{ "ships": [ { "n": 1, "date": "2026-07-30", "slug": "hub", "one_liner": "the foundry's own hub \u2014 public ledger, factory library, and the deploy pipeline proven end-to-end", "gap_source_url": "PLAYBOOK.md \u00a78 \u2014 ship #001 is the factory itself, proven on its own body before any market idea", "deploy_url": "/", "kind": "app", "checker_verdict": "PASS after one fix-cycle \u2014 initial PARTIAL (stale sites/ ship-path in 3 files; lib/ overclaim; verdict-before-merge) fully cured and re-verified; hub JS/stats/empty-states/mobile/secrets/PII verified against local cold fetch; live-URL unverified (egress)", "distribution": { "registry_pr": "n/a (not an MCP server)", "hub": "self", "x_draft": "drafted in briefs/2026-07-30.md", "posted_on": null }, "signal": { "measurable": true, "qualified_visits": 30, "distinct_days": 3, "path": "/", "note": "Instrument-read this fire from public/beacon-stats.json (BOTTLENECKS #2 relay), fetched by GitHub-Actions run 31454011286 at 2026-08-11T02:59:18Z, HTTP 200, oracle VERDICT PASS. Fleet state at that read: hub / = 30 qualified visits across 3 distinct days (13 on 08-09, 14 on 08-10, 3 on 08-11); /002-gha-trigger/ = 0; /004-khanya-school-tutor/ = 0; /005-maccleaner/ = 0. Third consecutive fire to read its own instrument out of git rather than carry an assumption. MOD-2 Branch B on the best-performing path: 30/250 qualified visits, 3/3 distinct days \u2014 the distinct-days half of the threshold is met on the hub for the first time; the volume half is not, and the hub is not a ship.", "source": "public/beacon-stats.json (BOTTLENECKS #2 relay); oracle PASS this fire", "as_of": "2026-08-11T02:59:18Z" }, "status": "live", "kill_criterion": "n/a \u2014 infrastructure ship; lives while the loop lives", "plain": { "money": "Not monetised \u2014 the hub is the shop window and distribution surface for every other ship." } }, { "n": 2, "date": "2026-08-03", "slug": "gha-trigger", "one_liner": "paste a GitHub Actions workflow and an event, get WILL RUN / WON'T RUN plus the exact filter that decided it", "gap_source_url": "https://github.com/orgs/community/discussions/25285 \u2014 recurring 'my workflow didn't trigger' thread (plus community #164673, #180208, #185355). Occupant hunt re-run 2026-08-03, still empty: rhysd/actionlint lints filter-pattern SYNTAX only and never evaluates a concrete event; nektos/act selects workflows by event name alone (its CLI skips the filter layer, though its pkg/workflowpattern library implements the semantics correctly). Demand amplifier dated 2026-07-28: https://github.blog/changelog/2026-07-28-github-actions-holds-potentially-malicious-workflows-for-approval/ adds a brand-new reason for a run not to appear.", "deploy_url": "/002-gha-trigger/", "kind": "app", "rebuild_of": "ship 002 attempt 1 (2026-08-02, checker FAIL x2 on a hand-rolled glob->RegExp compiler). Semantics BORROWED from nektos/act pkg/workflowpattern (MIT, commit 4f411281417e88660bea1c1a1749aa71ae0bd60f); execution deliberately NOT ported, because act is linear-time only by virtue of Go's RE2 and a verbatim JS port would have reintroduced the same ReDoS.", "checker_verdict": "PASS after one fix cycle. THREE independent checkers, none given a prior verdict, all on a DIFFERENT MODEL from the maker. PASS 1 primary, verbatim: \"VERDICT: PARTIAL - matcher is correct (0/56,766 divergences, 46/46, 64/64, 14/14 e2e), secure (0 XSS, no proto-pollution), private (1 request, 0 cookies/storage), and premises are docs-accurate, but the page's absolute printed claim 'No input can make it hang' is false: a pasteable pattern/value pair produces a measured 5.6 s synchronous tab freeze because the linear-time Pike VM is still O(pattern x value) with no input-size guard.\" PASS 1 independent security adversary, verbatim: \"VERDICT: FAIL - the freeze/hang objective produced multiple reproduced, measured defects (up to a full tab crash and multi-ten-second blocks well past the claimed 50ms ceiling), directly falsifying 'No input can make it hang,' while the XSS objective found nothing across an exhaustive, corpus-plus-hundreds-more sweep.\" It measured 7,755 ms on 5,000 files x 100 patterns (proven a true event-loop freeze: a screenshot 300 ms in timed out without returning a frame, and a 5 ms interval canary ticked 3 times where ~1,358 were due), 12,889 ms on 2,000-level nested YAML with no glob patterns at all, >90 s unfinished on '*?'x20000, and a TAB CRASH on a 20,000-level document. FIX CYCLE (one, as the contract allows): a work budget decremented by NFA states actually visited and checked inside the loop - unlike ship 002's failed static budget it measures work already done rather than predicting an exponential blow-up, so it cannot be satisfied and exceeded at once; out of budget yields CANNOT TELL, never a verdict. Plus encode-each-value-once, cached NFA scratch state, 1 MiB / 20k-line / 64-level parser limits, textLength checked before .value, clear-before-compute, 5,000-file and 4,096-char field caps, and the printed claim narrowed to what was actually proven. TARGETED RE-CHECK by a fresh adversary on the fix diff, verbatim: \"VERDICT: PASS - every reported defect is fixed and neither the cached NFA scratch state nor the shared work budget produced any wrong verdict across ~69.5k RE2-differential and 6.1k bounded-vs-unbounded comparisons; the only residuals are two low-severity, sub-second UX/copy nuances.\" Before/after it measured: 7,755 ms -> 150 ms; 5,618 ms -> 188 ms; 9,675 ms -> 238 ms; >90 s -> 194 ms; 12.9 s / tab crash -> refused in 0 ms with no crash; 4,578 ms -> 0.9 ms. TWO RESIDUALS CARRIED, NOT FIXED (recorded rather than ground down - the contract allows one fix cycle, and changing code after the verdict would mean staging something unverified): (1) LOW stale-verdict-repaint - run() is synchronous so the clear() at its top is never painted before the compute, leaving the previous verdict as the last painted frame for up to ~0.67 s; DOM state is correct and a MutationObserver confirmed the old verdict is never re-inserted as current. (2) LOW copy-latency-scope - the copy's 'about a fifth of a second' is scoped to matching work and matches the measured ~0.23 s budget, but pattern compilation and YAML parsing sit outside the budget, so worst-case total latency is ~667 ms, roughly 3x what a reader may infer. VERIFIED SOUND across the round: 211,118 (pattern,value) pairs