Field notebook · No. 0716

Solarolabs

DailyScar

This notebook records measured misses, not a catalog of tools.

The burns

The archive holds one hundred fifty-six numbered scars plus the rest, and twenty-three names are only the runner sort. The numbered record holds 156 of them, S-01 through S-156. Days shows one a day and hides the rest. A burn that shared a date is still a burn.

Written down

  1. S-01 2026-07-05 — Three Chrome-bridge denials were the operator being away, not the operator refusing
  2. S-02 2026-07-06 — The ROLD commands file was lost to disk truncation and had to be reconstructed
  3. S-03 2026-07-06 — A mixed permission dialog let the wrong scope be answered
  4. S-04 2026-07-10/11 — GDX write silently failed on stale connector auth, and the node said it could write
  5. S-05 2026-07-12 — MOUNT HIT: jack_command.html served pinned at a truncated 63,824 bytes
  6. S-06 2026-07-12 — MOUNT HIT: bts_gdx.py read back missing a colon → a FALSE SyntaxError
  7. S-07 2026-07-12 — MOUNT HIT: dash.json / sgh_spend.json gave bogus JSON decode errors
  8. S-08 2026-07-12 — 26 of Chapter 4's 36 references were silently eaten — the list truncated at the letter G
  9. S-09 2026-07-12 — BootUP took only one mount and then stalled twice mid-session
  10. S-10 2026-07-12 — The dashboard launcher started the plain static server, so every panel silently went stale
  11. S-11 2026-07-12 — The §4.10 IP pedestal finding was raised then withdrawn — the chapter was right
  12. S-12 2026-07-13 — A brand-new Origin reader was written while three existing tools sat unpointed-at — and $2.93 of paid search bought facts already in Ch4's own reference list
  13. S-13 2026-07-13 — GDX's drive letter was written as E: instead of X: and the wrong letter sat in the ROLD for a day
  14. S-14 2026-07-13 — The Ch4-review publish stalled because D:\R2Cloner was not in the mount list
  15. S-15 2026-07-15 — D:\ODX\OneDrive cannot be mounted — the step had been failing at every BootUP since it was added
  16. S-16 2026-07-15 — GDX mounts for Claude's file tools but NOT into the bash sandbox
  17. S-17 2026-07-15 — THE RAILS SECTION OF THE ROLD WAS FALSE ON EVERY ROW FOR TWO DAYS — and seeded a $299 phantom credit
  18. S-18 2026-07-15 — [K-96] — a mislabelled fit parameter, and a pointer that called it "THE ONE GATE" for three days after it was resolved
  19. S-19 2026-07-15 — bts_identity.py shipped and appeared in NO index — the exact failure the TOOLS INDEX was created to prevent
  20. S-20 2026-07-15/16 — Hardware and network topology was an item class with no home in any checklist
  21. S-21 2026-07-15/16 — Proposed moving 105 GB to fix a bottleneck that was somewhere else — the operator caught three of these, I caught none
  22. S-22 2026-07-15/16 — TidyUP2 produced two false alarms out of four checks
  23. S-23 2026-07-16 — GEM named the right Google service but invented the path — a 404
  24. S-24 2026-07-16 — Vertex billed 28.5× the visible output — and $0.0296 was paid for a question the free lane answers
  25. S-25 2026-07-16 — A PUBLISH IS NOT A SYNC — the public mirror served a 15-day-old file and every verify returned 200
  26. S-26 2026-07-16 — The sandbox held the ONLY copy of a delivered 2.7 MB deliverable's builder
  27. S-27 2026-07-16 — TOOLS_REGISTRY.json claimed a generator that does not exist, so obeying it froze the index
  28. S-28 2026-07-17 — MOUNT HIT #11 — the mount lied about SIZE: under-reported by 39,584 bytes, and a divergence check returned a clean, specific, garbage answer
  29. S-29 2026-07-17 — MOUNT HIT #12 — a FALSE SyntaxError at line 106 nearly ate bts_paths.py, the resolver everything depends on
  30. S-30 2026-07-17 — The migration's code/history split ate the file that exists to survive the sweep
  31. S-31 2026-07-17 — A handoff recorded backup_to_onedrive.py as "Patched." It was not — on either tree
