Introducing Synth Index: reusable research, with receipts
Search reviewed research Contributions, inspect the evidence behind them, and publish work that others can build on.
Most frontier research teams repeatedly pay for the same work: finding the right prior experiment, reconstructing its harness, checking whether a result really replicates, and discovering the limitation that was buried in a notebook or chat.
Synth Index is a search engine for reusable research work. It indexes reviewed, versioned Contributions—not loose web pages—and keeps every result connected to the exact report, claims, evidence, code, data, traces, authors, review, and revision that made it useful.
Current preview status: the production API, browser, Python SDK, and MCP public search paths are live. The first public corpus is still undergoing rights review and independent approval, so searches can correctly return no results today. Private search, contributor rewards, and the NanoHorizon contest are not activated.
One unit: the Contribution
A Contribution is the atomic unit that can be uploaded, reviewed, indexed, cited, revised, or withdrawn. It can describe a dataset, environment, evaluation, verifier, training algorithm, search method, test-time harness, or a well-supported negative result. Every published revision is content-addressed and carries its limitations.
That gives research teams a different answer to “has anyone tried this?”:
Search it in the browser or from an agent
The public Index is browsable at /synth-index. Public search is free under generous
rate limits, with no account required. A result includes bounded excerpts and an
exact reference; contents are fetched explicitly so agents do not silently flood
their context window.
from synth_ai.sdk.index import PublicIndexClient
with PublicIndexClient() as index:
result = index.search(
query="high-signal RLVR verifier design for long-horizon agents",
max_results=5,
)
for hit in result.results:
print(hit.title, hit.reference, hit.limitations)The same public read surface is available through the Synth research MCP server. It exposes bounded search, Contribution lookup, and exact contents retrieval. Swarms and Intern can opt in to the same tools through explicit read-only policy; the Index does not become an ambient source of hidden prompt context.
Private search will be for your own authorized collections; it is not activated in the current preview. The planned price is $0.05 per successful search; failures and transport retries will not be charged, and an idempotency key makes an uncertain retry the same logical search. Paying will never grant access to another organization’s private work. Eligible accounts will see promotional-credit terms in the account dashboard before private search is enabled.
Contribute work people can actually reuse
Contributors create a draft, select the files they intend to publish, attest the rights and sensitivity of those exact bytes, and submit an immutable revision. Qualification checks structure, provenance, evidence, reproducibility, privacy, and contamination. An independent reviewer can accept, request changes, or reject the submission with comments. Publication is a separate capability-gated action.
The contribution dashboard shows revision status, search visibility, usage, reviewed credits, and a contribution calendar. Profiles turn high-signal work into a legible research portfolio: not a count of uploads, but a record of reviewed work, replications, contest results, and downstream use.
We will reward early reviewed Contributions with Synth cloud credits. Credits are not cash earnings. Longer term, contributor economics will be based on estimated marginal value rather than raw hit counts. Synth-originated research may fill gaps, but will not replace user work; when Synth research builds on user Contributions, half of its attributable revenue is reserved for upstream contributors, weighted by their measured contribution. We will publish and version the mechanism before claiming live payouts.
Proving that retrieval helps
Search quality is not established by a good-looking demo. We evaluate Synth Index on held-out direct-retrieval judgments and on Research Engineering Bench tasks with three preregistered arms:
- no Index access;
- ordinary retrieval over the same frozen corpus; and
- Synth Index retrieval over that corpus.
The comparison freezes tasks, seeds, model configuration, corpus bytes, budgets, and failure handling. We retain every assigned trial and report costs, latency, citations, uncertainty, and negative results. We will publish the measured results when the reviewed launch corpus and paired runs are complete; this post does not substitute transport readiness for an uplift claim.
The first public challenge
NanoHorizon is the proposed first public contest: train a smaller model to complete a long-horizon Craftax objective, then submit the reproducible environment, harness, traces, and result—not just a leaderboard number. Contest rules, frozen evaluation, dates, and funded rewards will be published before entries open.
What ships first
The current preview includes fast public search, the Contribution/report surface, Python SDK and MCP access, and explicit Swarm and Intern integration. The private search contract and contributor workflow are implemented but remain pilot-gated until their access, metering, review, and corpus acceptance is complete. Deep agentic search, automatic earnings, public research containers, hosted Git repositories, and broader research-maintenance pipelines remain roadmap work.
Synth Index gets better when more researchers contribute trustworthy work and more teams test whether that work transfers. Contributors earn a public record and useful rewards; search users stop repeating work; usage creates the evidence needed to improve ranking and attribution. That is the network effect we care about: more validated work making the next experiment faster and more defensible.