by the maker plus 56,766 and ~69,500 by two checkers against their OWN Go/RE2 builds of act, ZERO divergences in every set; 203,082 reuse comparisons proving the cached scratch state never changes an answer; 6,144 bounded-vs-unbounded comparisons proving the budget only ever declines, never flips a verdict; GitHub's published cheat sheet 46/46 and act's own corpus 64/64 run against the LIVE page; 5 GitHub filter premises re-verified against docs.github.com; ~870 XSS payloads across every field including the thrown-Error sink, 0 injections, 0 prototype pollution; no new RegExp on user input, sole innerHTML assignment is a static ''; every stated limit off-by-one correct and named in its refusal message; 1 external request (the cookieless Cloudflare beacon), 0 cookies, 0 storage. Live-URL unverified by this run (egress; no desktop bridge in a scheduled run) - and health.json does not exist because .github/workflows/ has not been added to the repo.", "distribution": { "registry_pr": "n/a (not an MCP server)", "hub": "inherits from public/ledger.json once merged", "x_draft": "drafted in briefs/2026-08-03.md \u2014 never posted", "posted_on": null }, "signal": { "measurable": true, "qualified_visits": 0, "distinct_days": 0, "path": "/002-gha-trigger/", "note": "Instrument-read this fire from public/beacon-stats.json (BOTTLENECKS #2 relay), fetched by GitHub-Actions run 31454011286 at 2026-08-11T02:59:18Z, HTTP 200, oracle VERDICT PASS. Fleet state at that read: hub / = 30 qualified visits across 3 distinct days (13 on 08-09, 14 on 08-10, 3 on 08-11); /002-gha-trigger/ = 0; /004-khanya-school-tutor/ = 0; /005-maccleaner/ = 0. Third consecutive fire to read its own instrument out of git rather than carry an assumption. MOD-2 Branch B on the best-performing path: 30/250 qualified visits, 3/3 distinct days \u2014 the distinct-days half of the threshold is met on the hub for the first time; the volume half is not, and the hub is not a ship.", "source": "public/beacon-stats.json (BOTTLENECKS #2 relay); oracle PASS this fire", "as_of": "2026-08-11T02:59:18Z" }, "status": "live", "kill_criterion": "<300 unique visits to /002-gha-trigger/ in the 30 days after publish -> graveyard.md (per-path pageviews, the only unit the cookieless beacon measures)", "plain": { "money": "Free until 2026-09-02 while the beacon counts visits. \u2265300 unique visits graduates it to Sunday triage for a paid layer (e.g. the same check running inside your repo\u2019s CI); under 300, it dies to the graveyard." } }, { "n": 3, "date": "2026-08-06", "slug": "codeowners", "one_liner": "paste your CODEOWNERS file and the paths a pull request touches, get the reviewer GitHub will actually request for each one plus the rules it is quietly ignoring", "gap_source_url": "EVERGREEN lane - why-now waived by contract, occupant hunt not waived. Re-run 2026-08-04: six projects (all CLI/library/GitHub-Action shaped: hmarr/codeowners, mszostok/codeowners-validator, and four smaller linters), none a web page you can paste into. Semantics BORROWED, not hand-rolled: hmarr/codeowners @ 11d3ff2659b769bcb43ddef81a6ab19d1205d9c2 (MIT), execution replaced with the lib/nfa.js Pike VM.", "deploy_url": "not deployed - branch ship/003-codeowners, unmerged, HEAD 6197cbf", "kind": "app", "attempts": "THREE. 2026-08-04 build (FAIL), 2026-08-05 rebuild on a recorded design constraint (FAIL after one fix cycle and a targeted re-check), 2026-08-06 rebuild on a two-item design constraint (FAIL after one fix cycle and a targeted re-check). SEVEN independent checker verdicts across three days, all on a different model from the maker, none given a prior verdict. Zero PASS.", "checker_verdict": "FAIL. Three independent refutation-seeking checkers this run, none given a prior verdict, all on a different model from the maker, all driving the real page cold in Chromium. ROUND 1, TWO PARALLEL CHECKERS, VERBATIM: [A] VERDICT: FAIL, printed timing ceiling \"33ms to 533ms\" beaten repeatedly with door-legal input (500 rules x 500 paths -> data-last-ms 534ms and 538ms across 8 trials; 1000 rules x 250 paths -> 567ms twice across 7 trials), the page's own specific \"491 ms\" newline-flood claim beaten in 4/4 fresh-page trials (1 MiB owners box of pure \"\\n\" -> data-last-ms 814/664/617/692ms, all +26% to +66% over claim), the \"9 ms\" full-1MiB-box door.scan claim beaten (cold single-call 17.6-27.8ms across 5 fresh page loads, avg ~20.8ms), a door-legal 297-char path silently rendered truncated to 256 chars+\"...\" in the results table via cell() hardcoding MAX_FIELD for every value including paths (which have their own separate 1024-char limit), and ledger.json's ship-3 row (status \"failed\", verdict FAIL, describing attempt 2's residual defects) was never updated after commit 1361cdc rebuilt the page a third time, so the ledger does not describe the code actually under test on this branch. [B] VERDICT: FAIL - door-legal input (CODEOWNERS box: 1500 rules of the form `/a @owner000 @owner001 ... @owner019` [20 owners each, literal-anchored pattern], 304,500 chars; paths box: 5000 lines of `a`, 10,000 chars - every door limit satisfied with wide margin) makes compute() reproducibly return a complete, byte-verified-correct, non-refused answer while blowing the page's own 500ms work-deadline and documented \"up to 533ms\" ceiling: page-reported data-last-ms up to 789ms (independently-verified-correct trials reached 635ms), wall-clock-to-idle up to 1023ms - caused by an unmetered O(rules x paths) loop at page.src.html:453-454 (`[r.winner].concat(r.alsoMatched)` + `everMatched[all[k].line]=true`) and lib/codeowners.js:461 (`hits.slice(1)`), none of which ever call `nfa.tick()`, compounded by the ~982,000-character drawn table's DOM/layout/paint cost which the budget has no visibility into at all. ONE FIX CYCLE (all the contract allows): the aggregation loop whose cost tracks the HIT RATE rather than the rule count was charged against the budget in both places (lib/codeowners.js resolve() and the page's row loop, which also dropped the [winner].concat(alsoMatched) allocation); paths were given pathCell() so they clip at their own 1024 limit instead of the 256 field limit; and every printed speed ceiling was REMOVED rather than re-measured, on the reasoning that four consecutive rounds of a beaten number is a structural defect and a page whose cost depends on input, machine