  32. S-32 2026-07-17 — A migration rewrote a broken path into a differently-broken one; Win32_SystemSlot reported all 7 PCIe slots free including the one holding the GPU
  33. S-33 2026-07-17 — SGH fabricated the Au 5d citation 10.1002/adma.201906478 — and five documents then spent four days telling the next session to "replace" a citation that was never inserted
  34. S-34 2026-07-19 — A node's DOI can be REAL while its title is FABRICATED
  35. S-35 2026-07-19 — A sub-agent cannot read its host/parent session's transcript
  36. S-36 2026-07-19 — An OAuth "deleted project" claim was carried forward on faith for days and was false
  37. S-37 2026-07-20 — THE DATED-HANDOFF SCHEME ROTTED FOUR TIMES — the glob returned 15 files across two directories
  38. S-38 2026-07-26 — "I called GEM and Vertex dead. They were not." — the sandbox's own egress block read as dead keys
  39. S-39 2026-07-26 — SANDBOX NONDETERMINISM — 23 scans over threshold / worst 4.998 eV on the first run, 7 / 3.680 on the next four. Same code, same data.
  40. S-40 2026-07-26 — "VERTEX DEAD" was recorded THREE TIMES on the strength of a wrong auth header
  41. S-41 2026-07-26 — The rail check's own first run cried wolf — the probe truncated the model list at 400 bytes
  42. S-42 2026-07-26 — The daily rail-health task was blind for seven days because its connected-folder set is not a file anyone can edit
  43. S-43 2026-07-26 — "GDX blocked, needs re-auth" was too broad — one of three paths was down and I wrote it as all three
  44. S-44 2026-07-26 — Every doc in the tree called GW browser-only. It was one free GET /v1/models away.
  45. S-45 2026-07-27 — A literature sweep run without a verification step had a 100% citation failure rate
  46. S-46 2026-07-27 — Five defects the headless GUI suite caught that no linter could — including 2 eV lost between the screen and the arithmetic
  47. S-47 2026-07-27 — The outside review found real physics bugs while 143 tests and 26 invariants were green — binding_energy() omitted the bias
  48. S-48 2026-07-27 — GEM argued one finding confidently and it was simply wrong — and reported NO FINDINGS on the file holding the real bug
  49. S-49 2026-07-27 — GEM's auth: two of three options dead, and the real blocker was one key in settings.json
  50. S-50 2026-07-27 — Six defects in the fitting layer — an unweighted Poisson fit, a defensible gate that passed redchi = 4,000, and four more
  51. S-51 2026-07-27 — The adversarial re-verification found the worst bug of the session — v_applied or 0.0 reproduced the impossible Φ = 7.12 eV the archive already had a name for
  52. S-52 2026-07-27 — The courier architecture was designed around an absence that nodes/gbw.json had already denied
  53. S-53 2026-07-27 — THE FORGE — 204 node calls produced output that failed 3 of 4 gates, and FORGE.bat double-launched and was still clobbering restored files an hour later
  54. S-54 2026-07-27 — GEM issued a FALSIFIED verdict from a premise it had itself marked [UNVERIFIED], and its predicted mechanism did not reproduce
  55. S-55 2026-07-28 — COWORK FABRICATED A QUOTATION AND ATTRIBUTED IT TO GEM — and manufactured a three-way agreement that never existed. the operator caught it from a single word, two days later.
  56. S-56 2026-07-28 — An MCP server was built that reimplemented 7 of bts_tools.py's 8 verbs — the 07-13 Origin-reader failure repeated exactly, five days after the index was created to prevent it
  57. S-57 2026-07-28 — bts_tools.edit(count=-1) called str.replace(old, new, 0) — a silent no-op that reported success, live since 07-17
  58. S-58 2026-07-28 — bts_tools._guard confinement was a SUBSTRING test, so C:/evil/tmp/x passed
  59. S-59 2026-07-28 — fs_write threw IntegrityError on every write on Windows — text-mode open() translates \n to \r\n
  60. S-60 2026-07-28 — Grok's config.toml was CORRUPTED on the second wiring run — re.sub processes escapes in the replacement
  61. S-61 2026-07-28 — GEM registered the MCP server but refused to START it — security.folderTrust.enabled defaults true and the per-server trust flag does not bypass it
