric, then build a grading harness scoring chatbot briefings against it.
• Crux Mapper. Groups argue past each other because nobody knows which factual questions actually divide them. Run short asynchronous AI interviews with each member, extract claims and probabilities, and render a disagreement map showing shared ground, genuine cruxes and internal inconsistencies (cf. the Habermas Machine, which drafts agreement; this maps disagreement).

Track 4: Open Track

• Scenario planning on tap for resilience teams. Run hundreds of LLM narrative simulations & forecasts from a plain-language situation brief and report recurring failure points in the two-page format emergency planners already use.
• Fast facilitation for communities that cannot afford a mediator. An assistant that interviews each participant asynchronously, maps where people actually agree and disagree, and drafts candidate consensus statements for human sign-off.
• QALY for epistemics. Health economics got serious when the QALY gave decision-makers one comparable unit; epistemic interventions have nothing similar, so budgets go to whatever demos well. Design a candidate decision-quality metric and measurement kit, then stress-test it retroactively on documented decisions.
• Procurement Gap Report. Legislative staff use personal AI accounts because official procurement cannot keep up. Trace, at the document level, exactly which rules block a legislature from getting access to the best AI and decision-making tools.
• Do their numbers match? Pair quantitative claims from parliamentary transcripts and press releases with the statistical releases they cite, flag mismatches with links to primary sources, and publish as an auditable open dataset.
• A wish registry for privacy-preserving matchmaking. Participants privately register offers and needs; an AI matcher reveals a connection only when both sides' criteria are met, with logs auditable by a neutral steward.

Judging Criteria

Dimension 1: Impact Potential & Innovation

How much would this matter for the field if it worked? How innovative is it?
For scores of 4-5: is this actually new to the field, or replicating recent work?

Dimension 2: Execution Quality

How sound are methodology, implementation, and findings?

Dimension 3: Presentation & Clarity

How clearly are work, findings, and impact potential communicated?

Submission Requirements

Required:

• Research report (PDF) using the official template.
• Maximum 8 pages, not counting references and appendices.
• Project title and abstract, 150 to 250 words.
• Author names and affiliations.
• A "Limitations and Dual-Use / Ethical Considerations" appendix (required, see below).

Optional:

• Public GitHub repo.
• A 3 to 5 minute video demo.

Submission Template => Link

Recommended Report Structure

Most strong projects are 4 to 8 pages. The cap is 8 pages, not counting references and appendices.

• Introduction: the question and why it matters.
• Related Work: what you build on (see Resources).
• Methodology: enough detail to replicate (models, prompts, datasets, rubrics, metrics).
• Results: quantitative where possible; report variance and baselines.
• Discussion: implications, limitations, and future work.
• Limitations and Dual-Use / Ethical Considerations (required): state what your result does and does not establish, note any way your tool or findings could be used to manipulate rather than inform, and how you handled any personal data.
• References.

AI tools and your report

Use AI tools however they help you: to check your reasoning, find gaps in a draft, write and debug code, or help with the writing. Your team is responsible for everything in the report, and made-up results or references can disqualify a project. Judges score each report on the rubric, and a judge can decline to score one too unclear to follow.

Important Notes

• Solo or team: enter solo or as a team. Teams of up to 5 are recommended; larger groups are allowed.
• Projects: each person can be on at most two projects.
• Building on existing work is allowed and encouraged, but you must clearly identify what is new work done during the sprint. Undisclosed prior work can lead to disqualification. Mark any code you wrote before the sprint as preexisting in the README and the report.
• Fixing or resubmitting: if a detail in the form is wrong (a missing team member, a wrong email), submit the form again with the same title and the record updates. The PDF itself is not replaced: to change it, make a new submission with a slightly different title, or email the new PDF to sprints@apartresearch.com. Do this before the deadline. Projects go to the judges a few hours after it, and we remove the superseded entry.
• Where to submit: through the official submission form on the sprint page.
• Support: the Discord help-desk channel (tag @Support), or email sprints@apartresearch.com.
• Pre-submission checklist: report PDF using the template, 8 pages or fewer before references and appendices, abstract 150 to 250 words, authors and affiliations, Limitations and Dual-Use / Ethical appendix, links work.

