Work log

Append-only running record of what was done, in order. Its job is to let a session that has lost its conversation history pick the work back up without guessing.

Read this file, and DECISIONS.md, at the start of any working session. Append to it at the end of every completed unit of work - not at the end of a session, which may not arrive cleanly.

Entry format: what changed, what it unblocks, what is next, and anything surprising found on the way. Surprises are the most valuable field; they are what a fresh context cannot re-derive. This table is generated. Edit an item under worklog/, then run tools/split-doc.sh --index orbistoun WORKLOG 2 worklog.

date entry
2026-08-19 Scaffold
2026-08-19 Terminology and documentation pass
2026-08-19 Concept intake and roadmap resequence
2026-08-19 Test corpus and the accuracy suite
2026-08-19 Observability and per-title overrides
2026-08-19 Real material arrived; plan revised against it
2026-08-19 Crunch: phase 0c, 0e, and most of phase 1
2026-08-19 Crunch continued: phases 2, 0e, and 3
2026-08-19 Crunch continued: worker mode and image placement
- Open at end of session
2026-08-19 Guest code executes
2026-08-19 (later) - The guest runs, and says what it wants
2026-08-19 (later still) - The name search
2026-08-19 (evening) - Names, and proving they are ours
2026-08-19 (late) - The loop, made runnable
2026-08-19 (night) - Making the loop canonical
2026-08-19 (very late) - The first implemented function, and what it taught
2026-08-19 GPU instrumentation: counting before translating
2026-08-19 The encoding table stops being a guess
2026-08-19 Operands, and a surface to see them through
2026-08-19 Operand layouts, and the limit of the per-family model
2026-08-19 Per-opcode operands, solved rather than written
2026-08-19 The execution model, decided and stubbed
2026-08-19 Translated shaders can now be run, not just validated
2026-08-19 The first guest instruction, translated and executed
2026-08-19 The worklist starts moving
2026-08-20 Two wavefront models, agreeing
2026-08-20 Factoring, floats, and the solver earning its refusals
2026-08-20 Flat memory layouts, and the same lesson learned twice
- Guest memory, and two driver faults that a builder could have caught
- The worklist blockers, and a solver that was quietly wrong about three of them
- The last three encoding families, and one of them was wrong
- The execution mask becomes real
- Comparisons, and a conditional shader with no branch in it
- Lanes that do different things
- Control flow: a switch inside a loop
- The seam: a submitted command buffer reaches the GPU
- The condition code, and the first compiled shader to translate
- Vector arithmetic, and the bits nothing was looking at
- The condition code, everywhere it was missing
2026-08-20 Attribution, the submit function, and honest word lists
- The second sub-encoding, and a check that was lying
- Two routes to a shader, because the loader thread named the entry points
- Reconciling with the loader thread
2026-08-20 (later) - A home for what we learn
- G9: the capture format, and a guard that was right
- G13 measured rather than assumed, and the work that does have an oracle
- The local data share, and an operand that was never in any sample
- Using the oracle properly
- The generator found a real bug, which is what it was for
- The target generation was never checked
- Named instructions, because there are two targets
2026-08-20 (evening) - The wall had a name, and it was C++
- The dispatch speaks names
- The retarget, and the tree is red on purpose
- The retarget landed
- Threads are real, and the entry point was reading garbage
- The crash was the harness rebuilding Vulkan ninety-six times
- The retarget was not finished, and cargo had been hiding it
- The entry image is built, and measured out of the running
- The division sequence, two thirds of it
- The tail found the wall, and the wall moved once
- A debugger, of the cheap kind - and it named the parked bug in one run
- The number nobody had recorded: every guest call was misaligned