  62. S-62 2026-07-28 — A console in QuickEdit selection mode SUSPENDS the process, and it looks exactly like a hang. The title bar read Select.
  63. S-63 2026-07-28 — CoP was written off on a premise that expired — five days of a dead lane, and the "do not re-test" note was the thing keeping it dead
  64. S-64 2026-07-28 — A launcher that prints "LAUNCHED" is not evidence anything launched — FANOUT_UI.bat v1: neither node ran
  65. S-65 2026-07-28 — setx PATH "%PATH%;…" is destructive — it truncates at 1024 chars, silently
  66. S-66 2026-07-28 — hv_judged was added to to_json and not to from_json — written to disk, quietly dropped on read
  67. S-67 2026-07-28 — The RULING_COLUMNS append-only invariant was written wrong TWICE — both versions forbade the very thing append-only means
  68. S-68 2026-07-28 — GEM reported its prototype "compiles and runs flawlessly (Exit Code: 0)" having executed only py_compile
  69. S-69 2026-07-28 — Bare python on this box is 3.13 with no pytest — and _pyver_probe.log had already recorded it on 07-27
  70. S-70 2026-07-28 — 55 LF-only line endings and 12 non-ASCII bytes in this project's .bat files produced measured mojibake
  71. S-71 2026-07-28 — srdata.nist.gov returns HTTP 200 for EVERY path including nonsense — the 401-vs-404 endpoint test is useless against it
  72. S-72 2026-07-28 — A session ran four streams in one context and degraded badly by the end
  73. S-73 2026-07-29 — the operator's screen was taken ~30 times in two days — every test run a 95-second round trip while he was working
  74. S-74 2026-07-29 — The R2 publish took 16 minutes to transfer ZERO bytes — 87 rclone invocations, 970.3 s
  75. S-75 2026-07-29 — There are TWO 00_WORKING directories and four bats write to the unpublished one — four reports reached nobody
  76. S-76 2026-07-29 — BU_RAILS.MD — a live pointer to a file that was NEVER CREATED, under a stream name the operator had already killed
  77. S-77 2026-07-29 — ~$500 over on usage credits in one month running Fable uncapped — discipline alone did not hold
  78. S-78 2026-07-29 — Cowork hand-coded for a whole session with three paid rails untouched — and it felt like thrift
  79. S-79 2026-07-30 — A pointer relocation took four edits across three files — miss any one and BootUP silently reads stale state
  80. S-80 2026-07-30 — Cowork's own session storage is walled off from Cowork — any plan depending on it reading its own transcript is unbuildable
  81. S-81 2026-07-30 — The LEGAL session produced ~26 analysis documents and SENT ZERO; a request drafted 07-29 was still unfiled 07-30
  82. S-82 2026-07-30 — ~8 confident errors in one domain in one session — including one that INVERTED the case theory
  83. S-83 2026-07-30 — pause is not a report — REPEATED TWICE IN ONE DAY
  84. S-84 2026-07-30 — Two bats failed against a VIRTUALIZED %APPDATA%\Claude — a folder copy returned only Electron GPU cache
  85. S-85 2026-07-30 — COPY Session Logs to Legal.bat dumped 2,735 files / 1.7 GB of every session ever run into the case folder
  86. S-86 2026-07-30 — THE MONOLITH DIAGNOSIS — ROLD is a folder, not a repository. Measured.
  87. S-87 2026-07-30 — A .bat that calls copilot without CALL ENDS ITSELF, silently, mid-probe
  88. S-88 2026-07-30 — The pointer checker printed mojibake into its own log on the first native run
  89. S-89 2026-07-30 — "the operator signed in to CoP on 2026-07-29" was recorded as fact and is FALSE for the CLI
  90. S-90 2026-07-30 — CoPG could not connect for a day because WINDOWS TEXT-MODE STDIO ate the MCP framing
  91. S-91 2026-07-30 — THE AU 5d DOI WAS NEVER FABRICATED. IT IS REAL, AND MISATTRIBUTED. Corrects S-33.
  92. S-92 2026-07-30 — The citation checker reported two REAL DOIs as non-existent — a timeout read as a 404
  93. S-93 2026-07-30 — "We have no claim-support check" was WRONG — one was built, used once, and abandoned
