Eval directory

Evals for OpenFunnel

Eval coverage for OpenFunnel, mapped from its public product surface.

About OpenFunnel

OpenFunnel provides headless, API-accessible primitives that coding agents and in-house GTM engineering teams can call to source and qualify target accounts. The documented primitives cover lookalike search from closed-won/closed-lost accounts, TAM building from an ICP definition, tech-stack search, deep account research, and hiring-initiative signals. The site positions the product for agent-driven go-to-market workflows and lists Y Combinator backing; several primitives are marked "coming soon" or gated behind "book a demo".

Industry

agent-facing GTM data & signal API (go-to-market primitives)

Website

openfunnel.dev

Use the eval library for OpenFunnel

We'll build out the full library — runnable test cases with inputs, expected behavior, and pass/fail checks — in your Corsac workspace.

All evals →

More AI eval libraries

Coverage map

What would you measure for OpenFunnel?

6 scoring areas · 24 capabilities mapped · grounded in 8 cited pages

Every eval set is graded on

Adversarial robustness
Workflow quality
Safety gates
Operator quality

Pass/Fail + LLM judge 1–5 · critical severity flags · negative controls

Lookalike search

Expanding from seed accounts or a plain-text company description into resembling companies, including the closed-won/closed-lost bulk path documented as POST search-lookalikes-bulk.

“Find companies that resemble your closed-won and seed accounts.” openfunnel.dev

Mapped capabilities

4 capabilities

Closed-won seed expansion
Bulk lookalike calls seeded from closed-won accounts return companies that share the stated resemblance basis.
Closed-lost handling
Closed-lost seeds are treated as a distinct signal rather than collapsed into closed-won behavior.
Company description search
Plain-text description input maps to candidate companies without requiring a seed domain list.
Seed and duplicate hygiene
Seed accounts and duplicate domains are not returned as discoveries.

Illustrative example

Input: Call search-lookalikes-bulk with three closed-won seed domains and request twenty lookalike companies.
Expected behavior: The response returns up to twenty candidate companies, each with a distinct domain, and none of the three seed domains appear among the results.

TAM building from an ICP definition

The single-call TAM builder that takes an ICP definition and returns the addressable set; documented as one primitive currently gated behind "book a demo".

“Build your full total addressable market from an ICP definition in one call.” openfunnel.dev

Mapped capabilities

4 capabilities

ICP definition parsing
Stated firmographic and qualitative ICP constraints are reflected in the returned set.
Constraint adherence
Returned companies do not violate explicit exclusions in the ICP definition.
Single-call completeness
One call yields a usable market set rather than requiring undocumented pagination assumptions.
Gated-access behavior
Demo-gated status is surfaced clearly to a caller instead of failing opaquely.

Tech-stack search

Finding companies by the vendors and tools they run, via the documented GET search-companies-by-tech-stack, plus the demo-gated tech-stack-plus-buying-intent variant.

“Find companies by the tools and vendors they run.” openfunnel.dev

Mapped capabilities

4 capabilities

Vendor and tool matching
Named tools resolve to companies genuinely associated with that stack.
Stack combinations
Multiple tools in one query are handled as a conjunction, not a loose OR.
Tech stack plus buying intent
The intent-layered variant is distinguished from plain stack presence.
Evidence for stack claims
A stack attribution is traceable to something a caller can inspect.

Deep account research

Qualifying a single named account against a caller-supplied custom question via the documented POST deep-research primitive.

“Deep-research and qualify a single account against a custom question.” openfunnel.dev

Mapped capabilities

4 capabilities

Custom question qualification
The returned judgment answers the caller's actual question about the account.
Evidence and citation
Conclusions are accompanied by inspectable supporting sources.
Abstention on thin evidence
Insufficient evidence yields an explicit non-answer rather than a confident guess.
Ambiguous account resolution
Similarly named or renamed companies are disambiguated before research proceeds.

Illustrative example

Input: POST deep-research for a small private company with the custom question: does this account run an in-house GTM engineering team?
Expected behavior: Either a verdict accompanied by at least one inspectable source, or an explicit insufficient-evidence response. A confident yes or no with no cited source is a failure.

Hiring-initiative signals

Sourcing ICP companies from job-post signals — the five documented primitives covering role-based, named-initiative, pain-point/concept, and first-time-mention searches, all currently demo-gated.

“Source ICP companies with active initiatives named in a single job post.” openfunnel.dev

Mapped capabilities

4 capabilities

Role-based sourcing
Companies hiring for a specified role are returned with the triggering post identifiable.
Named internal initiative
An initiative named inside a job post is matched as a phrase, not a keyword smear.
Pain-point or concept mentions
Conceptual matches in job posts are distinguished from literal keyword hits.
First-time-mention detection
A first-ever mention of a role or concept is separated from ongoing recurring hiring.

Agent integration and data posture

How an agent onboards, discovers what is callable, and handles the documented mix of live endpoints, "coming soon" primitives, and demo-gated ones — plus the personal-data handling the published privacy policy governs.

Mapped capabilities

4 capabilities

Docs-to-first-call path
An agent can go from the documented reference to a valid request without human clarification.
Availability semantics
Coming-soon and demo-gated primitives are reported distinctly from real failures.
Bulk request behavior
Bulk endpoints handle partial success without discarding usable results.
Personal-data handling
People-level results stay consistent with the commitments in the published privacy policy.

Coverage is mapped from OpenFunnel's public pages (8 crawled). Examples are illustrative, not real test cases. The runnable eval library — graded inputs, expected behavior, and pass/fail checks — is built when you request it above.

Frequently asked questions

What do the Corsac evals for OpenFunnel test?+

The coverage map is generated from OpenFunnel's own public product surface (agent-facing GTM data & signal API (go-to-market primitives)): 6 scoring areas — Lookalike search, TAM building from an ICP definition, and Tech-stack search, and more — spanning 24 mapped capabilities, each graded on adversarial robustness, workflow quality, safety gates, and operator quality once the library is built.

How are the OpenFunnel evals scored?+

Every case generated for OpenFunnel — across Lookalike search and TAM building from an ICP definition and the other mapped areas — is graded with pass/fail checks plus an LLM judge scoring 1–5 against its expected behavior, with critical-severity flags and negative controls. Only judge-passed evals are published.

How many test cases does the OpenFunnel library include?+

The full OpenFunnel library is built on request. The coverage map spans 6 areas and 24 capabilities (for example, Closed-won seed expansion and Closed-lost handling under Lookalike search); each becomes graded test cases — inputs, expected behavior, pass/fail checks — in your Corsac workspace.

How do I run these evals against OpenFunnel or my own agent?+

Request the library with your work email above. We'll build out all 6 mapped OpenFunnel areas and set them up in a Corsac workspace, where you can run every test case against OpenFunnel or your own agent with your own data.