A sandboxed prototype of the founder analytics layer described in the Team Interviewing Framework (the “Agent 3” idea) plus the intangibles layer and the success definition we worked through. Everything here is local, self-contained, and fully fabricated. Nothing reads from or writes to Affinity, Notion or Box at runtime, or anywhere else.
Every company (all 119, “invested” and “passed” alike), every fund label, every founder score, every intangible, every outcome, every MOIC mark, and every rater is fabricated. Nothing in this repository is derived from any real company, fund, or CRM record. No real founder or partner name appears anywhere. Raters are “Rater A” to “Rater F”.
The synthetic data is built to contain specific patterns so the dashboard has something to show. Ceiling and Talent magnetism are designed to predict graduation; Customer obsession is designed to be over-weighted in decisions but weakly predictive; the portfolio has almost no variance on Ceiling because the generator selects on it, so the “selection gap” between the full-population view and the portfolio-only view is visible by construction. When you present this, say so. The point of the demo is the shape of the analysis, not the numbers.
The generator uses seed 33. That seed was chosen from a sweep of 60 so that the designed ordering is legible at this sample size. Other seeds show the same structure with more noise, which is itself a useful demonstration of what N of 85 resolved outcomes looks like. Change SEED in data/generate_synthetic.py to see it.
A company is “graduated” (raised its next priced up-round) or “not graduated”, or it is “pending / censored” if neither has happened yet. Whether it’s still pending depends on a window that starts at entry and is keyed to stage_at_entry: Pre-Seed gets 18 months, Seed gets 24, Series A gets 30 (WINDOW_MONTHS in data/generate_synthetic.py, also in SPEC.md section 3.4). Once entry date plus that window has passed with no round either way, the company is called “not graduated”; before that, with no round yet, it’s pending.
These specific numbers (18/24/30) are an assumption, not pulled from Motive’s actual historical time between rounds, the reasoning was that earlier-stage companies have less runway and need to return to market sooner, later-stage companies have more capital and more time. If this ever runs on real data, these should be replaced with Motive’s own median time-between-rounds by stage; every graduation rate and the pending count in this dashboard depend on where this line is drawn.
Option 1, single file: open dist/founder-signal-mvp.html in any browser. It needs internet access once to fetch Chart.js from a CDN; everything else is inline. Share this one file with the team.
Option 2, source: open app/index.html in a browser (it loads ../data/data.js relative to itself).
Option 3, Cowork: the same page is saved as the artifact “founder-signal-mvp-demo” in the Cowork sidebar.
The ground truth read (no number, always open). Three plain-language cards, computed from the same data as everything below: which dimensions actually predict graduation (top 3 by lift), what the firm may be over-weighting in decisions without it predicting graduation, and the veto/spike numbers. This exists so the headline finding is never buried under process detail. Read this first; every other section is the evidence behind it.
Supporting detail (collapsed by default, click to expand): persona by vertical grid, Stage 1 to Stage 2 drift, intangibles (word frequency, theme, worry, would-back-again), the MOIC panel, rater calibration, and data completeness. These are the exploratory and QA views; nothing in them is needed to answer “what predicts success,” so they stay out of the way until you want them.
Top bar: permanent synthetic-data banner, Anonymize toggle (swaps company names for Company 01 to 119, for presenting outside the team), and filters for decision, stage at entry, vertical and cohort.
app/analytics.js computes two different lifts per dimension, and the ground-truth headline is the difference between them: dimension lift is “does a high score here predict graduation” (drives the main chart, the Dimension lift table, and the “What actually predicts graduation” card); decision lift is “does a high score here predict getting invested at all” (a new, smaller calculation, dimensionDecisionLift, used only to power the “What we may be over-weighting” card). A dimension that scores high on decision lift but low on dimension lift is one the firm leans on when deciding who to back, without it actually forecasting who succeeds, which was the original question this whole project was built to answer.
python3 data/generate_synthetic.py # writes data/data.js and data/data.json
node tests/test_analytics.js # 18 groups, 167 checks
python3 build/build_dist.py # rebuilds dist/founder-signal-mvp.html
Tunable constants are at the top of generate_synthetic.py: dimension loadings, rater biases, pass reason distribution, graduation windows, pre-money centres. SPEC.md documents every rule.
SPEC.md schema, generation rules, analytics definitions
data/portfolio_roster.json fabricated roster (names, fund, vertical, current stage only)
data/generate_synthetic.py deterministic generator
data/data.js, data.json generated data
app/analytics.js pure analytics functions (Wilson CIs, Katz lift CI, all metrics)
app/app.js rendering, filters, explorer, form
app/index.html, styles.css
tests/test_analytics.js node tests on data invariants and analytics
tests/smoke_app.js optional headless render test (needs jsdom)
build/build_dist.py single-file build
dist/founder-signal-mvp.html the shareable demo
It does not connect to live data. It scores only the CEO of each company (co-founders exist with archetype only). It does not run a predictive model; every number is descriptive with a confidence interval, which is the right posture at this N. It does not persist assessments anywhere except browser localStorage.
The path in the Notion plan still applies: create the Founder Assessments and Company Outcomes databases in Notion, backfill the outcome panel from the Box Valuation Committee memos, capture forward assessments (including passes at Chemistry or later), and point this dashboard at the real data. The analytics layer here is data-source agnostic: anything that produces the schema in SPEC.md section 2 will render.