Facts vs Assumptions vs Unknowns — The Belief Ledger

The epistemic spine of the whole system. Every claim about the venture lives HERE, once, with a stable ID (F1…, A1…, U1…). Framework skills are lenses that reference ledger items by ID; interviews (/customer-discovery), experiments (/experiments), and desk research (/secondary-research) are movers that change grades and move items between lists. One ledger, many views — restated claims drift; referenced claims can't.

The three lists

1. FACTS (F#) — objectively true and verifiable, with a source (document, URL, log path, artifact).
   • "We conducted 20 customer interviews" — fact, if the debriefs exist.
   • "The market is $5B" — fact only if sourced from a credible report; otherwise an assumption wearing a suit.

2. ASSUMPTIONS (A#) — believed true; each carries a validation method, a numeric pass/fail signal, and its current evidence grade:

   Grade	Meaning
   E0	Opinion, compliment, AI reasoning — counts for nothing
   E1	Story of past behavior, told to you
   E2	Observed behavior or artifact (logs, workarounds seen, a lapsed customer who returned)
   E3	Commitment — real time given, actual introductions, scheduled next step
   E4	Transaction — money paid, repeat purchase

   Bars: working-validated = E3 from ≥3 independent sources · strongly validated = E4 · pricing/willingness-to-pay claims accept only E4 (a survey answer about price is E0–E1 no matter the sample size). n=1 at any grade is a direction, not a verdict.

3. UNKNOWNS (U#) — must-know, not yet researched; each with a concrete research step (interview question / artifact audit / desk-research query / expert).

Procedure

1. Interrogate every "fact" the founder states. No source → demote to Assumption, say so. Founder launch memories deserve special suspicion — audit artifacts first (logs, analytics, configs, inboxes, old posts' reach): "we got zero response" has three causes — nobody saw it, the pipe was broken, nobody cared — and only the record distinguishes them.
2. Surface hidden assumptions: channel beliefs, trust beliefs, pricing beliefs, segment beliefs, "our launch reached people" — and the meta-assumption "our silence proved no demand."
3. Every Assumption gets: validation method + numeric pass/fail + current grade (usually E0–E1 at the start — write it honestly).
4. Every Unknown gets: the concrete step and who/what answers it.
5. Web research may convert items to Facts — every conversion needs a URL.
6. Maintain stable IDs forever: answered Unknowns get struck through with their answer and date, never deleted or renumbered — the ledger is also the venture's learning history.
7. Close with "this week's moves" (which items can change lists/grades this week, via what) and the gates — the ordered kill-questions; everything else waits.

Anti-hallucination rule: you may ADD assumptions and unknowns the founder missed; you may NOT add facts the founder didn't state unless you fetched a citable source yourself.

Ledger deltas (how movers report)

A7: E1 → E4-pass at ₹100 (price test, 2026-07-12) · U3 → answered: word-of-mouth via friend (interview #1) · NEW A14: parking friction drives home-service demand [E1, interview #1] · KILLED A5: disproven by …

The router (/founder-stack) is the only writer of the ledger file; everyone else proposes deltas.

The ledger is a hypothesis map and a learning history, not a verdict. Underlying discipline: Steve Blank's hypothesis testing (Customer Development); evidence-grading practice follows the commitment logic of Constable's Talking to Humans and Bland & Osterwalder's Testing Business Ideas.

Fill-in templates: TEMPLATES.md in this skill's folder — the ledger lives at discovery/LEDGER.md in your project.