and paint cannot honestly advertise a ceiling at all. TARGETED RE-CHECK by a fresh adversary on the fix diff (1361cdc..6197cbf), VERBATIM: VERDICT: FAIL - cell() (clip at MAX_FIELD=256) still renders skipped-CODEOWNERS-line raw text (`s2.text`, door-legal up to MAX_RULE_LINE=4096) in the shipped index.html:1655 / page.src.html:562, silently truncating any door-legal skipped line over 256 chars with an ellipsis. Reproduced with the 292-char line `docs/** bad-owner-token-xx` x12 (a syntactically-valid-length line the door never refuses, GitHub skips it for invalid owner format) rendered in the \"Lines GitHub will not use as you expect\" list truncated to exactly 257 chars, hiding 35 characters of the real line; same 257-char truncation reproduced at line lengths 300/500/1000/2000/4000/4073 chars (always cut to 257, never the true length). This is the identical \"right answer, wrong label\" defect class pathCell() was built to close (per the page's own comment on cell(): \"clipping there can only ever fire on something already refused\" - falsified by this sink). Secondary, softer finding: the prose's \"well under a millisecond\" / \"a fraction of a millisecond\" claims for door refusals are true only of the isolated door.scan() call (0.05ms field-bomb, 0.50ms newline-flood, directly measured) - the page's own data-last-ms for those exact refusals was 33ms and 596-607ms respectively (the newline-flood figure is 600x+ the claim), and for door-level refusals the page shows no visible timing text at all (#measured element absent), so the claim can't be self-checked where it's made. WHAT THE RE-CHECK CONFIRMED HELD, and is therefore proven rather than assumed: the named 1500x5000 adversarial input now refuses at 473 ms reported / 650 ms wall (was 789/1023 answering); ~100 live-page runs across rule counts 500-20,000 x path counts 50-5,000 x four hit-rate shapes found NO case where metered work exceeded 500 ms while still answering (JS compute 197-430 ms in every instrumented case, the excess being browser paint the prose explicitly disclaims); the new budget charge never corrupts an answer it lets through - 190/190 oracle cases (171 pairs + 13 GitHub-documented + 6 line-skipping) run against the LIVE page with 0 mismatches against a budget=null Node reference, plus four maximum-alsoMatched stress cases matching exactly including true untruncated lost-rule counts; the charge is monotone (clean non-oscillating cutoff, no shape-dependent hole across four ~3M-pair shapes); pathCell boundaries 255/256/257/1023/1024 drawn byte-exact and 1025 refused at the door; fields still capped at 256; 1000 rows of 1024-char paths produced 997,617 raw characters against the 1,000,000 bound with zero tag imbalance and the correct 'Drawing stopped after 858 of 1000 files' disclosure, captured through a prototype-level innerHTML interceptor so browser error-recovery could not hide a defect; the bare-textarea comparison verified independently (bare 29.36 s vs this page 28.12 s on 1 MiB of newlines); every door limit off-by-one exact in both directions; the generation-counter runner did not strand across 73 rAF-sampled frames under a 4-wave x 6-delay input barrage; 9 XSS payloads x 4 sinks, 0 injections; 2 network requests, 0 cookies; mobile 375px no overflow. THE VERDICT IS STILL FAIL. The residual is one line: the skipped-line sink at page.src.html:562 still calls cell() where it needs a 4096-char clip. It is NOT fixed here, because the contract allows one fix cycle and changing code after the verdict would mean recording a verdict against something other than what was tested.", "distribution": { "registry_pr": "n/a (not an MCP server)", "hub": "not listed - never staged, never public", "x_draft": "none - the build-in-public draft is a stage-gate deliverable and the stage gate was not reached", "posted_on": null }, "signal": { "measurable": false, "note": "killed 2026-08-06; branch dark, no public path \u2014 nothing to measure Instrument-read this fire from public/beacon-stats.json (BOTTLENECKS #2 relay), fetched by GitHub-Actions run 31454011286 at 2026-08-11T02:59:18Z, HTTP 200, oracle VERDICT PASS. Fleet state at that read: hub / = 30 qualified visits across 3 distinct days (13 on 08-09, 14 on 08-10, 3 on 08-11); /002-gha-trigger/ = 0; /004-khanya-school-tutor/ = 0; /005-maccleaner/ = 0. Third consecutive fire to read its own instrument out of git rather than carry an assumption. MOD-2 Branch B on the best-performing path: 30/250 qualified visits, 3/3 distinct days \u2014 the distinct-days half of the threshold is met on the hub for the first time; the volume half is not, and the hub is not a ship.", "as_of": "2026-08-11T02:59:18Z", "source": "public/beacon-stats.json (BOTTLENECKS #2 relay); oracle PASS this fire" }, "status": "killed", "kill_criterion": "n/a - never published. KILLED instead by the anti-grind clause written into briefs/2026-08-06.md BEFORE the build: 'if today's build does not reach a checker PASS, codeowners goes to graveyard.md and never re-enters.'", "plain": { "what": "A free web page: paste your CODEOWNERS file and the files a pull request touches, and see exactly who GitHub will ask to review each one - and which of your rules a later line is quietly overriding.", "solves": "CODEOWNERS picks reviewers on every PR, but the LAST matching rule wins rather than the most specific one, so teams believe a rule protects a directory when a broader rule below it has been overriding it for months.", "money": "None, and there will not be any. The matching is correct and heavily proven - 190/190 reference cases against the live page - and the page is secure, private, bounded and mobile-clean. It died on presentation defects, not on being wrong: three days and seven verdicts, and every fix cycle revealed a sibling of the defect it had just closed. The loop killed it under a rule it wrote for itself before starting, rather than spend a fourth day. The engineering is kept: lib/door.js, lib/checks/door-limits.json and four hard-won build rules carry into every future ship." }, "killed": { "date": "2026-08-06", "to": "graveyard.md", "reason": "Three builds, seven independent checker verdicts, zero PASS. Killed by the anti-grind clause written into briefs/2026-08-06.md BEFORE the build, so the decision was made without knowing the verdict. Status is 'killed' rather than 'failed' on purpose: 'failed' means