# What this lane got wrong, and what stopped it Written 2026-09-04, the last step of this lane's work. Rendered on the site verbatim. Every claim below carries the path of the receipt that settles it. Nothing here was volunteered by a reviewer or found in review: each item was found by something running and failing. Where a number is given, it was measured this turn, not recalled. Read this before the theorems. If it is not the most useful page here, the rest is not worth much. --- ## 1. Nine claims that would have shipped, and the check that stopped each This is the part that matters, because in every case the wrong version was already written down. **A hero sentence citing a number its own receipt does not contain.** The first draft of the site package said the covert-bits ceiling was "corrected upward from 34.55", with `artifacts/covert_fbl_v98c.json` as its receipt. That artifact holds the corrected figure and not the retracted one; 34.55 lives in the dossiers. **Stopped by:** `scripts/site_package_check.py`, the rule that every load-bearing number in a claim must appear in that claim's own receipt. The claim was changed, not the check, and the retracted figure moved to a row whose receipt does hold it. **A theorem stated more generally than it was proved.** The register says a duplicate-eliminating receiver "must distinguish 3^w states". The Lean kernel checks `2^w ≤ s` for **every** width and the sharper `3^w` bound **only at width two**. The general case was stated and never machine-checked, and it would have gone onto the site as the hero. **Stopped by:** reading the elaborated statements in `artifacts/backends/lean_credit_audit.json` rather than the theorem names — which is precisely what this lane built that audit to do, and it caught this lane. `dataroom/STALE_CLAIMS.md` S29. **A headline that the Linux kernel refuses our compiled monitor.** It does not. Our loader applied none of the object's twelve relocations, so the kernel read through a literal zero we handed it. **Stopped by:** `llvm-readelf -r` on our own object, run because the result looked too good. We retracted it from a relocation table before anyone checked; another lane then measured the same object accepted at 182 instructions under a relocating loader. Our controls had proved the verifier was *reached* and never that the loader was *faithful*. `monitorgen/artifacts/kernel_load.json`, `NEGATIVE_RESULTS.md` §S06-6. **The acceptance figure is not ours to receipt:** `182` appears nowhere in this repository (`grep 182 monitorgen/artifacts/kernel_load.json` returns nothing). It was measured by another lane on its own hardware and recorded outside this tree. What this repository can regenerate is the rejection under its own loader — and a number whose receipt we do not hold is a number a reader should discount accordingly. **A defect reported against another repository that was not one.** Our scanner paired a producer with an artifact it merely *names* as a baseline witness and never opens. **Stopped by:** the target lane checking before acting on it. Two further notes against us: the reason we were given for the refutation was itself wrong (that script does carry the option it was said to lack), and we said so rather than accepting the tidier story. **Three flags out of five sent to other lanes as findings.** Two were real. Three were numeric co-occurrences between an option and a field sharing no name — 24 rows from 3 seeds is not a defect. **Stopped by:** hand-checking every flag against the producer's source before anything was addressed to anyone, and then a confidence tier that separates a name-matched pairing from a coincidence. **A public test-count badge derived from a broken collection.** The tool ran `pytest --collect-only` under an interpreter that cannot import the packages. pytest said `14 tests collected, 2 errors`; the regex took the 14 and returned it as authoritative, for a suite of 67. **Stopped by:** comparing the tool's output against the suites we had just run by hand. The badge said 849. It is **817**. **A self-test that destroyed the artifact it was testing.** The `--control` leg pointed the census at an absent directory, and the skip branch wrote its stub over the real result. **Stopped by:** `git status` showing a tracked artifact modified after a control run. This is the same defect this lane had fixed in a different file three steps earlier, reappearing inside the instrument meant to catch it. **Six `pyproject.toml` files broken by our own fix.** A regex insertion leaked a backslash escape and TOML stopped parsing on every interpreter. **Stopped by:** the Python 3.9/3.10 import control we had just written and run, within a minute of making the mistake. **A paper edit that would have shipped stale.** `oss/README.md` declares `paper/` a copy of `protocol-bench/paper/`; the edit went into the copy. **Stopped by:** a file-by-file blob comparison against the live repositories before pushing. A declared-copy guard now blocks staging while the two differ, and it was proved to fire by planting a drift. --- ## 2. Three filters we wrote that hide true things, with the cost as a number A filter that only removes noise is a filter nobody has measured. **The estate scanner's write-context rule costs 93 percent of its reach.