- Float controls, then narrow wavefronts
- Starting the GUI found the leak before writing a line of GUI
- The subgroup level runs
- Phase 2b: a window, and it launches guests
- The last unblocked instruction, and the buffer descriptor's operands
- The shell grew a menu, a toolbar, and an argument about dead controls
- Untyped buffer access
- The worklist ranks by reachability first
- What downloadable homebrew can and cannot do for us
- Library rows were being rebuilt sixty times a second
- The library folder is a setting now, and refresh is a button
- A filesystem, and the guest read every file it asked for
- Review queue, and the shader loop gets a verdict
- The bisection lever had never been connected
- The two rungs
- One name out of a quarter million, and sixteen useful refusals
- The name sweep could not have found the answer
- The decision log had been quietly losing its own references
- Video-out handles, two more names, and a wall that will not move
- The sub-encoding list was never needed
- The oracle for hidden side effects was in the fixtures already
- Notes from the obSCEne thread: one declined, one implemented
- A recorded gap was recorded wrongly
- The dispatch loop was not where the cost was
- The log line became a check
- A stack trace, at last
- Call sites, aliasing, and a wall that has taken eight attempts
- Three assumptions asked a question they had only ever answered themselves
- The filesystem is cleared, and there are four walls
- The most common instruction in the set had no operands, and every test was green
- The split opcode, and a test that asked to be deleted
- Two titles that never parsed, and an abort reported as an illegal instruction
- Typed buffers: operands solved, formats measured
- The decision-number ceiling drops to thirteen
- A documentation sweep, because six units of change left claims behind
- The run now says what to do about itself
- Typed buffer translation, and the last thing in this lane that needed nobody
- The decode side is finished
2026-08-21 Behavioural provenance (D180)
2026-08-21 Run conditions and the discount on the headline (D181)
- A provenance rule that was stated too narrowly, and a refusal defended for the wrong reason
2026-08-21 The compatibility record (D182)
- The third-source idea had no consumer, and I had already written that it did
2026-08-21 snprintf_s, and what "implemented" does not mean (D183)
2026-08-21 The abort-at-53 is one bug, in two Unity titles
- The interop contract, pinned before either side has code
- The probe-record reader, built against transcripts and no hardware
2026-08-21 The abort-at-53, diagnosed (D186, D187)
2026-08-21 The database was there all along (D188)
2026-08-21 A shape, not a word list (D189), and an audit that was already red
- Provenance across the boundary, and the demotion that has to stay visible
2026-08-21 memalign, and one allocation path (D190)
2026-08-21 The nine-attempt wall: a clean negative, after three false starts
- The corpus is records all the way down, and the reader was looking in the wrong place
- From "a check passed" to "this function returns this, and here is how well we know it"
- A record becomes a knowledge entry, graded, or it does not become one at all
- Symbols are graded differently from values, and four record kinds were left unparsed
- Per-area coverage, and what a skip is not
- A probe cannot certify its own machine, and the grading was resting on it
- orbistoun drives the session now, and D207 said it never would
2026-08-22 FreeBSD harvest: the walker ignored its own rule (D191)
- call and read are live, and the fixture for one of them is broken
- The streamed report needed no change, and now there is a test saying so
- The pinned defect fired, and the replacement pins the shape instead
- Three ways of not knowing, kept as three
- A shared bridge file, and the open-enum rule caught a live bug here
2026-08-22 Closing the automation gaps (D193, D194, D195)
2026-08-24 PROJECT_STATUS.md rewritten, because it opened with a falsehood
- A claim made to the other thread, then tested
