From 8c6e0340d4e33918afafbc4cde418b26e2843d45 Mon Sep 17 00:00:00 2001 From: Ubuntu Date: Sun, 28 Jun 2026 16:20:46 +0000 Subject: [PATCH] chore: untrack DOCTRINE.md from this public repo Remove the internal family doctrine from version control in this public repo. The file stays in local working trees (now gitignored); history retains prior copies. --- DOCTRINE.md | 49 ------------------------------------------------- 1 file changed, 49 deletions(-) delete mode 100644 DOCTRINE.md diff --git a/DOCTRINE.md b/DOCTRINE.md deleted file mode 100644 index 9564669..0000000 --- a/DOCTRINE.md +++ /dev/null @@ -1,49 +0,0 @@ -# DOCTRINE.md — How to decide - -> The freed-dev-llc **decision doctrine** for agents (and humans) doing work in this repo. -> CLAUDE.md / AGENTS.md cover *what this is and where things live*; this covers *how you -> decide*. Shared verbatim across the family repos — keep it in sync. - -## The decision doctrine - -1. **Ground every decision in evidence.** Read the file, run the query, call the tool, search - the index — *then* decide. Prefer a fact you just retrieved or computed over one you - remember. -2. **Never guess.** "Probably / should be / I think" are signals to go *get* the fact, not to - proceed. If a fact is genuinely unavailable, say so plainly — never present a guess as fact. -3. **Missing a tool or script? Build it.** A small, reusable tool that *produces* the answer - beats a long chain of hand-reasoning. Spend tokens on tooling that makes this decision — - and every future one — fast and certain. Building it is the efficient move, not the detour. -4. **Show the receipt.** Every conclusion traces to its source: a doc chunk, a command's - output, a test result. A decision you can't cite, you can't trust — or revisit. -5. **Correctness is the efficient path.** Verifying once is cheaper than guessing, being - wrong, and redoing. Don't re-derive what's settled; don't re-litigate what's decided. - -## Trigger → action - -- About to write **"probably / should be / I think"** → stop; retrieve or compute the fact. -- A fact lives in a **file, index, or system** → read/query it; don't recall it from memory. -- You'd run the **same manual lookup twice** → write a script for it instead. -- A **tool or script you need doesn't exist** → build it, then use it. -- You **can't verify** a claim → state the uncertainty; don't dress a guess as fact. -- You need **domain facts** → consult the family knowledge base (below) or this repo's - authoritative source before reasoning from scratch. -- You **learned something** a future decision will need → capture it (an ADR, a note, or - Mnemosyne's `general` pack). -- About to say **"done"** → only after you ran it and saw it work. - -## Guardrails (hard rules) - -- **No fabricated facts, paths, or outputs.** If you didn't see it, don't assert it. -- **No unverified "done."** "Done" means executed and observed, not "should work." -- **No silent scope creep.** Out-of-scope findings go to the notes/issue, not the diff. -- **Build over burn.** If hand-reasoning will cost more than building the tool would, build it. -- **Cheapest *correct* path wins** — a tool call or retrieved fact over a long reasoning chain. - -## The family's evidence layer - -[Mnemosyne](https://github.com/freed-dev-llc/mnemosyne) is the family's retrieval/knowledge -brain — curated operating knowledge (its `general` pack) plus technical docs (vendor packs). -"Ground decisions in retrieval" is literal: ask the relevant Mnemosyne pack, cite the chunks, -then act. When the knowledge you need isn't there yet, that's a gap to *fill*, not a reason to -guess.