** After the refutation above we required a producer to be shown *writing* an artifact before pairing with it. Precision went to 1 finding out of 1 verified. Recall collapsed. Across the nine repositories the scanner walks **14,386 JSON files** and now examines **11 producer–artifact pairs** — down from 168 under the broadest pairing and 59 under the version that was sealed before any foreign tree was read. A producer that writes through a helper function this cannot follow is invisible to it. Four of the nine repositories yield zero pairs — f2, f6, lattice and vllm — and that is a property of our scanner, not of those repositories. `artifacts/backends/scope_defaults_estate.json`. **Our own package checker verifies 5 percent of the numbers on the page.** It compares a number to its receipt only for decimals or integers of four digits or more, and only inside six named fields. The package contains **242 numeric literals**. **12** are compared against a receipt. **1** is exempted by name with a stated reason. **229 are never compared to anything.** The rule that makes the check meaningful is also what makes its coverage small, and a reader should assume any two- or three-digit number on the page is unchecked by machine. Measured with `scripts/site_package_check.py`'s own regex against `site-program/package.json`. **The MCP census refuses more than it counts.** To stop counting audit scripts that merely quote `@mcp.tool` in a comment, decorators must appear at the start of a line; to stop the scanner counting itself, its own file is excluded by path. Both rules are right and both discard true mentions. The census reports **10 servers, 0 declaring a hint**, and it reached that number only after **four** detector blind spots were caught by a leg that requires it to find every server already known to exist — the first version reported zero servers for three repositories that each ship one. `artifacts/backends/mcp_annotation_census.json`. --- ## 3. The numbers on that page we are least sure of, and what would settle each **The 3^w attribution floor.** We state it for every width and the kernel has it at width two. What would settle it: prove `attribution_floor_at_width_w` for general `w` in Lean, or withdraw the general form. Until then the site says what the kernel read. **5.673 dB per decade, and the 18.7858 / 26.4421 dB walls.** Exact inside a declared model — Rician K = 10 dB, no diversity, rate 1/2, CSIR, no CSIT — and that model is a choice. What would settle its relevance: a measured link at 28 GHz with a real front end, which this lane has never had. Nothing here is over the air. `artifacts/wall_vs_reliability.json`. **T\* = 0.17707 ms.** A Bessel correlation crossing against a chosen minimum usable correlation. The crossing is exact; the threshold is a modelling decision, and a different one moves the number. What would settle it: a paired over-the-air experiment at 28 GHz and 10 m/s. `artifacts/staleness_floor_v93.json`. **57 informative impossibility slots.** 62 are impossible and 5 of those sit on compositions with no modelled behaviour, so we quote 57. That split is our own judgement about which wins are vacuous. What would settle it: an independent reviewer re-deriving the degenerate set from the artifact. `artifacts/backends/impossibility.json`. **0.840000503469 as "84.0% of the feasible savings".** The artifact holds three variants — 0.8459, 0.8415 and this one — and we quote the complex-unconstrained figure. The number is not wrong; the choice of which of three to lead with is ours. `artifacts/rounds_converse_v54.json`. **"All nine public repositories have green CI."** True when measured on 2026-09-04 at the commits named. CI state is not a property of a commit alone; a dependency change turns it red without anyone touching the code. Assume it decays. **The 383,805-word distributed I/O floor.** A model composed with a published single-device theorem that is not ours, validated against brute-force minima at pinned configurations. It has never been compared against a measured interconnect. `artifacts/distributed_attention_floor_v101.json`. --- ## 4. What a buyer's engineer finds in the first hour that we did not put on the page **The file the page invites them to check is not there.** The verifier section names `oss/protocol-bench/examples/bfs_submission.json` as the sample certificate to drop in. It was created at the last step, committed to the private repository, and never pushed. Fetching it from the public repository returns **HTTP 404**. The invitation is addressed to a reader who cannot accept it. We found this in the last hour of the last step, added a check for it, and left that check **red** rather than quietly pushing the file — nothing was permitted to leave the repository in this step. One line on the next push list fixes it. `scripts/site_package_check.py`, the reachability leg. **The first command they type fails.** Clone `minicheck` and run `python3 -m pytest tests -q`, which is what anyone does: **9 errors during collection**. The suite needs an editable install first, the README's own test line says so, and the badge says 275 tests pass. Both are true and the first impression is a broken repository. **The version in the source is not the version on the index.** `pyproject.toml` says `0.4.1`; PyPI serves `0.4.0`. The README correctly says 0.4.0 — a test enforces it — but a reader who trusts the source tree will install something that does not exist. **Three checks in this repository are red and stay red.