2026-08-24 The unknowns become a queue (D196)
2026-08-24 The loop, written down; and a documentation audit (D197, D198, D199)
2026-08-24 The repository's own layout, audited (D208)
2026-08-24 orbistoun-gen: the table generators, and the seam that makes them checkable (D209)
2026-08-24 The obSCEne bridge caught a live memory bug (D210)
2026-08-24 What observed actually meant, and the ceiling that was not one (D213)
2026-08-24 A language-model service, isolated from everything
2026-08-24 The first proposer, and the optimisation that had to be reverted
2026-08-24 The toolbar captures the window, and recording is greyed out (D215)
2026-08-24 The wall was readable all along (D217)
2026-08-24 Four eliminations, no confirmations (D218)
2026-08-24 A diagnostics toolkit, and the registry that should have come first (D220, D221)
2026-08-24 A build says which build it is (D222)
2026-08-25 The wall moved, and the address had been right all along (D223, D224)
- The console gets a console
- The rule became code, and a convenience script nearly corrupted the fixtures
2026-08-25 two eliminations that were never measured
- Every diagnostic axis, and a dispatcher that turns the loop
- Two files that disagreed, and the last unexplained sweep result
- The last channel into the wall, and two words that paid for themselves
- Pointing the model at the right question
- Seven names, two words, and a stale binary
- Asking whether the next name needs a word or a shape
- Three measured shapes, and a delimiter that could not spell a library
- The model gets a door of its own
- Watchpoints, and what the first one found
- A library that was written, declared, tested, and never registered
- The sweep gains a second dimension, and a wall opens
- The loop turns itself, and clears the wall doing it
- Promotion: what a turn measured, and what it did not
- The merge rule leaves the shim
- Every floating-point function worked, and every one was reported missing
- The dispatcher becomes its own crate, and the loop becomes a command
- Teach a man to fish: the stub policy learns to write
- Tier one: the loop writes its own patch
- The learned file becomes something you can send somebody
- The loop finds a gap nobody had looked at
- Two hypotheses instead of one, and the oracle that could not tell them apart
- The oracle arrives, and immediately corroborates
- Five hours to fourteen seconds
2026-08-26 The loader learned to read the other half of the world
- The vocabulary regrew, and the only alarm was a clock
- The shell button had nowhere to go
- One word for the shell, and the window had already taken it
2026-08-27 Asked the guest what it wanted instead of guessing
2026-08-27 Data imports finally get something they can dereference
2026-08-27 Two backlog items, and a method that was quietly unsound
2026-08-27 A model that needs no key, and the parts not worth copying
2026-08-27 The benchmark disagreed with me twice before it was right
- A controller subsystem, and a name I talked myself out of
2026-08-27 The payloads started working
2026-08-27 Asked to drop an engine, and the engine was fine
- Two modelled states made real, after being told they were only modelled
2026-08-27 The tags outlived the instruction that removed the thinking
- User support, and the third thing found modelled and inert
2026-08-27 klogsrv prints its banner
- The POSIX names were unserved, and they were never missing behaviour
2026-08-27 The guest names its own requirement
2026-08-27 The BSD harvest was names-only, and the work outgrew it
- Coverage crunch: libc, elf, hle
2026-08-27 The harvest became a command, and writing it twice found a bug
- Coverage crunch: kernel and thunk
2026-08-27 Proof of sourcing for the harvested constants
2026-08-29 The null jump is not the sysctl refusal
2026-08-29 The null jump was the shortcut coming due
2026-08-29 Markers in .bss, and the SDK's logging helper
2026-08-29 Gate green
2026-08-29 The payload wall, answered by asking the guest
2026-08-29 Sockets, and the file calls that go with them
2026-08-29 klogsrv imports nothing that is not implemented