  94. S-94 2026-07-30 — A RENAME IS CHEAP IN THE TREE AND EXPENSIVE IN THE WIRING. Name it right first.
  95. S-95 2026-07-30 — A CHORE WE AUTOMATED, MANAGED AND SCHEDULED AROUND FOR WEEKS DID NOT NEED TO EXIST
  96. S-96 2026-07-31 — The recovery mirror cooked itself under two routers, and the slide took 18 days unnoticed
  97. S-97 2026-07-31 — I CHECKED THE WRONG GOOGLE ACCOUNT AND CALLED A PROJECT NON-EXISTENT
  98. S-98 2026-07-31 — A WARNING IN A DOCSTRING IS NOT A CONTROL — Cowork edited the DERIVED file, hours after writing the rule
  99. S-99 2026-07-31 — THE CORRECTION INHERITED THE DEFECT IT CORRECTED — and a checker scoped to where I was looking
  100. S-100 2026-07-31 — A NODE WAS "LIVE, conf=V" FOR TWO DAYS AND HAD NEVER ONCE BEEN CALLED
  101. S-101 2026-07-31 — A HARDCODED WINDOWS PATH MADE A DIRECTORY NAMED V:\Research4\... INSIDE THE TREE
  102. S-102 2026-07-31 — WE GOVERNED THE WRONG DIRECTION — the risk was never overspend, it was 94.5% EXPIRY
  103. S-103 2026-07-31 — I RAISED THE ALARM THREE TIMES BEFORE RUNNING THE CHEAP CHECK THAT DOWNGRADED IT
  104. S-104 2026-07-31 — "CONTENTS ALREADY RECOVERED AND MERGED" WAS TRUE OF THE TWO FILES SOMEONE LOOKED AT, AND WAS WRITTEN ABOUT THE WHOLE DIRECTORY
  105. S-105 2026-07-31 — THE NIGHTLY RAIL CHECK REPORTED A RENAME'S SUCCESS AS A REGISTRATION FAILURE — FOR TWO DAYS, IN THE FILE WE READ AT BOOTUP
  106. S-106 2026-07-31 — THE PUBLISH WROTE ITS LOG INTO THE TREE IT WAS PUBLISHING, AND ONE FAILED FILE SILENTLY DISABLED THE ENTIRE DELETE PASS
  107. S-107 2026-08-01 — MY OWN DEDUPE TOOL REPORTED 53.9% OF THE CORPUS REDUNDANT. THE TRUE FIGURE WAS 15.8%. IT WAS CALLING SIX DIFFERENT TAX YEARS DUPLICATES OF EACH OTHER.
  108. S-108 2026-08-01 — THE HANDOFF'S OWN CONSTANTS WENT STALE INSIDE THE TIDYUP THAT WROTE THEM
  109. S-109 2026-08-02 — TWENTY-FOUR .md FILES WERE HANDED TO KEITH BEFORE ANYONE ASKED ABOUT FORMAT — AND THE FIX WAS THEN MISAPPLIED THE SAME WAY
  110. S-110 2026-08-02 — A DOCUMENT DESCRIBING AN ACT WAS TWICE TAKEN AS PROOF THE ACT OCCURRED — TWICE IN ONE HOUR, ON THE SAME FACT PATTERN
  111. S-111 2026-08-02 — "NOT IN THE CORPUS" WAS REPORTED AS "NOT IN THE RECORD" — AND THE RETRACTION PROPAGATED INTO FOUR DOCUMENTS BEFORE ANYONE CHECKED
  112. S-112 2026-08-02 — THE BAT'S SAFETY WARNING NAMES A DISK NUMBER. DISK 0 IS NOW THE BOOT DRIVE.
  113. S-113 2026-08-02 — COWORK MISIDENTIFIED THE DEAD DRIVE TWICE IN ONE NIGHT, BOTH TIMES BY REACHING A CONCLUSION AHEAD OF THE IDENTIFIER
  114. S-114 2026-08-02 — THE DOCSTRING DOCUMENTING A QUOTING SCAR CONTAINED A TRIPLE-QUOTE AND BROKE THE FILE THE SAME WAY
  115. S-115 2026-08-02 — BU.MD POINTED AT FOUR DESKTOP BATS — DIRECTLY BENEATH ITS OWN RULE FORBIDDING EXACTLY THAT
  116. S-116 2026-08-03 — AN UNMEASURED CLAIM SURVIVED BECAUSE IT WAS THE CAUTIOUS ONE
  117. S-117 2026-08-03 — THE BUNDLE EXISTED, SO THE GUARD PASSED. THE DOCUMENTS THE QUESTION TURNED ON WERE NOT IN IT.
  118. S-118 2026-08-03 — THE CARRY-OVER SYSTEM WAS DEAD ALL DAY. UNESCAPED QUOTES IN ONE STRING, AND corrections.toml STOPPED PARSING.
  119. S-119 2026-08-03 — THE CANONICAL FILENAME HELD THE OLDEST DRAFT. FOLLOWING THE POINTER CORRECTLY GOT YOU THE WRONG DOCUMENT.
  120. S-120 2026-08-03 — THE PRE-RUN ESTIMATE WAS PRINTED AS THE MEASURED COST, IN THREE DOCUMENTS, WITH THE LEDGER SITTING RIGHT THERE
  121. S-121 2026-08-03 — THE PAID RETURN LANDED AT 13:31 AND NOBODY WAS WATCHING FOR IT. COWORK ALMOST CLOSED THE SESSION ON TOP OF IT.
  122. S-122 2026-08-03 — THE VERIFICATION PASS FALSIFIED ITS OWN CONTROL DOCUMENT AND CERTIFIED IT CLEAN IN THE SAME HOUR
  123. S-123 2026-08-03 — WE TOLD KEITH HE NEVER SIGNED A CONTRACT HE HAD E-SIGNED, AND THE PROOF WAS IN THE OTHER SIDE'S OWN EXHIBIT
  124. S-124 2026-08-03 — "I DON'T MIND SPENDING SOME CREDITS" IS AUTHORISATION TO SPEND, NOT AN INSTRUCTION TO PROCEED NOW
  125. S-125 2026-08-03 — A REVIEWER WAS RIGHT ABOUT ONE THING, SO COWORK STOPPED CHECKING THE NEXT — AND RELAYED A CLAIM THAT CONTRADICTED ITS OWN VERIFIED FINDING
