A decision without a falsifier is a preference
Every record carries the same five parts: the context, the decision, the why, a falsifier (what would prove this wrong, and by when) and the consequences. Plus who decided, and how reversible it is.
The falsifier is the part that costs something. It is written at decision time, before the answer is known, with a number and a date attached: n ≥ 3 real projects, revisit 2026-11-18. That makes finding out you were wrong cheap and scheduled rather than expensive and accidental, and it makes the record impossible to quietly reinterpret later.
Supersede, don't edit.
A decision that turns out wrong gets a new record that supersedes it. The original stays exactly as written. A decision log you can revise is a log that agrees with you.
Eight you would actually feel
The record id is there so you can go and read the full reasoning, the consequences and the revisit date; all twenty-one are in docs/decisions/, including the internal ones that wouldn't mean much from outside.
It will never message you while you're away
No push, no phone, no "while you were away." This binds even where the notification would be genuinely useful, a red build or a failed deploy, because those are exactly the cases someone would use to reverse it. Coming back is when you get told, and never as a summons to return. Falsifier: a real founder, unprompted, asks to be told something while away, and what they want isn't already served on return.
It tells you where you are. It never scores you.
No score, letter, percentage, maturity level or traffic-light, ever, and it states what it can't see, which is usually usage, retention, cost and why you built it that way. You get two or three options with what each changes about your week, always including "nothing yet; keep building." And if you say yes, it does the migration; naming a problem and leaving you with it is a critique, not help. Falsifier: an observed founder runs it and asks for a straight verdict anyway.
Reading about a market never counts as evidence
Desk research is context and lives with the claim it informs, carrying a source and a date. The evidence ladder stays about people. No fourth rung, no "research grade": a ladder that grades two different kinds of thing grades neither. Learning three rivals charge $40 is context; watching a customer switch is evidence. Falsifier: across three real projects, research consistently has no home and gets orphaned.
It won't assume you're building a company
A company, a co-op, or a commons, whichever you're actually building, gets the same ladder. The money question stays live for all of them, and for the non-commercial ones it's harder, not softer: what keeps this alive, who else could carry it, what happens when you get bored or busy. Most open-source projects die of maintainer exhaustion, not of a missing business model.
It won't pick your licence, or open your repo for you
"Optionality by default" used to quietly mean private repo, proprietary licence. That pre-loaded answer is gone. The principle keeps its real content (decide late, on evidence) and stops deciding enclosure on your behalf. Falsifier: asked a genuinely two-sided question, founders pick proprietary at the same rate anyway, and the neutral ask changed nothing.
It grades your evidence honestly, including when you'd rather it didn't
Three rungs, stated-pain → observed-behavior → commitment, and the grade is what somebody did, not how enthusiastic they sounded. "Great idea, I'd definitely use that" lands on rung one and stays there. The only reward for real commitments is that the conscience gets quieter.
It stopped defaulting to its own favourite framework
The canvas became an answer store with interchangeable frames, and the humane frame stopped being the default; it became the floor. The framework the author built, and prefers, lost its privileged position because it hadn't earned it. Falsifier: two independent failure modes, either of which reopens it.
You can run the judgment it will apply to you
The conscience's own eval suite moved to the public tier, on the reasoning that it's tests rather than positioning. Falsifier, and it's a live risk: if by the revisit date nobody outside this machine has run, cited or corrected a case, the claim that the evals are inspectable was decoration.
What the records don't have
Outcomes. Almost none of these falsifiers has come due, because BOSS has very few users. The records show judgment under uncertainty and a pre-commitment to finding out, not a track record of having been right.
They also aren't a complete history. Twenty-one decisions were recorded; more were made. The honest reading is that the ones written down are the ones someone thought were hard to reverse at the time.
The rest of how BOSS decides what it believes: every practice is attributed, dated, and has a review trigger; the humane line and its mechanism; and the technical positions, and who each was learned from.