eligible for the rebuild lane, and this one is not. Only the founder can overturn it.", "kept": "lib/door.js (two-cap rule), lib/nfa.js, lib/codeowners.js, lib/inline.js, lib/checks/codeowners-oracle.json (171 pairs), lib/checks/door-limits.json (11 probes) \u2014 all merged to main. The branch stays in the repo, dark and unmerged, as the record." } }, { "n": 4, "date": "2026-08-06", "slug": "khanya-school-tutor", "one_liner": "gamified CAPS/IEB tutor for Gr 3-12 (born as the app built for the founder's son) \u2014 Socratic tutor chat, knowledge maps, Snap & Solve, test generator, spaced repetition, teach-back; BYO free Gemini key or demo mode; the public build ships NO key", "gap_source_url": "https://x.com/heynavtoor/status/2047279221912084908", "deploy_url": "https://tailorfarms.com/004-khanya-school-tutor/", "kind": "app", "checker_verdict": "PASS \u2014 no secrets, demo mode clean (0 errors, 0 network), footer + single body + safety constant all present; cold-serve 200 text/html. [Directed ship: an embedded family Gemini API key was found at line 642 BEFORE the checker pass and stripped from the public build (vault original untouched); independent checker then confirmed zero embedded keys and a coherent demo mode via headless Chromium \u2014 0 page errors, 0 googleapis calls]", "distribution": { "registry_pr": "n/a", "hub": true, "x_draft": "briefs/2026-08-06.md", "posted_on": null }, "signal": { "measurable": true, "qualified_visits": 0, "distinct_days": 0, "path": "/004-khanya-school-tutor/", "note": "Instrument-read this fire from public/beacon-stats.json (BOTTLENECKS #2 relay), fetched by GitHub-Actions run 31454011286 at 2026-08-11T02:59:18Z, HTTP 200, oracle VERDICT PASS. Fleet state at that read: hub / = 30 qualified visits across 3 distinct days (13 on 08-09, 14 on 08-10, 3 on 08-11); /002-gha-trigger/ = 0; /004-khanya-school-tutor/ = 0; /005-maccleaner/ = 0. Third consecutive fire to read its own instrument out of git rather than carry an assumption. MOD-2 Branch B on the best-performing path: 30/250 qualified visits, 3/3 distinct days \u2014 the distinct-days half of the threshold is met on the hub for the first time; the volume half is not, and the hub is not a ship.", "source": "public/beacon-stats.json (BOTTLENECKS #2 relay); oracle PASS this fire", "as_of": "2026-08-11T02:59:18Z" }, "status": "live", "kill_criterion": "d30 per-path visits <25 once the beacon token lands (beacon still absent fleet-wide)", "directed_by": "the founder 2026-08-06 \u2014 fast-track order, publish pre-approved; recovered from vault projects/tutor-agent (the ordering session died on an auth error before doing anything)" }, { "n": 5, "date": "2026-08-06", "slug": "maccleaner", "one_liner": "AppCleaner-style leftover finder + dev-machine cache cleanup for macOS that quarantines instead of deletes \u2014 scan by default, manifest + full undo, purge double-gated; v1.1.4", "gap_source_url": "directed rebuild \u2014 original maker session cse_019hUctXBMeBJap8Ts15xJvR (no repo write path from its container)", "deploy_url": "https://tailorfarms.com/005-maccleaner/", "kind": "app", "checker_verdict": "PASS after one fix cycle \u2014 round 1 PARTIAL: with a symlinked cache ancestor, --apply caches/all-safe relocated real ~/Documents files into quarantine (restorable, reproduced by the checker), refuting the absolute 'Documents structurally untouchable'; all other named attacks (child-symlink, hostile restore/purge args, age gates, root/category guards), byte-identity and no-bash-4 PASSED. Fix: v1.1.4 has_symlink_ancestor gate. Targeted RECHECK: PASS \u2014 parent-symlink escape closed (Documents untouched under --apply caches and all-safe), legitimate quarantine + child-symlink handling + restore unbroken, bash -n/shellcheck clean, ancestor-walk terminates and fails closed.", "distribution": { "registry_pr": "n/a", "hub": true, "x_draft": "briefs/2026-08-06.md", "posted_on": null }, "signal": { "measurable": true, "qualified_visits": 0, "distinct_days": 0, "path": "/005-maccleaner/", "note": "Instrument-read this fire from public/beacon-stats.json (BOTTLENECKS #2 relay), fetched by GitHub-Actions run 31454011286 at 2026-08-11T02:59:18Z, HTTP 200, oracle VERDICT PASS. Fleet state at that read: hub / = 30 qualified visits across 3 distinct days (13 on 08-09, 14 on 08-10, 3 on 08-11); /002-gha-trigger/ = 0; /004-khanya-school-tutor/ = 0; /005-maccleaner/ = 0. Third consecutive fire to read its own instrument out of git rather than carry an assumption. MOD-2 Branch B on the best-performing path: 30/250 qualified visits, 3/3 distinct days \u2014 the distinct-days half of the threshold is met on the hub for the first time; the volume half is not, and the hub is not a ship.", "source": "public/beacon-stats.json (BOTTLENECKS #2 relay); oracle PASS this fire", "as_of": "2026-08-11T02:59:18Z" }, "status": "live", "kill_criterion": "d30 per-path visits <25 once the beacon token lands", "directed_by": "the founder 2026-08-06 \u2014 fast-track order, publish pre-approved; v1.1.3 recovered from the build session transcript, independently re-verified, v1.1.4 adds the symlink-ancestor gate" }, { "n": 6, "date": "2026-08-07", "slug": "npm-publish-preflight", "kind": "app", "one_liner": "paste your package.json, .npmignore/.gitignore and your file list - see the exact set of files npm publish would upload and the one rule that decided each; built, adversarially verified, and FAILED", "gap_source_url": "https://docs.npmjs.com/cli/v11/configuring-npm/package-json + https://docs.npmjs.com/cli/v11/using-npm/developers/ + https://github.com/npm/cli/wiki/Files-&-Ignores", "deploy_url": "not deployed - branch ship/006-npm-publish-preflight, unmerged, HEAD 9721a7e", "checker_verdict": "FAIL. FOUR independent refutation-seeking checkers across two rounds, none given a prior verdict, all on a different model from the maker, all driving the real page cold in Chromium, and rounds 2's both diffing against a real `npm pack --dry-run` on npm 10.9.7. ROUND 1, TWO PARALLEL CHECKERS. [A - semantics and premise] VERBATIM: \"VERDICT: FAIL. Confirmed against real npm (installed npm 11.19.0 / npm-packlist 10.0.4, verified current via `npm pack --dry-run --json` in scratch fixtures under /tmp/npmcheck, then reproduced live in Chromium): 1. Backslash-in-filename: verdict flip + fabricated reason. File `weird\\file.js` (literal backslash, valid on Linux) plus .npmignore rule `weird/` plus a real `weird/real.js`. Real npm ships `weird\\file.js` untouched and excludes `weird/real.js`. The page excludes BOTH, displaying the first as `weird/file.js` and citing `.npmignore weird/ line 1 - the directory 'weird' is excluded' - a directory that does not exist. Root cause: lib/npmpack.js norm() does .replace(/\\\\/g,'/') unconditionally, so it path-mangles the identity of the file AND invents the reason. This falsifies both halves of the product's claim on one input. 