** The published Merkle closure root no longer equals what the live tree folds to; a claim-trust guard is orphaned; one gate's vocabulary contradicts four register rows. Each is disclosed, none is repaired, and repairing the first means re-pinning a certificate, which is not ours to do. **Most of the evidence does not survive a clone.** 15 of 17 radio runs are git-ignored. 589 tracked artifact-like files are named in no ranking document. The manifest pins 385 leaves; the repository holds far more than that. **The benchmark is small and one of the datasets is synthetic.** protocol-bench is 15 tasks, of which 2 are violations. specforge is generated, so its difficulty is a property of our generator. Neither is a corpus, and a system tuned on either will overfit within a week. **There is nothing to buy.** No customer, no revenue, no pilot, no third-party audit, no over-the-air measurement, no silicon. The two domains bought on 2026-09-04 are domains, not products. --- ## 5. The program that grades us made the same class of error, and that belongs here too An estate that publishes only its own errors is curating. These were made by the orchestrator of this program, and two of them touched this lane directly. A census of which of the estate's packages are on PyPI was answered **from thirteen recalled names instead of an enumeration of the trees**, and both halves of the answer were wrong. The true count, enumerated and probed: nine packages, eight predating this program — four of them this lane's, all published 2026-07-30 — and one published by it. That is the rule this program enforces on every lane, broken by the instrument that enforces it. A lane caught it. A concurrency fix was declared done and was not: making each write atomic does not make a read-modify-write atomic. Seven lanes' state records were silently erased and four turns died after running and being paid for. A lost update announces nothing, which is why it needed a control that reproduces it — 1 of 7 records survive under the old pattern, 7 of 7 under the new. Three smaller ones that reached this lane: a branch was asserted never to have been pushed when it was pushed on eight of nine lanes, and we read the remote ourselves and reported the contradiction; a ruling required `url: null` for artifacts under a nonexistent organisation, which cannot validate because the schema types that field as a string; and a stranger-verifier claim was validated from a developer's environment with the tools already on the path. --- ## 6. This file was checked twice, and the second pass changed it This estate's record is that one pass was never enough, so this program did not assume two were. The second pass re-read every sentence above against its receipt rather than against memory, and it found five things — two of them errors in the first draft of this very file. **It said two of the nine repositories yield zero pairs. It is four** (f2, f6, lattice, vllm), read back from `artifacts/backends/scope_defaults_estate.json`. The sentence understated the reach this lane's own filter gave up, in the section whose entire purpose is to state that cost. **It said "twelve sealed bars", and there are eleven files**, of which nine were sealed during this program and two predate it (`ls artifacts/prereg/`). That figure had been repeated across several of this lane's own step records before anyone counted the directory. A recall figure quoted from an earlier step was incomplete: the write-context filter was described as taking pairing from 168 to 11, which is true of the broadest pairing but omits that the **sealed** version examined 59, and both belong in the sentence. A "first hour" item asserting that bare `pytest` fails was written from expectation and then actually run against a fresh clone, which is where the exact figure of 9 collection errors comes from. And the reachability of the sample certificate had not been checked at all until the second pass asked how a stranger would obtain it — which is how the 404 above was found, at the last possible moment, in the deliverable of the previous step. Three of those five are counting errors in a document about counting errors. That is the honest rate, and it is why the number of passes is written down. The second pass is why this file names a defect in the step that produced the page it accompanies. --- ## 7. What is still not true Nothing of this lane's evidence is public. The repository is private and returns 404 to a stranger, so the register, the sealed bars and every correction above are visible to nobody outside. What is public is tooling and two small benchmarks. Four packages sit one patch version behind the code in this tree, because uploads are blocked and we will not publish a README quoting a version the index does not serve. Two floating action tags point at commits three and two behind, the older carrying a real path-resolution bug, and we did not move them because moving a published tag is a forced update. Both browser demos still vendor the old and new wheel at once. The strongest offline verifier this lane has — 49 conformance vectors, stdlib only, exit codes that distinguish refused from undetermined — is not published, and whether it should be is an open question rather than a claim. `artifacts/prereg/` holds **11** seal files, of which **9** were sealed during this program; the other two predate it. Each records only the git commit at sealing time. A stranger can verify that a seal has not been altered. Without an external time anchor they cannot verify that it preceded its run, and we have not anchored them.