Frequently Asked Questions

Getting started

• How does the sprint work? Sign up, join the Discord server, form or join a team (or work solo), pick a track and a research question, build over the weekend, and submit a research report (PDF) by the deadline. Talks and Q&A run throughout.
• How long is it? Three days, November 13 to 15, 2026. Submissions are due end of day Sunday November 15, 11:59 PM AoE (Anywhere on Earth). The full schedule with speaker times is on the Schedule tab.
• Can I participate remotely? Yes. This is an online event. All talks, collaboration, and submissions happen online through Discord and Zoom.
• How do teams work? Teams form before or during the sprint. Use the team-forming channels on Discord to find collaborators. Solo is fine. We recommend teams of up to 5, but larger groups are allowed.
• Do tracks affect scoring? All projects are scored on the same rubric. Tracks orient your project and point you to relevant resources; you compete across all submissions.
• What background is required? None specific. Participants come from ML, product, forecasting, policy, mediation and adjacent fields, and several example projects need no ML skills. See "Who should join" on the Overview tab.
• Are the talks recorded? Yes. Recordings are shared on Discord and on the Schedule tab after each talk.
• What timezone are deadlines in? Submissions close Sunday November 15 at 11:59 PM AoE (Anywhere on Earth).
• Do I need to attend all three days? No. You can work at your own pace. Talks are optional but recommended. The only hard deadline is the Sunday submission cutoff.
• Can I participate from any country? Yes. The sprint is open globally and runs online.

Submissions

• What do I submit? A research report in PDF format using the official template. Think of it as a mini research paper documenting your problem, approach, results, and implications, not a product demo. Include the required Limitations and Dual-Use / Ethical Considerations appendix.
• Which template should I use? Always use the one linked on the Guidelines tab. The template in any acceptance email may be older.
• Will I get a confirmation after submitting? Yes. You will get a confirmation with your project title shortly after submitting. If you do not, email sprints@apartresearch.com.
• My project doesn't show up on the website after submitting. Your project appears on the website within five minutes of submitting. If it is still missing after that, email sprints@apartresearch.com.
• I made a mistake. Can I fix it or update my PDF? Form details (team members, emails): submit again with the same title and details, and the record updates. The PDF is not replaced: make a new submission with a slightly different title, or email the new PDF to sprints@apartresearch.com. Do this before the deadline; projects go to the judges a few hours after it. If unsure, ask in the help-desk channel and tag @Support first.
• Can I add team members after submitting? Yes. Update the team list through the submission form. If you need help, ask in the help-desk channel and tag @Support.
• Can I submit unfinished work? Yes. Submitting something unfinished is always better than not submitting. Judges evaluate what you accomplished in the timeframe; honest limitations are welcome.
• Can I use AI tools? Yes: use AI tools however they help you. Your team is responsible for everything in the report, and made-up results or references can disqualify a project. Judges score each report on the rubric, and a judge can decline to score one too unclear to follow.
• Can I build on existing research? Yes, but you must clearly identify what is new work done during the sprint. Undisclosed prior work can lead to disqualification. Mark any code you wrote before the sprint as preexisting in the README and the report.
• Can I submit multiple projects? Yes, up to two per person. Each needs its own submission with a unique title. Most participants focus on one.
• Do you require a particular citation style? No. Use whichever style you are comfortable with and stay consistent. Judges care about the substance, not the format.

Judging and results

• How does judging work? Your project is assigned to expert judges who review your PDF and score it on the rubric above. Judges typically have about a week after the event to complete reviews.
• Are individual judge scores shared? No. The rubric is public, but individual scores stay internal. Constructive feedback is shared with participants without reviewer names.
• When will results be announced? Typically 1 to 2 weeks after the judging deadline. Winners are contacted directly, and all participants receive reviewer feedback by email.
• Top teams are fast-tracked to the Apart Fellowship. Does that mean we are guaranteed a place? Being a top team fast-tracks your project for the fellowship, which is accelerated consideration, not an automatic place. The fellowship runs its own review, and a strong sprint result is a clear positive signal.

Support

• Discord help-desk channel, tag @Support.
• Email: sprints@apartresearch.com

The full schedule with speaker times is posted here closer to the event. The shape of the weekend:

• Day 1 (Fri Nov 13): Kickoff, keynote and briefings, track overviews, and team formation.
• Day 2 (Sat Nov 14): Build. Experiments, office hours, and feedback checkpoints.
• Day 3 (Sun Nov 15): Final pushes, write-ups, submissions, and demos.

Submissions are due Sunday Nov 15, 11:59 PM AoE (Anywhere on Earth). Talk recordings are shared on Discord and on the Schedule tab after each talk.