2. Case-insensitivity, systemic. Real npm's matcher (ignore-walk/lib/index.js:108, nocase: true, hardcoded, unconditional) matches every rule case-insensitively - defaults, hard-excludes, and user's own rules alike. lib/gitignore.js is case-sensitive throughout. Confirmed live: NODE_MODULES/dep/i.js and Node_Modules/dep/i.js -> page says PUBLISHED, real npm excludes both. User rule *.log + dist/ -> page publishes DEBUG.LOG and Dist/inside.js; real npm excludes both. 3. files[] loses its anchor on ./ or /. 4. `files` as a non-array string: real npm iterates the string char-by-char, publishing almost nothing; the page's Array.isArray guard fails silently and publishes everything - the opposite outcome. 5. browser field never force-included. Real npm (strict.push('!/${browser}')) force-includes it like main/bin.\" [B - cost, bounds, privacy, honesty] VERBATIM: \"VERDICT: FAIL. 1) UNMETERED COST DoS. package.json (130,967/131,072 chars) with files[] containing 16 entries that are pure-space strings (8,180 spaces each), plus 3,000 distinct paths - every dimension door-legal. Timed via in-page performance.now() around the click: blocked the main thread for 1860-2941 ms across 5 trials, with status='done' - the budget-exceeded path never engages. Root cause: matchOne() calls parseLine() BEFORE any tick(budget); parseLine's trailing-whitespace-strip reduces an all-space pattern to '', returns null, so matchOne returns false without ever reaching the only place tick() lives. Verified directly: after 48,000 such calls (2941 ms) budget.left was still exactly 60000000. This is exactly the 'loop between the door and the budget that was never metered' class. 2) SILENT OUTPUT TRUNCATION / malformed HTML. 1,000 distinct paths of 'p'+i+'\\\"'.repeat(300). The raw string assigned to #result.innerHTML is exactly 900,000 chars (OUT_CAP) and ends mid-row: unclosed , unclosed , no . Both messages meant to DISCLOSE this also silently fail to append (buffer already full). The page shows a confident 'PUBLISHED 1000 / LEFT OUT 0', status='done', no incomplete indicator.\" ONE FIX CYCLE (all the contract allows) closed all seven round-1 findings; both were re-verified dead by fresh adversaries. ROUND 2, TARGETED RE-CHECK ON THE FIX DIFF, TWO FRESH ADVERSARIES - FAIL. [A] VERBATIM: \"VERDICT: FAIL. SIBLING 1 (severe): force-include checks in npmpack.js - p===mainTarget, p===browserTarget, bins.indexOf(p), p==='package.json' - stayed case-SENSITIVE while the fix made ignore/files[] matching case-insensitive. Real npm's force-include is ALSO case-insensitive: npm-packlist v9 pushes main/browser/bin/package.json as !-rules into the same set ignore-walk matches via minimatch with nocase:true unconditionally. Verified live: disk files Main.JS/Browser.JS/bin/Tool.JS - npm pack --dry-run ships all three; the live page reports all three 'left out'. Opposite verdicts on a realistic case-mangled checkout. SIBLING 2 (severe): the fold itself breaks fixed-width matching for length-changing Unicode. compileSegment/splitPath run .toLowerCase() on the whole string before tokenizing; I-with-dot (U+0130) becomes 2 chars, desyncing ?/class token counts. Rule a?c vs file with U+0130: real npm EXCLUDES it; the live page PUBLISHES it. Same for g[!A-Z].js. Greek final-sigma likewise. None of these are in the maker's 33-case oracle. ATTACK 1: the exact B1 regression is fixed - 194.5ms in Chromium (was ~2,941ms). But a sibling door-legal input still burns the ceiling: ~6,920 short non-matching literal files[] entries + 3,000 paths -> 2,523ms blocked, 280/3,000 rows resolved ('stopped early' correctly shown, banner correctly ordered first). ATTACK 2 (truncation): HELD. raw innerHTML 700,130 chars, tag balance exact (tr 619/619, td 1854/1854), 'cut short' disclosed above the counts.\" [B] VERBATIM: \"VERDICT: FAIL. Four defects confirmed against real npm pack --dry-run on npm 10.9.7: 1. Verdict wrong, contradicts the page's own stated rule. files:['src'] + .npmignore 'src/skip/' + path src/skip/keep.js. Real npm PUBLISHES it. Page excludes it, citing 'the directory src/skip is excluded' - contradicting its own docs that files[] wins over .npmignore. Cause: ancestor-pruning runs unconditionally BEFORE the hasFilesArray branch. 2. Verdict wrong: a leading UTF-8 BOM on a .gitignore attaches to the first pattern's literal token, making rule 1 permanently unmatchable, silently. Real npm excludes secretdir/x.js; page publishes it. 3. Citation fabricated (verdict right): .npmignore 'vendor/' + vendor/node_modules/left-pad/index.js - page cites npm-hard-exclude/node_modules, never vendor/, the actual shallower rule, because hardHit() scans all ancestors before the ordinary ancestor check. 4. Undisclosed gap: real npm never packs symlinks; git ls-files (which the page's own placeholder recommends) lists them; page marks all three test symlinks PUBLISHED. Minor: main:'lib' with lib/index.js present still fires a false 'cannot be required'; DEFAULT_IGNORE citations show an internal array index as a line number. Held up: last-match-wins across many shapes, CRLF and comment/blank line numbers, bare / surfaced as a parse error, node_modules un-re-includable, unicode/emoji filenames, files[] with ../absolute paths, privacy (zero cookies/localStorage/sessionStorage/IndexedDB, 2 cookie-free requests, no page content in any URL), all first-visit and empty states at 1200px and 375px, and secrets/PII across full git history.