GLaDOS 1.2.29

A correctness release. One real bug, one new diagnostic, two report-only verbs.

Fixed: the adapter could not learn, and never said so

Dora::new started both low-rank factors at zero. Two zero matrices
multiply to a zero product and, worse, to a zero gradient — so the low-rank
branch of every QDoRA adapter this system has ever trained could not move.
What was left was the per-row magnitude, which rescales a row and does nothing
else.

Nothing failed. train adapter ran, reported a falling loss (the magnitude was
still training), wrote a file, and the judges compared the result against the
baseline and rejected it.

If you have been running trials, read this part. Those rejections were made
against an adapter that was structurally incapable of learning, so they say
nothing about the grid point they were testing. After updating:

godel forget

That clears the tried-markers and walks the search grid again. J4's cost
figures stand — the shapes did not change, only what the shapes could learn.

diag adapterinit is the check that would have caught this, and it ships here
too: it differences the adapter's own gradient and asserts the low-rank branch
is not identically zero on a freshly constructed one.

New: gpu

gpu

Reports what is on the PCI bus, whether every bridge forwards a bus that
answers, and — when the part is NVIDIA's — maps BAR0 and decodes
NV_PMC_BOOT_0 into a chip id. From ring 0, with no vendor driver anywhere in
the path.

It refuses to name a chip it cannot identify rather than guessing. Under QEMU
that is exactly what it does: the emulated VGA answers 0x001c222c, which is
not an NVIDIA boot register, and the decoder declines.

If you run this on real hardware, the output is useful. The binding
constraint on that work is that there is one machine here to test on and no
second sample. Every install that runs gpu and reports back is a data point
on whether a from-scratch, non-Unix OS can see and address a discrete GPU.

New: study and abstract

Both report-only in this release. They measure and print, and change nothing.

• abstract enumerates the subtrees of every program under /ai/tools,
  canonicalises them so structurally identical code with different names
  collides, and ranks what repeats by how much naming it would save.
• study and study seq measure what learning one field costs the fields
  already learned.

Tested under emulation, and not on hardware

Every check here ran under QEMU with the Windows hypervisor accelerator. The
boot selftests pass with no failures, and diag re-runs the suites by name.

None of it has run on real hardware. That matters most for gpu, which has
never seen a GPU: its refusal path is exercised and its success path is not.

Files

glados-1.2.29.efi	the kernel image, and what the in-OS updater installs
glados-1.2.29.efi.sig	its detached GLADOSIG signature
manifest	the signed manifest the updater reads
glados-1.2.29.iso	a full install image: kernel, model, tokenizer, roots. 576 MB

The ISO for this release was built by hand. From the next release CI builds
them, from a payload pinned outside the repository and verified against
recorded digests before the image is assembled.

Updating in place

On a machine that already trusts the signing key:

update check
update fetch
update stage <first eight characters of the digest>

then reboot. The image is replaced by the next boot, not the running one.
Installing fresh is the ISO.