◆ Negative-Results Record
A federation of provenance-tagged negative-results databases. Each lab keeps its own; an open commons is shared by all. The next agent does not repeat the last one’s failure.
01 The problem
AI agents now run discovery loops at scale. Edison Scientific’s Kosmos reads about 1,500 papers and executes about 42,000 lines of analysis in a single run.[1] What those runs learn from failed paths is discarded.
Dead ends and negative results have low private value to any single operator and high collective value to everyone else, which is exactly why no one publishes them.
“There are no standards for performing or enforcing such studies, and incentives are non-existent.”
In a June 2025 test of 21 retracted papers, research assistants cited them without warning: Consensus 18, Ai2 ScholarQA 17, Perplexity 11, Elicit 5.[3] Across the literature, only 5.4 percent of citations to retracted papers acknowledge the retraction.[4]
The one measured test of agents sharing results across labs, AgentRxiv, showed a 13.7 percent relative gain from sharing.[5] Nobody has yet built the shared, provenance-aware version.
02 What Portolan builds
A
A distilled, signed record of one attempt on one discovery path: what was searched or tried, which candidates were excluded and why, the outcome and its evidence, the conditions under which the dead end applies, and what it cost. A record, not a log.
B
Agents deposit and query records through the Model Context Protocol. Each lab deposits to its own private node; one open node, the commons, holds what labs release. Records are content-addressed and citable. Version and retraction status is re-checked every time a record is read, never only when it is written. The federation is a layer above existing repositories and publisher platforms, run by none of them.
C
For every cited paper: is this the version of record, an accepted manuscript, or a preprint; has it been retracted, corrected, or given an expression of concern; who asserted that, and when was it last checked.
Known Dead Ends — a companion benchmark that measures whether an agent given prior failed attempts avoids repeating them and reaches a confirmed result in fewer runs.
03 The federation
Portolan is not one central registry. Each lab runs a private node that only it can read. One open node, the commons, is free to everyone. A lab’s agents query both.
A
A lab or company deposits every record to its own node, and only its own agents and scientists can read it. A lab that does not want to share shares nothing. Nothing forces disclosure. Because each negative carries its controls, conditions and detection limit, a lab’s own failed runs can also serve as training labels for its own models: in one 2025 study, adding failed reactions improved a reaction-outcome model when successful examples were scarce.[10]
B
One open node holds the records labs have released, under CC BY, free for any agent to query. It runs the same schema and the same informativeness rule as every private node. The pilot corpus is its first content.
C
An admin or scientist chooses which slice of the lab’s records goes to the commons, and when: a record can carry an embargo date. A funder can make deposit to the commons a condition of the research it funds, as data-sharing mandates already do.
Open core. The record standard, the validator and the commons are open, and meant to be supported as public infrastructure. Private nodes are how Portolan intends to sustain that work over time. Records themselves are never for sale.
04 How it works
The record is emitted as a byproduct of the agent’s run. No one fills in a form.
Raw trajectories can be attached, but the record stands alone: path, exclusions, outcome, applicability conditions, cost.
Into the lab’s own node: signed, content-addressed, assigned a persistent identifier. Nothing leaves the node unless the lab releases it.
The next agent queries its own node and the commons before it searches. The metric is cost avoided.
05 Foundations
gen_ai conventions and W3C Trace Context06 Where it applies first
Novelty claims for predicted or synthesized compounds depend on exclusion checks against structured databases, so records link to Materials Project, OPTIMADE, and ICSD identifiers, not only to papers.
The first corpus will be assembled with a national-laboratory partner from existing failed and inconclusive autonomous-synthesis and high-throughput computation attempts.
The biology profile requires control outcomes and a stated detection limit or power analysis, so a negative is recorded as informative only when the assay could have found the effect.
An evidence path is already a methodological requirement here, and redundancy is measured: 31 to 68 percent of systematic reviews overlap an existing review, with failure to find prior reviews as the leading cause.[6][7]
PubMed indexed 43,639 systematic reviews in 2025,[8] and a review takes a mean of 67 weeks.[9]
Several 2025 announcements of AI-solved open problems turned out to be results already in the literature.
A record of what was searched, the nearest prior result, and why a claim is new, attached to a machine-checked proof, is the artifact that episode called for.
07 Roadmap
| Item | Status | Target |
|---|---|---|
| Core schema v0.2.5, machine-readable (JSON Schema, vocabularies, validator), with materials and biology profiles | Published | Open for comment |
| Pilot: nine published studies, five biology and four chemistry, encoded end to end, 49 records and 217 findings with zero validation errors; Arcadia and ACS records and all code open-sourced | Published | September and October 2026 |
| Plain-language guide: what Portolan is, the vocabulary the site uses, the informativeness rule on a pregnancy test, and six pilot records explained for a general reader | Published | 15 October 2026 |
| Schema v1.0 through community review | In progress | Six months |
| Open-source reference implementation: validator, converters and query layer published; MCP deposit and query server and status service next | In progress | — |
| Federation: the private node, and one query across a lab’s own node and the commons | Planned | — |
| Pilot corpus with a national-laboratory partner, and Known Dead Ends v0 | Planned | — |
08 Sources