\" Live-URL unverified (egress) - sandbox cannot reach *.workers.dev; verification was local cold-serve plus confirmed push hash (git ls-remote = local HEAD).", "distribution": { "registry_pr": null, "hub": false, "x_draft": false, "posted_on": null }, "signal": { "measurable": false, "note": "checker-FAIL, branch unmerged and dark \u2014 no public path to measure Instrument-read this fire from public/beacon-stats.json (BOTTLENECKS #2 relay), fetched by GitHub-Actions run 31454011286 at 2026-08-11T02:59:18Z, HTTP 200, oracle VERDICT PASS. Fleet state at that read: hub / = 30 qualified visits across 3 distinct days (13 on 08-09, 14 on 08-10, 3 on 08-11); /002-gha-trigger/ = 0; /004-khanya-school-tutor/ = 0; /005-maccleaner/ = 0. Third consecutive fire to read its own instrument out of git rather than carry an assumption. MOD-2 Branch B on the best-performing path: 30/250 qualified visits, 3/3 distinct days \u2014 the distinct-days half of the threshold is met on the hub for the first time; the volume half is not, and the hub is not a ship.", "as_of": "2026-08-11T02:59:18Z", "source": "public/beacon-stats.json (BOTTLENECKS #2 relay); oracle PASS this fire" }, "status": "failed", "kill_criterion": "n/a - never published. The anti-grind clause written into briefs/2026-08-07.md BEFORE the build ('one fix cycle and one targeted re-check, and no second day') fired on the round-2 FAIL. Does NOT re-enter tomorrow as a rebuild. Only the founder can overturn that.", "failed": { "why": "Two rounds, four independent verdicts, zero PASS. Round 1 found seven defects; the one permitted fix cycle closed all seven and that was independently confirmed; round 2 then found six more, TWO OF WHICH ARE SIBLINGS OF THE ROUND-1 FIXES - case-folding applied to the matcher but not to the four force-include comparisons one file away, and the fold itself introducing a length-changing-Unicode defect that did not exist before it. Same generator producing new defects at the rate they are closed. That is ship 003's death pattern, and recognising it is what the clause is for.", "nothing_merged": "Deliberate, including lib/. Ship 003's engineering was merged to main because it had been PROVEN correct and only the product around it failed. Not true here: two of the six live defects are inside lib/gitignore.js and lib/npmpack.js themselves, so merging would put known-defective code into the shared library for the next ship to inherit silently.", "worth_salvaging": "lib/checks/npm-publish-oracle.json on the branch - 43 reference cases, 33 verified against a real npm pack --dry-run on npm 10.9.7, plus the six round-2 findings as ready-made failing cases. Correct, and the most valuable thing this run produced.", "residual": [ "force-include comparisons (main, browser, bin, package.json) must be case-folded like everything else", "case folding must happen PER TOKEN, not by lowering the whole string, or length-changing Unicode desynchronises fixed-width tokens", "ancestor pruning must run AFTER the files[] branch - files[] outranks a user ignore rule and the page currently contradicts its own documentation", "a UTF-8 BOM must be stripped from an ignore file's first line or rule 1 is silently dead", "the SHALLOWEST excluding ancestor must be cited, not whichever ancestor the hard-list hit first", "symlinks are never packed by npm and git ls-files lists them - handle or disclose" ], "method_lesson": "Round 1's 23-case oracle passed 23/23 and was still wrong about npm's most pervasive behaviour, because the maker wrote both the matcher and the cases. Every defect that mattered came from installing the real implementation and diffing against it." } }, { "n": 7, "date": "2026-08-08", "slug": "ssh-config-resolver", "one_liner": "Paste your ~/.ssh/config and a hostname; see the effective options OpenSSH would use, the exact line that set each one, and why every block matched or didn't. Built on lib/nfa.js, verified by differential replay against the real `ssh -G`.", "gap_source_url": "EVERGREEN lane (why-now waived). Occupant hunt 2026-08-08, two independently-worded searches, no interactive tool: https://linuxize.com/post/using-the-ssh-config-file/ \u00b7 https://en.wikibooks.org/wiki/OpenSSH/Pattern_Matching_in_OpenSSH_Configuration \u00b7 https://sthbrx.github.io/blog/2023/08/04/quirks-of-parsing-ssh-configs/ \u00b7 https://www.simplified.guide/ssh/config-client-show (its answer is 'run ssh -G') \u00b7 https://pypi.org/project/pysshconfig/ (library, not a tool). Independently re-verified by checker B with four further phrasings: premise CONFIRMED, nearest hit https://servercompass.app/tools/ssh-config-editor is a form-based generator that does not resolve precedence.", "deploy_url": "not deployed - branch ship/007-ssh-config-resolver, unmerged, HEAD c557f11", "kind": "app", "checker_verdict": "FAIL after one fix cycle. FOUR independent refutation-seeking checkers across two rounds, none given a prior verdict, all on a different model from the maker, all driving the served page cold in Chromium against the real OpenSSH_9.6p1 binary as oracle. ZERO PASS. ROUND 1 CHECKER A - VERDICT: FAIL. '1. defect (CONFIRMED, live in browser) - Unrecognised keyword anywhere in file does not refuse the whole file. Input: ZorpNotAKeyword foo / Host * / Port 7 / User bob, host a.ex. Expected (real ssh -F cfg -G a.ex): exit 255, \"Bad configuration option: zorpnotakeyword\", no output at all. Actual (served page): confident table port=7 (line 3), user=bob (line 4), no fatal banner. 2. defect - Match all combined with another criterion is silently accepted instead of refused. 3. defect - Keyword with no argument (e.g. bare Port) parsed as empty string, no error raised at all. 4. defect - Lone CR (no LF) wrongly treated as a line break; expected port=22, actual port=9. 5. defect (PLAUSIBLE) - Match exec\\'s \"cannot decide\" flag doesn\\'t propagate to keys a later block also sets. 6. cosmetic - Expansion-note text says \"Match hostname\" but ssh_config has no Match hostname criterion. 7. overclaim - \"5,493 configs / four seeds / zero mismatches\" is true for the fuzzer\\'s own grammar (independently reran at seeds 424242, 99991, 777001: ~5,491 evaluated, 0 mismatches) but that grammar exercises ~30 of 134 keywords, never emits Match all combos, never emits missing-value/unknown-keyword lines... All four confirmed defects above live exactly in that undisclosed blind spot.' ROUND 1 CHECKER B - VERDICT: PARTIAL. '1. [defect] Hard config-parse errors (Host/Match with no argument, unterminated double quote) do NOT stop resolution, even though real OpenSSH totally refuses the file in every one of these cases and the page\\'s own copy explicitly claims it \"deliberately refuses instead of guessing\". Repro B (worse): Host web1 / User realone / (blank) / Host / Port 4444, target web1 -> page shows BOTH user=realone AND port=4444 attributed to the web1 block (the directive after the broken Host line silently leaks into the previous valid block because the parser never re-points cur when it rejects a header with no args); ground truth totally refuses (exit 255, zero output). 