vcscore — internal investor-diligence tool

Raising from VCs is hard, and the process is largely a black box. You fly mostly blind on who you're taking money from and how they actually behave once the wire clears, especially when things go wrong. There's plenty of public signal scattered across news, podcasts, and posts, but nobody has it organized for a founder trying to decide.

vcscore pulls that scattered public signal together for your own diligence. It's the mirror of dtscore: instead of scoring companies, it organizes what's publicly known about the people writing the checks.

The core question it answers: how founder-friendly is this investor — how do they actually treat founders, especially when things go wrong? Behavior in adversity (down rounds, recaps/washouts, founder removals, board conduct, follow-on support) is the sharpest tell, so it's weighted heaviest. Everything else is secondary to "how are they to founders."

This is an internal tool, not a published product. It organizes public information for your own decision. It asserts nothing of its own, and it is not a verdict on any firm. That scope is deliberate: it keeps you out of the reputation-publishing legal minefield. If you ever want a public version, that's a separate project with its own legal review.

How it works

  gather (Claude + web search)  ─►  verify (fetch each source)  ─►  score (deterministic)  ─►  report
  findings: claim + source +         keep only findings whose       verified findings →          scorecard
  sign + severity, per dimension     source actually supports them   0-10 per dimension          or timeline

• The LLM never sets the score. It gathers findings (each tagged founder-friendly / unfriendly / neutral, with a severity and a source URL). Deterministic code turns the verified findings into numbers, so the scoring is transparent and tunable in rubric.yaml.
• Citation verification fetches every source URL and keeps a finding only if the page actually loads and mentions the firm. This catches the failure mode that matters most with sparse data: a plausible-looking but hallucinated or non-supporting source. Findings it can't confirm aren't deleted, they're listed under "check these yourself" and don't score.
• No fabricated composites. If fewer than 4 dimensions have verified signal (common, the juicy stuff is often private), there's no score, the firm renders as a sourced incident timeline instead. It degrades to "here's the record," never to "here's a made-up number."

Usage

cd ~/vcscore

# Offline, no key needed:
./vcscore score firms/example-ventures.yaml --save-md   # score the fictional example
./vcscore rubric                                          # see/tune the dimensions + weights
./vcscore new acme-vc --firm "Acme Ventures"             # blank profile to fill by hand

# Needs ANTHROPIC_API_KEY + network:
export ANTHROPIC_API_KEY=sk-ant-...
./vcscore run "Acme Ventures"        # research + verify sources + score
./vcscore verify firms/acme-vc.yaml  # re-check the source links on a profile

score, new, rubric are fully offline. run/gather call Claude; verify fetches the source URLs.

Do this before trusting it

Whether this is worth using at all hinges on one thing: is there enough public, verifiable adversity signal per firm, or does it cluster on a few famous blowups? Before relying on it, spot-check by hand: pick a couple of firms you know something about, run vcscore run, and confirm the receipts are real and the framing is fair. If most firms come back as thin timelines, that's the tool being honest, not broken.

Files

• rubric.yaml — 7 dimensions, weights (adversity heaviest), bands. The tunable brain.
• gather.py — Claude + web search → tagged findings (model claude-opus-4-8, adaptive thinking).
• verify.py — fetch each source, confirm it supports the claim.
• scoring.py — deterministic finding → score, unknown handling, timeline fallback.
• report.py — terminal + markdown, receipts + always-on disclaimer.
• vcscore.py / vcscore — CLI + wrapper.
• tests/test_scoring.py — offline tests for scoring + verification.
• firms/example-ventures.yaml — a fictional example firm.

Roadmap

• X / Twitter ecosystem monitor (planned). Connect a bot to X to continuously watch the founder/investor ecosystem, surfacing fresh public signal about how firms treat founders and feeding it into the gather step, so a firm's picture updates over time instead of being a one-off lookup. Same rules apply: anything it surfaces still needs a verified public source before it scores, and it uses the official API, no scraping.

---

Sibling design philosophy to dtscore: LLM gathers the mess, deterministic code does the scoring.