VC open calls

Cold inbound, honestly#

Either read the call properly, or accept that it is optics.

Aide, not arbiter — by design#

Google Ventures built "The Machine" to score deals, watched it drift from a diligence aide into a de facto investment committee, and shelved it in 2022. That story is the design spec here, not the risk: independent AI reviewer roles prepare the evidence, the scores are advisory, the shortlist is built from partner decisions, and there is no mode in which the system says yes or no to a deal.

Conviction was never the part that needed automating. Reading was.

The five steps#

1 · Frame the program. Open call, scout batch, fellowship or demo day: your investment dimensions and thesis-fit criteria become the shared rubric, locked before the window opens. Your Typeform or Airtable intake stays; the batch is ingested.

2 · Every deck read in full as it lands. Every page, coverage logged, reports ready as submissions arrive. Signal engines cannot see a pre-seed founder with no web footprint; a full read can. Deck #1 and deck #400 meet the same standard.

3 · A brief per deck. Red flags, the three questions you would ask the founder, and a page-referenced quote behind every finding. Where reviewer roles disagree, the gap goes to your team as an open question rather than being averaged away. Screening calls stop being first reads.

4 · One ranked view, partners decide. The batch becomes one comparable board. Analysts review the borderline; partners confirm or override, and their calls build the shortlist. The evidence pastes straight into your own memo.

5 · Pass with feedback, under your control. Evidence-based pass feedback for every founder, reviewed and editable by your team before anything is sent, tone configured to your house style, opt-out per call.

What a rank traces to#

A worked example on Market: score 7.6, advisory, on your dimensions — finding bottom-up sizing grounded in a served niche; top-down claim unsupported, quote "…112 paying teams in vertical X, 9% m/m…" · page 8, red flag churn not disclosed anywhere in the deck, and a drafted founder question: "What's logo churn for the last two quarters?"

Quotes are verified against the deck before a finding stands. No quote, no finding.

Founder-facing feedback, with a safety catch#

AI-generated feedback going out under a fund's name is one hallucinated critique away from a thread on X. Four controls, by default:

• Reviewed before sent. Your team approves or edits every note; nothing leaves without human sign-off.
• Your tone, your template. House style configured once, signed the way you choose.
• Opt-out per call. Run a silent call when you need to — feedback is a switch, not a default you fight.
• Evidence-grounded. Every point traces to the deck, so a note cannot claim what the record does not hold.

What this does not replace#

Keep Harmonic, Specter or Affinity — different job. Signal engines discover companies through external data; a pre-seed founder who just applied often has no signal footprint at all, and nothing in that stack reads submitted decks against your criteria. Complementary, not competitive.

And this is not for the warm pipeline. It is for the programs you run for coverage, where inbound currently gets 150 seconds a deck.

Validating before it counts#

Run it in parallel on your next call: your analysts screen as usual, the panel reads the same decks, and you compare agreement and catches before committing anything.

Next steps#

• What EvalLens does not do — the boundary a skeptical GP will test.
• Read a report — the brief, layer by layer.
• Grants and prizes — the same evidence standard where an appeal is likelier.