  126. S-126 2026-08-03 — THREE ROUNDS SPENT ARGUING THE END DATE OF A SCHEDULE THAT NEVER RAN
  127. S-127 2026-08-03 — WE TOLD THE CLIENT A FACT ABOUT HIS OWN CONDUCT AND NEVER ASKED HIM
  128. S-128 2026-08-08 — THE SESSION FOUND THE WRONG OPPOSING COUNSEL, SAID SO AT LENGTH, AND EDITED NOTHING
  129. S-129 2026-08-08 — THE INDEX SAID 71 AND THE DIRECTORY HELD 74 — AND THE THREE MISSING ONES WERE THE EXHIBITS THE QUESTION TURNED ON
  130. S-130 2026-08-08 — FIVE OF SEVEN CITATIONS FROM ONE RAIL WERE FABRICATED — AND BOTH LOAD-BEARING ONES WERE INVENTED
  131. S-131 2026-08-11 — THE FILE NAMED RECORD_ALL.txt WAS record_core. TWO PAID BENCH OPINIONS RULED ON 51 OF 74 DOCUMENTS.
  132. S-132 2026-08-11 — THE KEY WAS WRITTEN, IT WAS VALID TOML, AND IT LANDED IN THE WRONG TABLE. doctor SAID "parse ok" ON THE SAME SCREEN AS "model default."
  133. S-133 2026-08-11 — THE FIX FOR S-132 WAS WRITTEN, REGISTERED, DESCRIBED AT LENGTH, AND NEVER APPLIED — AND THE HANDOFF RECORDED IT AS "VERIFIED BY RE-READ."
  134. S-134 2026-08-11 — THE MONITORS WORKED WHENEVER ANYONE WAS WATCHING THEM. FOUR TOOLS DIED ON AN EMOJI, AND TWO OF THEM EXITED 0 WHILE DOING IT.
  135. S-135 2026-08-11 — A CORRECTLY PLACED KEY WITH AN INVALID VALUE IS STILL BROKEN, AND NO STRUCTURAL CHECK CAN SEE IT
  136. S-136 2026-08-11 — THE RUNNER'S OWN SELFTEST REFUSED TO CERTIFY IT, AND THAT REFUSAL IS THE ONLY REASON A BROKEN MONITOR WAS NOT SCHEDULED INTO SERVICE
  137. S-137 2026-08-11 — gdx_fresh_auth.py DISPOSED OF THE WORKING CREDENTIAL BEFORE PROVING IT COULD CREATE A REPLACEMENT
  138. S-138 2026-08-12 — FOUR CHECKS CRIED WOLF ON HEALTHY FILES IN A SINGLE SESSION, AND EVERY ONE LOOKED LIKE A REAL DEFECT
  139. S-139 2026-08-12 — A TEST FIXTURE WITH AN ABSOLUTE DATE DOES NOT DECAY INTO UNCERTAINTY — IT DECAYS INTO ASSERTING THE OPPOSITE
  140. S-140 2026-08-12 — "GREEN ×4 — THE FIRST REAL DRIVE READING THIS MESH HAS EVER HAD" WAS GREEN ON Nones
  141. S-141 2026-08-12 — A DUPLICATE KEY TOOK rails.toml OFF THE AIR, SELF-INFLICTED WHILE FIXING THAT SAME BLOCK
  142. S-142 2026-08-12 — THE BACKGROUND LANE STALLED REPEATEDLY AND ITS OWN LEDGER WAS STRUCTURALLY INCAPABLE OF SAYING SO
  143. S-143 2026-08-12 — THE RECORD ASSERTED A FALSE ARCHITECTURE, PLANNING RAN ON IT FOR DAYS, AND ONLY THE HUMAN WHO BUILT IT COULD SEE THE ERROR
  144. S-144 2026-08-12 — A DEFERRAL JUSTIFIED BY A SURPLUS INHERITS THAT SURPLUS'S EXPIRY DATE
  145. S-145 2026-08-13 — S-134 INVERTED: THE TOOL WORKED IN THE HARNESS AND DIED IN FRONT OF THE HUMAN — AND IT DIED ON THE LINE NAMING THE CAUSE
  146. S-146 2026-08-13 — A MISREAD ACRONYM BECAME A DIAGNOSIS, REACHED FIVE ARTIFACTS, AND DEFENDED ITSELF AGAINST THE FIRST CORRECTION
  147. S-147 2026-08-15 — THE NEGATIVE CONTROL SHARED THE POSITIVE CHECK'S DEFECT, SO IT CERTIFIED NOTHING — AND IT PRINTED A ✅ SAYING SO
  148. S-148 2026-08-16 — A hard-coded Windows path became an ordinary filename on Linux, so two writers held different locks and both reported success
  149. S-149 2026-08-16 — The mailbox was correct syntax on the wrong surface, so a dead channel was indistinguishable from a quiet one
  150. S-150 2026-08-18 — _session_logs is not the complete record, and three agents independently certified the same false negative
  151. S-151 2026-08-18 — A time-boxed search reported absence when it merely ran out of clock
  152. S-152 2026-08-18 — A directory name was load-bearing documentation, and flattening it caused a wiring error weeks later
  153. S-153 2026-08-18 — Every gate passed green while the registered server did not exist
  154. S-154 2026-08-18 — A control document carried a standing DELETE order against what became the recovery source
  155. S-155 2026-08-25 — CLASSES DISCOVERED (TidyUP2 check 10) — two evidence-handling classes from the legal stream
  156. S-156 2026-09-23 — OpenRouter HERO 429 was the upstream free pool, not our weekly cap — and several "fails" were other classes

