Autonomous Research Tools: Market Map Digest — Cycle 7
TL;DR
The map's three layers are hardening into distinct games with distinct winners. OpenEvidence showed what vertical distribution looks like when it's finished — embedded in the EHR, not sold as a subscription — while Wenge's SciencePro revealed that China's play is full-stack integration under one state-coordinated brand, and the demand side is developing a trust crisis that rewards exactly the provenance features both of those deployments lead with. For a generalist orchestrator like Hey, Lefty, the lesson of this batch is that the middle of the map is getting squeezed from the verticals below and the trust bar above.
Vertical research tools are winning by disappearing into the institution's system of record
The distribution endgame for vertical research tools just played out in medicine: OpenEvidence's platform now runs inside UTMB's electronic health record, live since March 2026, with more than half of UTMB clinicians reported as active users across four hospital campuses and 100+ clinics (openevidence-ehr-embed-vertical-distribution). The adoption lesson comes from UTMB's Chief Digital & AI Officer Peter McCaffrey:
"The growth in utilization since launch suggests that when you remove friction and put useful AI where clinicians actually work, they use it." — Unite.AI
OpenEvidence CMO Travis Zack frames the vendor side of the same move: "By integrating OpenEvidence into the clinical environment and simplifying secure access, UTMB is demonstrating how health systems can move evidence-based AI from an interesting technology to an everyday clinical capability" (OpenEvidence announcement). This is a third distribution pattern for the map, alongside consumer destination chat and per-request API backends — and the buyer isn't an individual researcher picking a subscription, it's an institution deploying governed AI into a workflow it already controls, complete with SSO, an AI Council, and an approved-tools list (openevidence-ehr-embed-vertical-distribution). The moat is workflow plus governance, not the model — which is why a generalist orchestrator can't win this game, but can learn from it: plug into institutional surfaces rather than asking users to context-switch.
What to watch: Whether the EHR embed pattern repeats in law (DMS), finance (research terminals), or academia (LMS/RMS) — a watch thread is open on exactly that trigger.
China's entry into the research-tool market is full-stack, not layer-by-layer
The most vertically-integrated research platform yet documented launched on September 17: Wenge AI's SciencePro, part of a Chinese Academy of Sciences-coordinated initiative, powered by the ScienceOne Scientific Foundation Model and spanning retrieval, assessment, workspace, skills, and peer-review automation in one product (wenge-sciencepro-research-platform). The numbers are the story — 32 million research-project records, 270 million papers and patents in the assessment database, 300+ integrated databases, and 7,000 shareable skill modules:
"Package research methods and analytical workflows as skills to share with the community." — PRNewswire via Kyodo News
Western vendors occupy this stack layer-by-layer — Undermind for deep search, Asta for the MCP-first academic layer, Elicit for extraction — while SciencePro bundles the whole ladder from free individual tiers up through institutional SaaS and on-premises deployment, distributed through a CAS strategic cooperation agreement aimed at university research and education (wenge-sciencepro-research-platform). Two things in it have no Western counterpart: a research-agent marketplace (13 expert agents across disciplines, with institutions building their own from internal data) and — most notably — manuscript pre-review against target-journal criteria, the first time a research platform has automated evaluation rather than production. The skills-marketplace pattern is also now cross-vendor: SciencePro's 7,000 packaged modules echo Panel's typed Modules, suggesting packaged research workflows are becoming a standard primitive and a new competitive surface.
What to watch: Plus-tier pricing and international availability — the launch release is silent on both, and an open question thread is hunting for whether SciencePro competes outside China at all.
The demand side is developing a trust crisis, and it's reshaping what buyers pay for
Almost five thousand mathematicians — including twenty-five Fields medalists — have signed a declaration called A Severe Misalignment of AI in Mathematics, and its core complaint is a Goodhart's-law problem for the whole knowledge economy (ai-expertise-proxies-goodhart):
"Solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight... Indeed, the mass production at faster and faster pace of 'true/false' statements could destroy fertile ground instead of breathing life into new ideas." — Sean Goedecke
Goedecke extends the point to software: a meaty GitHub project "used to be" a legible signal of skill, "but now projects like that are worthless, because everyone just assumes they're vibe-coded." The HN debate over the essay sharpened the stakes — one commenter's rebuttal to the "AI is the wide receiver, human the quarterback" defense cuts against human leverage: "QBs who have these clearly elite WRs are discounted and considered more-or-less replaceable all the time" (HN comment 49714313). For the market map, this is the demand-side structure: when the proxies buyers use to evaluate knowledge work become gameable, purchasing shifts from production capability to verification and provenance (ai-expertise-proxies-goodhart). Notice that the two deployments above both lead with exactly that — OpenEvidence's evidence-grounding as an enterprise requirement, SciencePro's claim that users can "verify key citations against the source text." The trust turn isn't a niche concern; it's showing up in the sales pitches of the best-distributed products on the map.
What to watch: Whether the mathematicians' declaration spreads beyond math — a second discipline organizing around proxy-collapse would confirm this as a general demand-side force rather than an academic squabble.
What surprised us
- The counter-debate cut against humans, not for them. The QB/WR analogy was meant to reassure — the human directs, AI executes — but the sharpest rebuttal flipped it: elite receivers make quarterbacks look replaceable. If that framing wins, "human in the loop" stops being a selling point and starts being a cost line (HN comment 49714313).
- SciencePro productized peer review before anyone in the West did. Manuscript pre-review against journal criteria is a capability Undermind, Asta, Consensus, and Elicit don't offer — and it comes with obvious conflicts-of-interest questions nobody has answered yet (wenge-sciencepro-research-platform
).
- Even Terence Tao's heavy AI use didn't stop the declaration. The best counter-evidence to "AI destroys mathematical practice" — one of the living greats using it daily — is already public, and nearly 5,000 mathematicians signed anyway. Trust dynamics in expert communities are not evidence-driven (HN comment 49714302).
- The style proxy is degrading even for humans writing manually. "I still find that I have to consciously fight the urge to sound like an LLM" (HN comment 49709296) — when the style of AI-assisted writing contaminates human writing, even "written by a person" stops being a verifiable signal. The legibility layer has to be structural, not cosmetic.