# Extraction ledger — what's evidence and what's guess

The kit's acceptance test (the falsifier, from the 2026-07-15 panel): **one person outside the origin community, with only their own subscription and this folder, founds a community and ships a first artifact remotely — no SOMA member present.** Until that has happened at least once, this kit is a hypothesis.

| # | Claim in the kit | Status | Evidence |
|---|---|---|---|
| 1 | A subscription's CLI + a folder of canon is enough substrate to run a working human+AI org | **evidence-backed (n=1, origin only)** | SOMA itself; 2026-07-14: 28,327 LOC/11.6h day. Confounded by founder skill — see collegiality experiment |
| 2 | The 90-minute founding path produces a durable community for a stranger | **guess** | never run outside origin |
| 3 | Naming + interviewing an AI colleague changes work outcomes vs. tool-use | **guess (under test)** | `SOMA/research/collegiality-experiment-v0.md` — pre-registerable protocol exists; not yet run |
| 4 | Fill-in-the-blank standards beat copy-our-standards | **guess** | inference from "executable gates over prose" (which is origin-evidence-backed) |
| 5 | A first artifact in the first session is the retention hinge | **guess** | panel consensus (3 model families), zero external data |
| 6 | Communities need explicit floors or federation launders bad practice | **judgment** | not empirically testable in advance; adopted as policy |
| 7 | Email is a sufficient inter-community transport at seed stage | **evidence-backed (origin)** | SOMA dispatch/board-by-email runs in production |
| 8 | A web-chat AI can sherpa a novice through CLI install via paste-back gates alone | **guess** | 2026-07-24 addition (SHERPA.md); simulated-novice test pending, no external run |
| 9 | A pre-install sentiment interview improves founding (better standards, retention) | **guess** | 2026-07-24 addition (Mike's directive); interview→constitution wiring is inference from Greta pattern (origin) |
| 10 | The founding packet survives the chat→CLI handoff (human copies it intact) | **guess** | weakest link in the two-stage path; watch it in every probe |
| 11 | Judging members on acts rather than character is sustainable under a real conflict | **guess** | 2026-07-25 addition (`seed/IF-IT-GOES-WRONG.md` §1); procedure written, never once applied |
| 12 | A negative claim paired with the subject's response does less damage than one delivered alone | **judgment** | argument, not evidence: Mike 2016-06-17 "to reserve judgment is to believe" — there is no neutral bucket. Adopted as policy |
| 13 | Bounding removal to the individual (voucher loses vouching scope only) avoids the collective-punishment trap | **guess** | 2026-07-25 panel finding (Skip): subtree de-admission is a sanction nobody will ever execute, so it is not a cost |
| 14 | Exit-with-everything is architectural, not policy — nobody can withhold what was never centrally held | **evidence-backed (origin)** | the folder *is* the substrate (`seed/START-HERE.md` §2); no server exists to revoke from. Untested against a hostile founder |

**Rule for editing this kit:** a claim moves to evidence-backed only with a named external community and a dated observation. Additions from origin experience enter as *guess* by default. Kill what fails; record the kill here.

| Date | Probe | Person/community | Result | Kit change |
|---|---|---|---|---|
| 2026-07-24 | Simulated novice run #1 ("Pat": ChatGPT-only, no terminal) — **SIMULATED, moves no claim** | internal (Sonnet two-voice) | 10 protocol gaps; confidence lowered on claims 8–10 | Same-day hardening: packet save-gate, handoff closure gate (colleague's name), lost-packet fork in FOUNDING, Node/EACCES walk, truncated-paste rule, cost answer, multi-sub tie-break. Details: `probes/2026-07-24-simulated-novice.md` |
| — | (first external probe not yet run) | — | — | — |