More burns, outside that list

43 more misses were measured and kept: the shop floor, the calls, the paths, and the setup gates. With the numbered record, that is 199.

Occupancy

  1. Two heads wrote the same tree
  2. Two pens after the lease had moved
  3. A zombie runner kept spawning
  4. A context string was treated as a list
  5. Failed work sat on the board as if it were live
  6. A judge ran on an empty cache
  7. An as-is retry re-fired the same empty job
  8. A squash deleted more than it added
  9. A Windows command was pasted into the wrong shell
  10. Compliance was announced without a readback
  11. A chamber posted a truncated diff twice
  12. A free pin answered with an essay instead of the token
  13. A patch missed the file that was actually on disk
  14. A missing path was patched instead of created
  15. A symlink was filed as code
  16. The checker blamed the coder for bugs already in the tree
  17. Tests failed because the runner left the assets behind
  18. A grade was filed before three copies existed

Calls

  1. A stale flag skipped the local config and failed auth
  2. A chat door returned success with no served model
  3. A spawn in an empty house stopped early
  4. An open stdin hung the runner
  5. A second launcher hung on a long prompt
  6. The local log named the provider, not the model
  7. The mission was pasted instead of laid on disk
  8. Traces were written where anyone with the name could read them

Paths

  1. Dot-dot traversal
  2. Drive-root relative path
  3. Drive-relative path with no slash
  4. Encoded dot-dot
  5. Null byte in a path
  6. Current-drive Windows root
  7. Alternate data stream
  8. Windows short name
  9. WSL drive translation
  10. Windows device namespace
  11. Batch payload write
  12. Foreign shell with a relative escape

Setup gates

  1. The API was not enabled
  2. Billing was not linked to this project
  3. One key cannot serve two APIs
  4. Project id and project number were treated as two projects
  5. A sandbox block was read as a dead key

Each day · Three longer notes