Assay

Provenance + scientific-claim diligence for deeptech founders and companies.

Generic startup-intelligence APIs (Crunchbase, Harmonic, PitchBook) answer who raised and who invested. They don't answer the question that actually decides a deeptech deal: is the science real, does the founder own the IP, and has anyone non-dilutive already underwritten the work?

Assay has two modes, both built on free public APIs:

• Provenance — is the founder a real researcher, do they own the IP (or does their old lab?), and has anyone non-dilutive funded the work?
• Science (below) — take a technical claim and grade it against the primary literature, with independent replication as the headline signal.

Provenance triangulates three public data layers nobody bothers to wire together:

Layer	Source	Question it answers
Research lineage	OpenAlex	Is the founder a real, recent, cited researcher? Where?
Patent ownership	USPTO / PatentsView	Are they the inventor — and is the company the assignee, or their old lab?
Non-dilutive funding	SBIR.gov + NIH RePORTER	Has the technical claim been independently funded?

It returns a structured report with severity-ranked flags. The highest-value check is ip_owned_by_institution: the founder is named as inventor, but the patent is assigned to a university or national lab — meaning the startup needs an explicit field-of-use license before the IP is investable. That single flag has killed real deals.

All three layers are free, keyless APIs, so the engine runs with zero paid data subscriptions. The only optional key is Anthropic, for an LLM verdict on top of the deterministic flags.

Install

git clone https://github.com/ahammadshibil/assay.git
cd assay
pip install -e ".[llm]"      # drop [llm] if you don't want the synthesis layer

Use

# CLI
assay --founder "Jane Researcher" --company "Acme Bio" \
      --institution "CSIR-IGIB" --email you@fund.com

# add an LLM verdict (needs ANTHROPIC_API_KEY)
assay --founder "Jane Researcher" --company "Acme Bio" --synthesize

# attach a vertical for extra sources (biotech = ClinicalTrials.gov + openFDA)
assay --founder "Jane Researcher" --company "Acme Bio" --vertical biotech

# machine-readable, e.g. to pipe into Airtable
assay --founder "Jane Researcher" --company "Acme Bio" --json

# Library
import asyncio
from assay import run_check, synthesize

report = asyncio.run(run_check(
    founder="Jane Researcher", company="Acme Bio",
    institution="CSIR-IGIB", mailto="you@fund.com",
))
for f in report.flags:
    print(f.severity.value, f.message)
report.verdict = synthesize(report)   # optional

Science mode — confirm the claim, not the founder

The founder being real is table stakes. The question that decides a deeptech deal is whether the underlying science is real, published, and corroborated by someone other than the founders. Science mode takes a technical claim and grades it against the primary literature (OpenAlex), with independent replication as the headline signal.

# CLI
assay-science \
  --claim "CSP-TTK21 activates p300/CBP acetyltransferase and promotes axon regeneration" \
  --founders "Tapas Kumar Kundu, Snehajyoti Chatterjee"   # enables the independence check

assay-science --claim "..." --founders "..." --synthesize   # LLM reads actual support
assay-science --claim "..." --json

import asyncio
from assay import verify_claim
r = asyncio.run(verify_claim(
    "enFnCas9 high-fidelity CRISPR variant with improved specificity",
    founders=["Sundaram Acharya", "Debojyoti Chakraborty"],
))
print(r.grade.value)          # REPLICATED / SINGLE_GROUP / PREPRINT_ONLY / ...

Grades (the unit is the claim, the evidence is the literature):

Grade	Meaning
REPLICATED	peer-reviewed primary article from a group other than the founders
SINGLE_GROUP	peer-reviewed, but only the founders' own lab — no independent replication
PEER_REVIEWED	peer-reviewed primary exists; independence not assessed (pass --founders)
PREPRINT_ONLY	only preprints — not yet peer-reviewed
UNSUPPORTED	no literature found for the claim's terms
CONTRADICTED	a supporting work is retracted
TOO_GENERIC	the claim is all buzzwords ("an AI platform designs molecules") — it matches the field, not the company

Generic-claim guard. A claim built only from buzzwords has nothing specific to verify and otherwise greenwashes to REPLICATED off unrelated papers — so a claim carrying no distinctive term (a named compound / target / mechanism) grades TOO_GENERIC instead. Caught a real portfolio case where "a generative-AI platform designs optimized molecules" was matching the entire AI-drug-discovery literature. Subtler topical mismatches (a real term, but off-topic papers) need --synthesize to catch — the deterministic grade can't read.

The independent_replication flag is graded conservatively: reviews mistyped as articles are excluded, a single lightly-cited corroboration is downgraded to replication_thin, and — per the same integrity rule as provenance mode — a literature search that errored is flagged literature_source_unavailable, never silently graded UNSUPPORTED.

--synthesize: adversarial adjudication (structural → verified)

The deterministic grade answers "does independent peer-reviewed literature exist?" — but it can't read, so a review or a keyword-only match can inflate it. --synthesize closes that gap with a loop adapted from DeepMind's Co-Scientist (Nature, 2026): for each surfaced paper it reconstructs the abstract and runs a self-adversarial pass — steelman support, steelman refutation, then rule — yielding SUPPORTS / PARTIAL / IRRELEVANT / CONTRADICTS per paper. A meta-review aggregates these into a verified stance:

Verified stance	Meaning
INDEPENDENTLY SUPPORTED	a non-founder paper actually supports the claim on reading
SUPPORTED (founders' group only)	only the founders' papers genuinely support it
NOT SUPPORTED	literature matched keywords but none establishes the claim
REFUTED	a paper contradicts it

So REPLICATED (structural) + NOT SUPPORTED (verified) is a real and common outcome — keyword hits that evaporate on reading. Needs ANTHROPIC_API_KEY; the aggregation logic is pure and offline-tested. A title/abstract read is still not a substitute for reading the papers — it's a sharper triage.