2026-08-29 What about klogsrv
2026-08-29 The syscall boundary, built and not yet reached
2026-08-29 A setting consulted nowhere, for the fourth time
2026-08-29 A fault report that names our own code
2026-08-29 klogsrv opens a port and something connects to it
2026-08-29 ftpsrv talks, and asks to be root
2026-08-29 A one-line disagreement about zero
2026-08-30 An FTP server, and the six things in front of it
2026-08-30 The listing
2026-08-30 The download, and three blind instruments
2026-08-30 When, not which
2026-08-30 Ask the guest, one field at a time
2026-08-30 What real hardware sent back
2026-08-30 The probe answers a question about itself
2026-08-30 Seven placeholders, retired by one file
2026-08-30 The constant was a handle, and the wall was an argument
2026-08-30 A landing zone, and the first syscalls a payload has ever made here
2026-08-30 The work list could not see a system call
2026-08-30 The hardware had already answered; the results were on disk
2026-08-30 Word zero is getpid, and the payload's own arithmetic lands on real functions
2026-08-31 Four plans crunched: testable layout, profiles, port reporting, vaddr provenance
2026-08-31 obSCEne's hardware run absorbed: five vaddrs confirmed, and a payload retry
2026-08-31 Disassembled klog.elf around image+0x2708: kernel_copyout, setsockopt, and high-half kpipe_addr (D411)
2026-08-31 Implemented Syscall 477 (mmap) and wired to the syscall table (D412)
2026-08-31 (later) - payload retry after further changes: 7 -> 32 syscalls, into klog's socket loop
2026-08-31 Emulated kernel escape R/W pipe for dynamic symbol resolution (D413)
2026-08-31 (later) - orbistoun runs the whole obSCEne suite; obSCEne becomes its conformance oracle
2026-08-31 (later) - the test corpus becomes a verb (D414)
2026-08-31 (later) - COMPATIBILITY.md, generated from the records (D415)
2026-08-31 (later) - four HLE fixes from the hardware diff (D416)
2026-08-31 (later) - three more HLE fixes: thread join, mutex type, audio (D417)
2026-08-31 (later) - the census-control leak, root-caused and fixed (D418)
2026-08-31 SwVersion write refined from the hardware dump
2026-08-31 (later) - second payload run absorbed: five vaddrs confirmed (D419)
2026-08-31 (later) - the "surely there's more" pass found a wrong firmware value (D420)
2026-08-31 (later) - sceKernelGetSystemSwVersion made a profile setting (D421)
2026-08-31 (later) - per-title device sandbox, from the overlay already there (D422)
2026-08-31 (later) - C succeeded: the full current obSCEne runs as the payload
2026-08-31 (later) - video flip model: the first complete obSCEne run in orbistoun (D424)
2026-08-31 (later) - first fidelity fix off the complete run: GetModuleInfo refusal code
2026-08-31 (later) - resolution-status: implemented, and the "headless" skip debunked
2026-09-01 systematic hardware diff: video error family + single-output-ownership (D426)
2026-09-01 started the GPU subsystem: PM4 command builders (D427)
2026-09-01 real-title loop: libSceSysmodule (the load call nearly every title makes)
2026-09-01 reentrant guest execution + std::call_once (D430)
2026-09-01 sceKernelVirtualQuery + the C-runtime threading family (D431)
2026-09-01 Fault reports carry the faulting instruction (D432)
2026-09-01 PPSA28061's wall diagnosed: no guest thread pointer (D432)
2026-09-01 (/loop) - TLS block built; Windows fs-base limitation found; knowledge debt cleared
2026-09-01 (/loop) - Windows fs-base backstop; PPSA28061 FURTHER to the JSON/ULT wall
2026-09-01 (/loop) - libSceUlt mutex/condvar/ulthread; PPSA28061 wall is now online (D435)
2026-09-01 (/loop) - names search: PPSA28061's online import is un-nameable locally
2026-09-01 (/loop) - remaining walls characterised; all need obscene/external data (D436)
2026-09-01 (/loop) - obSCEne hardware memory data mined; map now starts at 0x10000 (D437)
2026-09-01 (/loop) - error-code sweep completed: 5/6 rejects-* checks match hardware (D438 cont.)
2026-09-01 (/loop) - file error codes now the measured errnos (D439 cont.)
2026-09-01 (/loop) - GNM dispatch header now byte-matches hardware (D439 cont.)
