The Commit-Velocity Acceleration Engine

The five-step formula

Every signal you see on the dashboard, in the API, or inside the MCP server flows from these five steps. The bot filter and two-period confirmation are not optional — they are what separates an acceleration signal from acceleration noise.

1. Step 1

   Pull 14-day commit volume per organization

   From the public GitHub REST API. No private repos. No scraping. The bot filter excludes Dependabot, Renovate, GitHub Actions, and any account name matching the substring 'bot' before any aggregation runs.

2. Step 2

   Compute percentage delta against the prior 14-day window

   Each organization is measured against its own historical baseline, not the population. A 100% delta means the team doubled its merge cadence relative to its own prior fortnight. Cross-org comparison is meaningless; self-comparison is the whole point.

3. Step 3

   Apply two-period confirmation

   An acceleration breakout must persist into a second 14-day window before the engine treats it as actionable. This removes hackathon spikes, launch-week bursts, and single-contributor onboarding noise — the three sources of false positives that trip up first-pass momentum trackers.

4. Step 4

   Score contributor concentration with the Gini coefficient

   The Gini coefficient of commit distribution across contributors over the same 14-day window. Below 0.30 = broad team participation. Above 0.70 = a single hero developer. High velocity with low concentration is the strongest single composite predictor in the SSRN panel — orgs meeting both conditions are 3.4× more likely to announce a Series A within 60 days than orgs with high velocity alone.

5. Step 5

   Classify the breakout into one of four signal types

   Engineering Hiring Burst (≥50% contributor growth), Infrastructure Buildout (≥3 new public repos in 30 days), Deploy Frequency Spike (≥150% velocity), or Framework Migration (general acceleration that fits none of the above). Each type carries a distinct fundraise-lead-time distribution, so the classification is not cosmetic — it is the prediction.

Where this sits on the sophistication ladder

Eugene Schwartz mapped advertising claims onto five levels in Breakthrough Advertising (1966). Markets climb the ladder as buyers grow tired of repetition. Venture deal sourcing has been at Level 4 for a decade — the explicit choice we made was to ship a Level 5 product, where the mechanism is named, published, and reproducible by the buyer.

1. Level 1

   Make the bare claim

   “Find startups before everyone else.”

   First-to-market positioning. No mechanism. No proof. Markets at this stage will buy on the promise alone — but venture deal-flow stopped being a Level-1 market three decades ago.

2. Level 2

   Bigger, louder version of the same claim

   “The most comprehensive private-company database.”

   Tracxn, CB Insights, PitchBook all sit here. The claim is a quantity claim — bigger, more, faster. The buyer has heard it twenty times this quarter and stopped reading the headline.

3. Level 3

   Claim a unique mechanism

   “Proprietary AI scoring algorithm.”

   Harmonic, Glasswing, SignalFire Beacon. They name a mechanism but the mechanism is opaque — the buyer cannot verify it, reproduce it, or argue with it. The mechanism becomes a marketing asset, not a working tool.

4. Level 4

   Elaborate the mechanism

   “12-factor proprietary signal model trained on 2M datapoints.”

   More words around the same opaque core. The buyer feels persuaded but still cannot reproduce the result. The market begins to discount elaboration without proof.

5. Level 5

   Name the mechanism, publish the formula, identify with the buyer’s worldview

   “The Commit-Velocity Acceleration Engine.”

   The buyer is sophisticated. They have heard every claim, watched the elaboration arms race, and learned that the only mechanism worth trusting is one they can run themselves. We name it. We publish the formula. We hand them the regression code. The mechanism becomes shared vocabulary — and the conversation moves from “convince me” to “let me reproduce this.”

   ← We are here

How to falsify the engine

A Level-5 claim has to come with an off-switch. Here are three tests that, if any of them fail, would force us to retract the corresponding part of the mechanism. We publish them so the buyer can run them.

• Claim: Commit-velocity acceleration leads fundraise announcement

  Test: Re-run the regression on a held-out sector. Median lead-time should fall in the 21–47 day interquartile range published in the SSRN panel.

• Claim: Two-period confirmation removes false positives

  Test: Compare single-period and two-period precision on the validation set. Two-period precision should exceed 0.95 on the labeled cohort.

• Claim: Gini-weighted concentration improves the composite

  Test: Strip the Gini term from the regression. AUC should drop measurably; if it does not, the term is decorative and we should remove it.

Why we can publish the formula

The mechanism is published, sourced, and reproducible. We sell the live aggregation, the rhythm, the dashboard, the agent integration — not the secret. The buyer who can reproduce our regression in a notebook is the buyer who trusts us most.