Emit into a diligence memo (--format l3)

Both modes can emit their findings as ready-to-paste memo sections instead of a console report — so a run drafts the deal memo rather than feeding a retype:

# Science → "§10 Claim Verification" (✅ checks out / ⚠️ inflated / ❌ couldn't verify)
#           + "Appendix A — Key Publications"
assay-science --claims-file deck_claims.txt --founders "..." --format l3

# Provenance → "§3 Founder — Research Lineage" + "§8 IP Portfolio"
assay --founder "..." --company "..." --institution "..." --format l3

--claims-file takes one claim per line and builds the full three-bucket ledger. Each claim's bucket comes from its evidence grade — or, if --synthesize ran, from the adversarial verified stance (so a keyword-only REPLICATED lands in ❌ "couldn't verify", not ✅). The integrity rule holds in the output too: an unreachable patent or literature source renders as UNVERIFIED, not clean. Markdown only — Assay prints the sections; it never writes into your vault.

Tests

pip install -e ".[dev]"
pytest -q          # 65 offline tests: provenance, integrity, parsers, science grader, adjudication, renderers

The suite is network-free — it pins the parser field-paths against fixtures and asserts the integrity property (failed source ⇒ *_source_unavailable, never "none found"), so API drift fails loudly in CI.

Evaluation

Science mode is benchmarked against SciFact, the standard scientific-claim-verification dataset — see eval/. The eval loop drove two improvements and then told us when to stop: a single full-sentence query scored 44% recall → adding a content-term query (_content_query) lifted it to 78% → adding PubMed reached 84%, at which point the NEI rate caught up and the eval showed retrieval is no longer the bottleneck — judgment is (the --synthesize layer). Every change has a number attached, including the decision to quit optimizing retrieval.

Verticals

The core engine is one domain-agnostic agent. Verticals are plug-in modules that attach extra sources behind it and contribute additional flags — they never replace the core. This keeps the surface area small: you build the provenance triangle once, and a vertical adds only what's genuinely sector-specific.

Shipped:

• biotech — ClinicalTrials.gov v2 (does the company sponsor registered trials, at what phase/status?) and openFDA drugsfda (any approved products?). The no_registered_trials WATCH is the useful one: a company positioning as clinical-stage with nothing registered is a discrepancy to reconcile.

To add one (e.g. space, semis): subclass VerticalModule in assay/verticals/, implement evaluate(ctx) -> VerticalResult, decorate with @register, and import it in assay/verticals/__init__.py. It then appears as a --vertical choice automatically. ~40 lines per vertical.

How it scores

Flags carry one of three severities:

• RISK — a hard gate (e.g. IP assigned to an institution, not the company).
• WATCH — verify manually (stale research, weak name match, grants to the wrong entity).
• INFO — recorded signal (h-index, patent count, non-dilutive validation).

Rules live in assay/engine.py and are deliberately readable — the judgment about what a flag means is the moat, not the data access, so the rules are meant to be edited per your thesis.

Honest caveats

• Live-test status (2026-06-11). OpenAlex and NIH RePORTER verified working end-to-end. ClinicalTrials.gov works but WAF-blocks some datacenter IPs (403) — a network condition, not a bug. SBIR.gov was returning 429 (service-side maintenance). The PatentsView PatentSearch host (search.patentsview.org) no longer resolves — see below.
• Unverified ≠ clean. When a source can't be reached, the engine emits a *_source_unavailable WATCH ("IP ownership is UNVERIFIED"), never the substantive "none found" flag. A failed lookup must never read as a clean bill of health — that distinction is enforced and tested (tests/test_engine.py).
• USPTO is mid-migration. PatentSearch is moving to the Open Data Portal (data.uspto.gov) with no public ETA, and the old keyless host is down. Point the client at the ODP endpoint once available via export ASSAY_PATENT_BASE=... and pass --patents-api-key. Until then the patent layer flags as unverified.
• US-centric grants. SBIR and NIH are US-only. Indian (e.g. BIRAC, DST) and EU (CORDIS) non-dilutive sources are not wired in yet — see roadmap.
• Name disambiguation is shallow. OpenAlex matching uses institution as a tiebreak; common names need a manual check (surfaced as a WATCH flag).
• This is a sourcing/triage aid, not legal or investment advice.

Roadmap

• More verticals behind the plug-in interface: space (Space-Track orbital data, FCC/ITU spectrum filings), semis (deeper patent/process analysis), robotics (citation-graph depth). Biotech is the reference implementation.
• Lens.org patent↔paper joins (links a patent to the scholarly work behind it).
• EPO OPS + Google Patents (BigQuery) for non-US IP.
• Indian grant registries (BIRAC, DST/SERB) + CORDIS (EU) in grants.
• Crossref / Semantic Scholar for citation-graph depth and influence weighting.
• Caching layer + proper author-ID resolution to cut repeat API calls.
• Confidence scoring on the disambiguation step.
• A thin web UI / Airtable sync for non-CLI users.

License

MIT.