[kishuxz]

Prose condensed only. No section removed, no figure removed, nothing collapsed behind a disclosure.

Where the lines went

Section	Before	After	What changed
Why supersession	21	13	Two paragraphs to one. Dropped the sup-01 worked example and the happy-endings close. Kept 60 of 273 and obligation-not-outcome.
Corpus separability	81	59	All four parts kept. Method to three sentences, the failure to one paragraph, the no-gate argument to two. Output block verbatim.
Grounding	27	21	Opening to two sentences, deadline-drift paragraph dropped. Span table and parse-time invariant kept.
Cost-weighted error	21	20	Intro condensed. Full weight table untouched.
Limitations	68	61	All 11 entries kept, each at two sentences or fewer, every figure and reason intact.
Roadmap + Kill criteria	31	15	Prose to bullets.
Total	339	279

Untouched as instructed: Findings (31), Status and scope (13), The corpus (38), License (3).

Where I stopped, and why

279, not the ~200 asked for. The brief said to keep the longer version where a cut would lose a specific number or a stated reason, and report where I stopped. That point is here:

• Limitations, 61 lines, is the binding constraint. Eleven entries, each already at two sentences or fewer, each still carrying its figure (174/84, 15 loops at 5.5% up from 2.7% against a 6% to 7% target, seven threads against a target of eight, 33 points, 15%) and its reason. Reaching 200 would mean dropping roughly four entries or stripping the reasoning from all of them.
• Separability, 59 lines, of which the output block is 10 and cannot be shortened, plus two subsection headings and the four required parts.
• Two tables, 12 lines, both explicitly protected.
• Untouchable sections plus the header are 91 lines, so the floor under these constraints is around 260.

Verified mechanically: every one of the 11 limitation entries is ≤ 2 sentences, the method paragraph is exactly 3 sentences, "It failed the first time it ran" is 1 paragraph, "Why it does not fail the build" is 2 paragraphs, and the output block is present.

Figures

Diffed every number in the file before and after. 117 → 116, and the single loss is 01 from sup-01, the worked example you asked me to drop. No numeric figure was lost.

All 40 substantive figures re-verified against live pnpm eval, pnpm stats and pnpm separability.

apps/web/README.md

It claimed "Status: stub. No implementation until Phase 3" and claimed Tailwind. Both false. Rewritten to describe what is actually there:

• Next.js App Router with output: "export", so the build emits static files and the deployment is a file server
• exactly two build-time imports: results/viewer-results.json and SCOPE_TEXT from @openloop-bench/eval/scope, with a note on why the scope text is imported rather than duplicated
• no model calls, no API routes, no database, no auth, no environment variables (process.env appears nowhere in the package)
• plain CSS in app/globals.css with custom properties on :root, no Tailwind, no CSS-in-JS, no UI library
• the "what this is not" scope limits kept from the original

One correction to the brief: the package does hold one piece of state, the failure gallery filter dropdowns in failure-gallery.tsx. Rather than repeat "no state", the README states this precisely and notes everything else is a server component over committed data.

Also

Root .gitkeep deleted, an empty placeholder from the initial commit.

Verification

• pnpm validate: 200 threads, 273 loops, 510 spans, all resolve. PASS
• pnpm check: 229 tests passing
• pnpm --filter @openloop-bench/web build: static export succeeds
• No em or en dashes in either README

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
openloop-bench	Ready	Preview	Aug 15, 2026 8:54am