Add the 44%-vs-85% cross-agent harness audit to §3, with deep note

[uipreliga]

Adds one §3 entry plus its deep note under notes/articles/. This is the follow-up @xdotli invited on #61, where he called this case "exactly the kind of war story §3 and §6 are short on." Thanks for merging that one.

Affiliation, per CONTRIBUTING: I work on Coder Eval and co-authored the linked post. Disclosing up front — the tool entry in §5a was self-submitted too.

What it is

We ran Codex and Claude Code over the same 417 tasks and got 44.1% vs 85.3%. Before publishing that, we hand-audited 227 of the 228 failures — no sampling, every label required to cite a mechanism from the run record — and sorted them into four buckets:

Bucket	Count	Share
Harness-unfair	144	63%
Ambiguous	49	22%
Genuine	25	11%
Strict criteria	9	4%

11% of the failures were the model. The ambiguous bucket turned out to be one cluster of 30 tasks downstream of a single broken tool, which puts ~89% outside the model's control. The largest single cause was one flaky file-edit backend: 106 of 228 failures. Two others were our own eval bugs — an activation check that only recognized Claude Code's Skill call (30 tasks lost regardless of what Codex did), and a command check that graded command text rather than effect.

Then we fixed the harness and re-ran: 44% → 76%, against 82%. The gap collapsed.

Why it clears the bar

• Show your work — a per-failure census with counts and named mechanisms, not a "harnesses matter" take. The post publishes the 22% we could not adjudicate rather than allocating it, states its own bias ("the incentive runs toward charity"), and says plainly that a 227-failure hand-audit does not scale and is a one-time instrument.
• Every number is quotable — all figures appear verbatim in the linked post; the note's Verified quotes section pins ten of them.
• It adds something §3 lacks. The section has the thesis from Lee, Brand's AlgoTune ranking flip, Hodgson's teardown, and Anthropic's VirBench spread. What it hasn't had is a practitioner reporting the number they nearly published and then itemizing why it was wrong — including a re-run that closes the loop. The reusable output is the inverse law of gap size: rank per-skill gaps, and the biggest ones are the least likely to be real, because infrastructure failures are binary and total while capability differences are graded. It comes with a decisive test (did that agent's tooling actually install for this skill?) and it points readers at their own most impressive-looking results.
• ⚠️ Caveat, stated in the note: one suite, two agents, one org. The taxonomy and the heuristic travel; the specific percentages should not be read as a general Codex-vs-Claude-Code result.

Housekeeping

• Note follows the house shape — Summary / Key points / Verified quotes (verbatim, with source URL) / What it adds / Themes.
• Bumped 143 deep reading notes → 144 in both places it appears (header line and § Deep notes).
• I did not cross-list into §6 even though you mentioned it, since CONTRIBUTING says cross-list only when genuinely warranted. Happy to add it there instead of, or as well as, §3 — your call.
• Did not add a (notes: ...) pointer on the entry; only 2 of 443 entries carry one and both point at a research/notes/ path that no longer exists. Say the word if you want it, and whether that path should be fixed.

Trim the annotation if it runs long for the section — it's above the §3 median and I'd rather you cut it than have it sit oversized.