2026-09-01 (/loop, user task) - orbistoun-fs test isolation + clippy debt fixed
2026-09-01 Shared-allocator diagnosis: wrong titles, real gap is Configured (D442)
2026-09-01 proc-param reader built; the mem-param oracle falsified the budget premise
2026-09-01 PPSA02664 goes FURTHER (233->1541 calls): policy region was on the guest's heap
2026-09-01 obSCEne is the oracle; flexible-memory fixed byte-exact against hardware
2026-09-01 stack read-ahead guard; obSCEne runs its whole suite under orbistoun
2026-09-01 sceKernelVirtualQuery sees the image and stack (obSCEne vq-text/vq-stack pass)
2026-09-01 sysctlbyname implemented; kern.osrelease answers (obSCEne osrelease passes)
2026-09-01 obSCEne oracle mined out for clean fixes; pivot to a real title next
2026-09-01 found PPSA02664 regression (1541->234); mechanism understood, fix deferred (net-zero turn)
2026-09-01 (/loop) PPSA02664's fault site is non-deterministic; the "regression" was a thread race
2026-09-01 (/loop) PPSA21564 boots: the sceLibcMspace allocator family + bcmp
2026-09-01 (/loop) scePthreadGetthreadid; PPSA21564 now runs at 0% stubs
2026-09-01 (/loop) POSIX TLS keys unblock Unity's TBB scheduler; PPSA21564 stops aborting/racing
2026-09-01 (/loop) time + fgets: PPSA21564 (Astro's Playroom) reaches main(), prints, parses args
2026-09-01 (/loop) POSIX unnamed semaphores; the Cond.cpp wall is not a traced HLE call
2026-09-01 (/loop) Fault reports name privileged instructions + distinguish emulator bugs
2026-09-01 (/loop) Fault reports name the null base register automatically
2026-09-01 (/loop) Execute breakpoints: capture a guest value where it is used
2026-09-01 (/loop) The call trace records what each call answered
2026-09-01 (/loop) Map direct memory into an existing reservation; crack image+0xafcc08
2026-09-01 (/loop) The reserve-then-map fix advanced three titles, not one
2026-09-01 (/loop) Overnight crunch: putchar and std::random_device
2026-09-01 (/loop) PPSA04263's map wall: a rigorous narrowing, not yet a fix
2026-09-01 (/loop) Made reservation failures legible, then cracked the collision they hid
2026-09-02 (/loop) strcpy_s implemented; new-thread TLS gap diagnosed (the real wall)
2026-09-02 (/loop) Per-thread TLS for spawned guest threads: PPSA04263 10k -> 333k calls
2026-09-02 (/loop) Blocking WaitEventFlag (kills a 304k spin) + atan2f/sincosf
2026-09-02 (/loop) PPSA25872's unnamed spin (naming-gated) + pthread attr setters
2026-09-02 (/loop) The last clearly-standard stubs: vsprintf_s, SetVirtualRangeName
2026-09-02 (/loop) The last clean functions; the oracle-free crunch is complete
2026-09-02 (/loop) Packed typed-buffer formats begin: UINT single-word load, GPU-verified
2026-09-02 (/loop) Packed typed-buffer SINT load: sign extension, GPU-verified
2026-09-02 (/loop) Packed typed-buffer UNORM load: the first converting kind, GPU-verified
2026-09-02 (/loop) Packed typed-buffer SNORM load: the clamp, GPU-verified
2026-09-02 (/loop) Packed typed-buffer FLOAT16 load: driver-widened halves, GPU-verified
2026-09-02 (/loop) The ctype tables, measured off hardware: table right, wall unchanged
2026-09-02 (/loop) The _Getpctype wall is a mislabel: measured, and it reframes four decisions
2026-09-02 (/loop) Every report today was a file from 01:50; the ctype work had moved the wall
2026-09-02 (/loop) The reporter could not survive an execute fault; fixed, and the ctype work turns out to be a 7x advance
2026-09-02 (/loop) 452 documented functions were knowable all along; the work list is now inventory-driven
2026-09-02 (/loop) Bulk port batch 1: the bounded string and memory functions
2026-09-02 (/loop) Bulk port batch 2: the runtime's out-of-line atomics, conversions and assert
2026-09-02 (/loop) Bulk port batch 3: the standard streams, real recursive locks, and the C++ runtime
