From 30a6a910f614415dde6ad226ff8e5973805ef5c8 Mon Sep 17 00:00:00 2001 From: "John M. P. Knox" Date: Sun, 9 Aug 2026 12:25:48 -0500 Subject: [PATCH] docs(README): first-run expectations + trim a repeat MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Applies the low-risk fixes from a cross-provider Impasse review of the README: - Add a "First run — what to expect" note under Requirements: you need an account with the reviewer's provider (separate from the host), it spends real tokens, the first send blocks for consent, and what a review does to your files (reviewer read-only; the host applies verified fixes to your working copy). Closes the onboarding/edit-authorization gap the review flagged. - Drop a near-verbatim repeat of "the reviewer never edits — critic and editor stay separate" from the flowchart caption (already stated in the opening). Deliberately NOT changed: the benefit-first hero, "decorrelation beats intelligence", and the CLAR priority claim (the term was introduced here — that is stated plainly, not hedged). The opening's confident framing is intentional; the "host-directed, not enforced" honesty sits prominently right below it. The new three-paper section (Xiang et al., Kumar et al.) was itself run through Impasse — it handles the inconvenient asymmetry result candidly — and both new arXiv citations were verified against the source abstracts (titles, authors, and headline figures match). Co-Authored-By: Claude Opus 4.8 (1M context) Claude-Session: https://claude.ai/code/session_01X8XaW2bqgYnp7oRVgfHXid --- README.md | 23 +++++++++++++++++++---- 1 file changed, 19 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index dae7cc6..49347c3 100644 --- a/README.md +++ b/README.md @@ -38,10 +38,9 @@ flowchart TB ``` The reviewer (indigo) proposes; the host (blue) verifies and applies the fixes they agree on; the -judgment calls come to you. The independent reviewer never edits — the critic and the editor stay -separate. A refutation only *drops* a finding when the host has contradicting evidence — a host -disagreement with no evidence isn't a refutation, so it goes back to the reviewer for one round, -then to you if neither side can win. +judgment calls come to you. A refutation only *drops* a finding when the host has contradicting +evidence — a host disagreement with no evidence isn't a refutation, so it goes back to the reviewer +for one round, then to you if neither side can win. **Status: pre-release.** The open implementation of the pattern — named in [the CLAR essay](https://www.movingavg.com/essays/cross-lab-adversarial-review.html) and told as a @@ -289,6 +288,22 @@ Impasse performs on *your* work. - Python 3 (standard library only — the shipped helpers have no pip dependencies). - macOS or Linux. Windows via WSL; native Windows is on the [roadmap](docs/windows.md). +### First run — what to expect + +- **You need an account with the reviewer's provider.** The cross-provider backend logs in with its + own credentials — a ChatGPT account or an OpenAI API key for Codex, a Claude account or an Anthropic + key for the Claude CLI (exact plan/entitlement is the provider's to set). That is separate from your + host, and reviews spend that provider's tokens. +- **It costs real tokens.** A small artifact at default effort is typically cents; large inputs or + high `--effort` cost more and take longer (a review can run from well under a minute to several, + depending on effort and size). It is not free to run. +- **The first send pauses for consent.** Reviewing sends your artifact to a third-party provider, so + the run **blocks the first time** until you approve the destination (see Data boundary & consent). + Nothing leaves your machine before you approve it. +- **What it does to your files.** The reviewer is **read-only** — it never touches your artifact. The + *host* applies the fixes it verifies to your working copy for you to review (like any edit your + agent makes); the review step itself changes nothing on its own. + ### How independent is it? Independence is a ladder, not a switch — and Impasse always tells you which rung you're on. A