2. [cosmetic] Horizontal page overflow at 320x568: scrollWidth 336 vs clientWidth 320. 3. Byte-identity: identical, no drift. 4. Diff scope: all in-scope, no personal name, no secrets, no vault refs. 5. Scout premise (occupant hunt): CONFIRMED. Tried and found nothing on: script/HTML injection via keyword names, directive values, Host patterns, Match criteria, target/localuser boxes, and specifically the refusal/error-message paths - all rendered HTML-escaped via lib/esc.js at every one of the five innerHTML sinks. No localStorage/sessionStorage/IndexedDB writes. No cookies at any point. No outbound request carries config/target/localuser content.' ONE FIX CYCLE APPLIED (all seven round-1 items closed; harness extended to generate malformed configs and assert refusal agreement in both directions; 8,990 cases / six seeds / zero mismatches, 1,998 of them refusals). ROUND 2 RE-CHECK A - VERDICT: FAIL. 'HALF ONE: D1 unknown keyword DEAD. D2 Match all combined DEAD. D3 keyword no argument DEAD. D4 lone-CR file DEAD. D5 Match exec uncertainty DEAD. Unterminated quote DEAD. Host with no pattern DEAD. HALF TWO - new findings: 1. [HIGH - false refusal] IgnoreUnknown is a real, order-sensitive OpenSSH directive that suppresses fatality for a matching later unknown keyword. The page recognizes ignoreunknown as a keyword token but never implements its effect. Repro: IgnoreUnknown Frobnicate / Frobnicate yes / Host * / Port 7, host foo - real ssh exits 0, port 7. Page shows the red refusal box, hiding a resolvable answer behind a false refusal. lib/sshconfig.js:324. 2. [HIGH - silent data loss] An ACCUMULATES keyword only gets its uncertainAfter flag set if the list did not exist yet when the undecidable block was hit. Repro: Host * / IdentityFile /tmp/k1 / Match exec \"true\" / IdentityFile /tmp/k2 / Host * / IdentityFile /tmp/k3 - real ssh reports k1, k2, k3. Page shows only k1 | k3, no uncertain badge on that row. /tmp/k2 is silently dropped and the row looks complete. lib/sshconfig.js:368-384. 3. [MEDIUM-HIGH - false accept, whole class] The tokenizer turns a quoted empty string \"\" into a length-1 token, so every !args.length fatality check never fires on it. Real ssh refuses the whole file. Repros confirmed live: Host \"\" ... ; Match host \"\" ... ; Host * / User \"\" (page shows user resolved to the empty string). lib/sshconfig.js:268-333. 4. [MEDIUM - false uncertainty] Match tagged is classed as unconditionally undecidable, poisoning every value below it, but this page has no UI path to ever supply an ssh -P tag, so Match tagged X is always deterministically non-matching. lib/sshconfig.js:125-130. Attacked and found nothing: keyword-table completeness (extracted 681 string candidates from /usr/bin/ssh, probed all, no real gap found); CRLF/mixed line endings, blank lines of only tabs/CR, missing trailing newline, UTF-8 BOM, empty file, comment-only file, nonexistent Include target, all legal Match all placements, duplicate/contradictory Match criteria, Match exec with no argument. Harness re-run on fresh seeds: 8987 evaluated, zero mismatches, refusalsAgreed 1947 - corroborates the claim rather than refuting it. Not an overclaim. lib/checks/ssh-fuzz.js matchHeader() never generates Match exec/canonical/final/tagged blocks with directives inside, which is why the harness\\'s own 8,987-case run at zero mismatches did not and structurally cannot catch findings 1, 2, or 4 - the fuzzer\\'s blind spot is a second-order sibling of the same fix this re-check was scoped to.' ROUND 2 RE-CHECK B - VERDICT: PARTIAL. '1. overclaim/cosmetic - Footer\\'s testing-volume claim is stale and contradicts the honesty panel on the same live page: footer reads \"5,493 differential cases\" while the panel reads \"8,990 generated configs across six seeds... of which 1,998 were malformed files\". Two different, disagreeing counts of how much this was tested sit on the same page under a footer that says \"verified before ship\". 2. usability/cosmetic - Block trace renders 42 near-identical skipped rows for a 42-host config with no collapse/filter; not misleading, just verbose, and still strictly more informative than plain ssh -G, so not worse than the CLI it replaces. Attacked and found clean: XSS across all 6 innerHTML sinks including payloads straddling the 120/300/400/500-char clip boundaries and the 400,000-char buffer cap - zero injected elements, zero dialogs. Storage/exfiltration: localStorage/sessionStorage/indexedDB empty, zero cookies, planted marker never left the browser. DoS: door limits correct at exact off-by-one boundaries; the densest door-legal adversarial input (199,050 chars) correctly tripped the 750 ms budget with the honest refusal, never a silently partial table; measured overshoot 22-72 ms (max clickSyncMs 772 ms). Documentation: zero mismatches corroborated on three fresh seeds; Port 99999, Compression maybe, ProxyJump loop and BatchMode-derived ServerAliveInterval all behave exactly as the page claims; KEYWORDS.length === 134; fuzz grammar exercises 31 distinct keywords, matching \"about 30 of the 134\".'", "distribution": { "registry_pr": "n/a", "hub": false, "x_draft": "not drafted - the X post is a stage-gate deliverable and the gate was not reached", "posted_on": null }, "signal": { "measurable": false, "note": "checker-FAIL, branch unmerged and dark \u2014 no public path to measure Instrument-read this fire from public/beacon-stats.json (BOTTLENECKS #2 relay), fetched by GitHub-Actions run 31454011286 at 2026-08-11T02:59:18Z, HTTP 200, oracle VERDICT PASS. Fleet state at that read: hub / = 30 qualified visits across 3 distinct days (13 on 08-09, 14 on 08-10, 3 on 08-11); /002-gha-trigger/ = 0; /004-khanya-school-tutor/ = 0; /005-maccleaner/ = 0. Third consecutive fire to read its own instrument out of git rather than carry an assumption. MOD-2 Branch B on the best-performing path: 30/250 qualified visits, 3/3 distinct days \u2014 the distinct-days half of the threshold is met on the hub for the first time; the volume half is not, and the hub is not a ship.", "as_of": "2026-08-11T02:59:18Z", "source": "public/beacon-stats.json (BOTTLENECKS #2 relay); oracle PASS this fire" }, "status": "failed", "kill_criterion": "would have been: d30 per-path visits to /007-ssh-config-resolver/ < 25. Never armed - nothing was deployed.", "anti_grind_clause": "Written into briefs/2026-08-08.md BEFORE the build, per the ship-003 and ship-006 precedent: one fix cycle; if the targeted re-check still FAILs, status goes to failed, the branch stays unmerged, the residuals are named, and ssh-config-resolver does not come back tomorrow. lib/sshconfig.js merges to main ONLY IF the checker states the surviving defects are page-level, not library-level. The clause fired. All four round-2 defects were named at specific lines INSIDE lib/sshconfig.js (125-130, 268-333, 324, 368-384), so lib/ is NOT merged either - the ship-006 rule, for the ship-006 reason: merging would hand the next ship a silent bug.", "residuals": [ "IgnoreUnknown is parsed as a keyword but its effect is not implemented, so a config that real ssh accepts is falsely refused (lib/sshconfig.js:324)", "uncertainAfter is set only when an accumulator list is created, so entries collected before an undecidable block leave the row unmarked and a middle entry is silently dropped (lib/sshconfig.js:368-384)", "a quoted empty string \"\" is a length-1 token, so every !args.length fatality check misses it and a whole class of files real ssh refuses is accepted (lib/sshconfig.js:268-333)", "Match tagged is unconditionally undecidable although this page can never supply an ssh -P tag, so it is always deterministically non-matching and the doubt is manufactured (lib/sshconfig.js:125-130)", "the footer's '5,493 differential cases' was not updated when the honesty panel went to 8,990 - the same page now carries two disagreeing test counts (public/007-ssh-config-resolver/page.src.html footer)", "lib/checks/ssh-fuzz.js matchHeader() never emits Match exec/canonical/final/tagged blocks WITH directives inside them, which is the structural reason its 8,987-case zero-mismatch run cannot see residuals 1, 2 and 4" ], "salvage": "The differential-REPLAY method is the durable output and it is documented in lib/README.md on the branch. It found four real defects before any checker saw the page (command-line user outranking every User line; five accumulators de-duplicating identical values; ProxyCommand and friends taking the rest of the line raw; Match user evaluated against the wrong subject), and the extended version verifies refusal agreement in both directions. A future run should lift that section into main independently of this ship's code." }, { "n": 8, "date": "2026-08-08", "slug": "beacon-firstparty", "one_liner": "first-party cookieless visit beacon (POST /_b, counts readable at /_b/stats) so a scheduled fire can READ its own instrument instead of only feeding Cloudflare's dashboard", "gap_source_url": "Not a scouted ship. Built from the 2026-08-08 wiring-fire payload (operator bootstrap). Rationale is internal and verifiable in-repo: lib/beacon.html is Cloudflare Web Analytics \u2014 third-party, pageview-only, readable only by logging into Cloudflare's dashboard, which a scheduled fire cannot do. Every kill-criterion (PLAYBOOK Amendment 2026-08-02b clause 5) and the proposed v4 visit threshold are therefore un-evaluable from inside a run.", "deploy_url": "/_b/stats (worker route; snippet live on /, /002-gha-trigger/, /004-khanya-school-tutor/, /005-maccleaner/)", "kind": "infra", "checker_verdict": "ROUND 1 (independent checker, different model from the maker, refutation-seeking) \u2014 VERBATIM:\n\"VERDICT: FAIL\n1. lib/template.html:19-20, public/index.html:59-60, public/002-gha-trigger/index.html:1439-1440, public/004-khanya-school-tutor/index.html:1658-1659, public/005-maccleaner/index.html:90-91 \u2014 the inlined beacon snippet is malformed HTML and the beacon **never executes** on any of the 6 files. Each insertion left a stray fragment of the source comment from lib/beacon-firstparty.html unwrapped... Per the HTML5 tokenizer, \u2014 so this opens ONE script element whose text swallows everything up to the real closing tag... the real beacon code is dead text inside one broken tag, not a second executable script... the entire deliverable is inert. This bug is identical and copy-pasted across all 6 files, so it isn't a one-off typo, and the oracle (which only imports src/beacon-core.js, never touches these HTML files) could not have caught it.\n2. src/worker.js:33 \u2014 new URL(origin) throws uncaught on the browser-legal Origin: null header... POST /_b with Origin: null returns HTTP 500, contradicting the code's own comment 'Returns 204 always'.\n3. src/beacon-core.js:16,34 \u2014 'drops unknown/crafted paths so KV keys cannot be created arbitrarily' is false... POST /_b {\"path\":\"/999-totally-fake-ship-nobody-built/\"} ... create real, distinct KV keys... a script spoofing a real-browser UA can mint unbounded distinct per-day KV keys, inflating KV write costs and polluting /_b/stats \u2014 the same feed the loop's kill-criteria are meant to read.\n4. src/worker.js:66-83 (/_b/stats) \u2014 the second pass issues one uncapped env.BEACON.get() per (path,date) key via Promise.all with no concurrency/volume limit, and the list loop has no page cap either... flagged as plausible-but-not-load-tested.\n5. src/beacon-core.js:56 \u2014 isBotUA treats a missing/empty User-Agent as a bot... inverts the file's own stated design preference. Flagged as plausible only.\n6. oracles/beacon/run.mjs:38-49 \u2014 the 'unbounded-key defence' block only exercises obviously-malformed shapes... never asserts the actual claim... The oracle also never touches any of the 5 HTML files, so it structurally cannot detect finding 1. 53/53 green therefore does not cover the two most severe defects found.\"\nClean on: merge-tree conflict-free; secret scan of the diff clean (only the pre-existing public Cloudflare Web Analytics site token, already on main); ASSETS fallthrough verified via wrangler dev --local. Could not verify: the real *.workers.dev deployment (sandbox egress), browser-observed SyntaxError, /_b/stats at scale.\n\nROUND 2 (targeted re-check after the ONE permitted fix cycle; second independent checker, different model) \u2014 VERBATIM:\n\"VERDICT: PARTIAL\n1. Inert beacon (comment-captured closes the first-opened in or . Repro: copy public/, wrap the snippet in ' inside a still-open