2026-09-02 (/loop) Bulk port batch 4: 20 POSIX delegations, and the arity check that refused two
2026-09-02 (/loop) Bulk port batch 5: the single-precision math family and byte order
2026-09-02 (/loop) A partial marker, and the inconsistency the audit found
2026-09-02 (/loop) Bulk port batch 7: scatter/gather, sync, and the ones refused instead
2026-09-02 (/loop) Bulk port batch 8: the posix_-prefixed family, 69 in one go
2026-09-02 (/loop) Bulk port batch 9: the attribute accessors, and three guards that fired
2026-09-02 (/loop) Bulk port batch 10: getsockopt, and what "missing" actually means
2026-09-02 (/loop) Bulk port batch 11: two inits resolved, and the cheap wins are exhausted
2026-09-02 (/loop) Bulk port batch 12: the lock-attribute families - and my stop recommendation was wrong
2026-09-02 (/loop) Bulk port batch 13: the timed condition wait and pthread_once
2026-09-02 (/loop) Bulk port batch 14: the timed acquisitions, and a guard that checked the wrong table
2026-09-02 (/loop) R0: the red gates cleared, and the knowledge file catches up
2026-09-02 (/loop) R1: the measure count is not a bug, and the name oracle already answered
2026-09-02 (/loop) R5: the hardware records finally get read, and the two runs disagree
2026-09-02 (/loop) R6: measurements become a work queue with a completion condition
2026-09-02 (/loop) R7: a live oracle at last, and it found four bugs in twenty-four cases
2026-09-02 (/loop) More differential cases, and the parallel run finds a real concurrency bug
2026-09-02 (/loop) The differential calls back into guest code, and qsort agrees
2026-09-02 (/loop) The sign-extension divergence was mine, not the console's
2026-09-02 (/loop) strtok needs a sequence, and the mutex type mapping becomes a test
2026-09-02 (/loop) Four outstanding items that were never hard, and the overlapping moves
2026-09-02 (/loop) The loader's answers were already right; the real gap is exports
2026-09-02 (/loop) Exports exist now, and the title's own module answers the wall
2026-09-02 (/loop) Three firmware directories, and a test that cannot tell them apart
2026-09-03 (/loop) A title's module is found by name, and fakelib/ proved the rule right
2026-09-03 (/loop) The wall resolves to a real address, and libc turns out to be the title's
2026-09-03 (/loop) The stub tables are global, so there is one of them (D484)
2026-09-03 (/loop) R9 was already closed, and twenty-one new differential cases
2026-09-03 (/loop) A third capture added 186 measurements and took two claims away
2026-09-03 (/loop) One stub table with a range per module
2026-09-03 (/loop) The float environment, and the bit that is not configuration
2026-09-03 (/loop) The title's own modules relocate
2026-09-03 (/loop) The worker links the title, and the verdict turns out to be noisy
2026-09-03 (/loop) The decision log had the answer
2026-09-03 (/loop) The guest runs its own code
2026-09-03 (/loop) The trace was naming the wrong functions
2026-09-03 (/loop) The instruction bytes answered it, and the tags refused the answer
2026-09-03 (/loop) A nid verb, and a hypothesis that did not survive it
2026-09-03 (/loop) .bss, and a blind spot of our own making
2026-09-03 (/loop) The instrument existed, and I mis-grepped it
2026-09-03 (/loop) The guest says which modules to start, and we do not start them
2026-09-03 (/loop) Six absences, and the tag that made them mean something
2026-09-03 (/loop) 24 claims, 24 non-claims, and a restore that broke a mutex
2026-09-03 (/loop) a conformance bug, a stale work item, and a lying instrument
2026-09-03 (/loop) The oscillation is one allocation round, and the backlog is now the whole list
2026-09-03 (/loop) What the 96% does not say, a stale red phase, and three more libc functions
2026-09-03 (/loop) The citations were already here, and R9 was already closed
2026-09-03 (/loop) The queue had permanent residents, and two asks were about the run
2026-09-03 (/loop) Skeletons for the gaps: 61 graphics names and 49 recorded imports
2026-09-03 (/loop) Twenty-eight libraries, and a coverage figure that got worse on purpose
2026-09-03 (/loop) The encoder measurements say the opposite, and thirty names go back
2026-09-03 (/loop) A second capture settles the third query field
2026-09-03 (/loop) A masked register, and the width where a shim shows itself
2026-09-03 (/loop) Eight bytes where documentation said four
2026-09-03 (/loop) A closed gap that both surfaces still called open
2026-09-03 (/loop) Two formatter bugs, in the padding nobody had tested
2026-09-03 (/loop) vsnprintf closes the list, and the break names the cases carrying it
2026-09-03 (/loop) The oscillation is settled: it was the guest's own allocator
2026-09-03 (/loop) The wall named: a module that was loaded, placed, and never started
2026-09-03 (/loop) The wall is down: the guest reached its frame loop
2026-09-03 (/loop) Nothing is ever pending: the guest left its frame loop
2026-09-03 (/loop) dlsym never knew the guest's own exports; the wall did not move
2026-09-03 (/loop) One writer, thirty-seven readers, and a guard that never passes
2026-09-03 (/loop) I blamed a guard that never ran
2026-09-03 (/loop) Starting every module early is not the missing ordering
2026-09-03 (/loop) No branch avoids the read: half the open question is closed
2026-09-03 (/loop) A fault now says what its registers point at, and the label is "None"
2026-09-03 (/loop) One thread call implemented, one deliberately not
2026-09-03 (/loop) The event queue exists so a handle means something
2026-09-03 (/loop) The vendor stat is not the POSIX one under another name
2026-09-03 (/loop) A positioned read took the guest into the GPU
2026-09-03 (/loop) The wall is sceAgcCreateShader, and a cap was hiding it
2026-09-03 (/loop) Two walls that need a measurement, routed to the mechanism that asks
2026-09-04 (/loop) The record format already carried what it was said to lack
2026-09-04 (/loop) The wide family verified, and the case that bites
2026-09-04 (/loop) The interleaved sequence, and the sixteen cases that were blind
2026-09-04 (/loop) strftime, and the question that found it
2026-09-04 (/loop) libm splits in two, and only half belongs here
2026-09-04 (/loop) Two rounding rules that pin each other
2026-09-04 (/loop) strdup covered, and a missing terminator is only caught by luck
2026-09-04 (/loop) One clock, two names, and one origin
2026-09-04 (/loop) The reasoning was in the code, and the record said nothing
2026-09-04 (/loop) Two thirds of the ask list is forty sentences
2026-09-04 (/loop) One question, written a hundred and forty-nine ways
2026-09-04 (/loop) The ask list was asking for a function that does not exist
2026-09-04 (/loop) A measurement was sitting in the list of things nobody knows
2026-09-04 (/loop) Six records said nothing was known while holding a measurement
2026-09-04 (/loop) dlsym succeeds where the console refuses
2026-09-04 (/loop) Forty-seven measured values that no test looked at
2026-09-04 (/loop) The relations are testable where the numbers are not
2026-09-04 (/loop) A permanent resident of the work queue
2026-09-04 (/loop) The relation needed a mount, not a guest
2026-09-04 (/loop) The wall is an experiment, and there is no guest to run
2026-09-04 (/loop) The oracle the roadmap said did not exist
2026-09-04 (/loop) The oracle draws
2026-09-04 (/loop) The export was already decoded, and its blocker had half expired
2026-09-04 (/loop) The fragment path does not need the feature
2026-09-04 (/loop) The translator draws
2026-09-04 (/loop) The oracle carries a varying
2026-09-04 (/loop) The wall was real, the record was twelve days stale, VINTRP translates
- 405. The out-parameter is the whole of the wall
- 406. A rung for the first frame, and a prop that was never counted
- 407. The Agc probe set, aimed from the guest's own arguments
- 408. The flip completion had no reader
- 409. The pthread family, and a prediction kept
- 410. The tail, and what was decided not to do
- 411. The record could not see the work
- 412. The whole corpus was twelve days stale
- 413. libSceUlt, and the quiet four gigabytes
- 414. The first command packets came back
- 415. The constructor was the wrong question
- 416. Seventy-eight percent of every call
- 417. A placeholder that says who
- 418. The classification failed, and found two bugs on the way
- 419. Six arguments, and a hypothesis that was wrong
- 420. The ring is circular, and the trade was false
- 421. The futex had a name, and the wait now waits
- 422. The stack the collector scans
- 423. The fault message was guessing, and it cost a finding
- 424. The protection call learns about the other half of the map
- 425. What the guest opened, and the two devices it wanted
- 426. The wall was the crash reporter, and the file path it died on has a name now
- 427. Reading the command buffer, and the two tools that had to be fixed to do it
- 428. The determinism bug had two halves and a third that is architecture
- 429. Autodebugging starts by not believing the run
- 430. The hardware answered, and the loop reads structures by itself
- 431. The path was globalgamemanagers, and yesterday's finding was wrong
- 432. The bytes were not what it wanted
- 433. The guest was telling us all along
- 434. The answer was in the title directory
- 435. The library was guest code all along
- 436. Stepping through the construction
- 437. Fifty-eight, not three
- 438. Four hundred and two bytes
- 439. Opened, never touched
- 440. Untouched, not zero
- 441. The other title
- 442. A different program
- 443. Re-derived
- 444. One branch
- 445. The head of the sequence
- 446. A conflict nobody stepped past
- 447. The largest record nobody read
- 448. Five files of twenty-eight
- 449. The probe runs here too
- 450. The probe was not calling
- 451. The encoding fell out of a pairing
- 452. The generous window was zero
- 453. The guest was carrying the answer
- 454. The handle was never consulted
- 455. The biggest number on the board
- 456. Advice that cannot succeed
- 457. One function, two answers
- 458. The ratchet cannot say worse
- 459. The branch that printed nothing
- 460. A record older than the repository
- 461. A hypothesis tested and rejected
- 462. An answer needs somebody who asked
- 463. The eboot was never stuck
- 464. Twenty blank pages
- 465. Every import accounted for
- 466. Four numbers
- 467. Two tables that answer each other
- 468. A third title at the same wall
- 469. The binding reached the relocation
- 470. The guest signals itself
- 471. The report can see a guest sitting still
- 472. The canary had a name
- 473. Every request came back at once
- 474. The target is not the caller
- 475. The report learned to name threads
- 476. The handler ran
- 477. Two readers, one corpus
- 478. The differential was measuring itself
- 479. The symbol level
- 480. FURTHER
- 481. Mirrors, and a new wall
- 482. The guest was talking all along
- 483. The guest is the only oracle left
- 484. The corpus has names
- 485. Three roots
- 486. The differential was comparing two sandboxes
- 487. Prospero, Trinity, Orbis, Neo
- 488. An origin list
- 489. The index was never mine to edit
- 490. Orbis, not neo
- 491. The vendor spelling of a socket, and the flag that hung the guest
- 492. orbistoun was calling itself a payload
- 493. The refusal a guest could not see
- 494. The pad structure had been measured
- 495. The audio drain was arithmetic
- 496. A weak symbol orbistoun says exists
- 497. The comparison ran on the wrong machine
- 498. Weak undefined symbols bind to zero
- 499. The corpus is GPU-bound, and a real header agreed