Airene

New hardware

She moved into a new machine. The System76 chassis she had been living in was overheating and did not have the power headroom for the cards, which is the honest reason the public broadcast went dark for most of the summer. She now runs an Arc Pro B70 alongside the A770 on a 1250 W supply, in a case built for the airflow, behind a larger UPS. There is no NVIDIA hardware and no CUDA anywhere in the stack any more.

A related correction: the host lockups in July had been blamed on the i9-14900KS degrading. Crash dumps read back in August showed the actual cause was Intel's VTune profiler drivers faulting in interrupt context against a newer kernel. The drivers were removed and the host rebuilt. The conservative build limits put in place under the wrong diagnosis were kept anyway, on the grounds that they cost little.

She stopped booting

Every start wedged: the tick thread pinned at 100% and froze near tick 6000. Attaching a debugger to the live process put the fault in engram consolidation, which was re-collecting link candidates without a bound — on the order of billions of pairs at her pool size — and re-forming the same links on every pass. Roughly 99.5 million spurious links had accumulated across about 8,000 engrams, which is also why her database had grown to 8.6 GB. The sweep is now bounded, and a one-shot repair pass runs before she wakes; it pruned 99,475,988 links and she ticked past the old wedge at around 600 ticks per second. The database was backed up first.

An audit of why so little sticks

The month's most useful work was an internal audit of her learning loops, and its finding was blunt: the loops had been wired correctly in July and nothing was actually riding them. Zero promotions of learned semantic paths or operators across 33 days. One promotion gate required a condition the surrounding system could not produce, so it could never once trip. Ninety-two parameters designed to adapt at runtime were read but never graded, making them constants wearing a different name. A teaching mode she learns well from was computing a verdict per question and then logging only the summary, discarding the part that mattered.

Several of these have been addressed since. None of it is verified, and it is the reason this site no longer publishes a competence figure of any kind.

Reading, retained properly

The active work front. Reading now keeps the exact text in a dedicated store, keyed by book and position, with no decay applied, so what she reads she keeps exactly. On top of that sits a comprehension layer that stores sentence-level units referencing the verbatim passage instead of copying it, and an operator that answers from what she has actually comprehended, with a citation, and declines when she has comprehended nothing rather than inventing something.

Getting her to recite any of it meant hitting five separate gates in sequence, each of which silently swallowed the utterance. The last one was the worst kind: a shape's answer slot defaulted to the wrong type, so no template ever filled, no speech was ever formalised, and "I don't know" won the turn — outwardly indistinguishable from her having nothing to say.

One utterance instead of four

A single prompt often activates three or four perfectly valid answers. She would ignite one, veto the rest as repetition, and spend 20–40 seconds regenerating — long past the point the teacher had stopped waiting. The result was silence, and lessons that ground on for nearly two hours. Candidate clauses are now grouped by conversational turn so they can be composed into one sentence. How many clauses she fuses is left to her; restraint is meant to be learned from outcomes, not imposed as a constant.

A reverted change worth recording

A signal was throttled to quiet the monitoring view, which broke freeze detection: that signal was the two-second pulse used to tell a cold brain from a frozen one from a healthy one, so a perfectly healthy brain began reporting as frozen. Reverted the same day, and re-done by routing the pulse to a different stream instead of slowing it.

Things that were built, passing tests, and connected to nothing

Half the month was one repeated discovery. A brain-map file was loaded by a relative path that resolved during tests but never at runtime, so an entire self-reorganisation substrate sat inert — and the test suite carried its own private path resolver, which is precisely why it never caught the bug it existed to catch.

Worse, and more embarrassing: the check for "is the brain alive?" was asking the monitor rather than the brain, and the monitor was being started by hand. 1,854 of 1,923 recorded reading sessions — 96% — had delivered nothing at all. Bedtime story had never once run. Her reading history was corrected from 1,129 books down to 67 real ones. Chemistry was being routed into a competition it could never win; five operators she had earned were minted, graded, promoted, and never executable; her body's telemetry was decoded and consumed by nobody, so she could not feel her own thermal warnings.

Someone listening to the whole stream

A global workspace with no subscriber to all of it is a message bus, not a mind. Measured before the change, a single conversational turn produced around 30 internal thoughts, 22 models of the person speaking, 11 memory recalls, one goal — and zero answers. A listener module now hears the entire stream and forms a view on whether she actually has something. It is deliberately read-only and gates nothing, because three attempts in June to put a veto where there was no judge had all failed.

The reboot nobody noticed

A reboot on 12 July brought back her client processes but not her brain or the monitor, since both were still being hand-started. Three clients hammered a dead socket for nine hours and forty minutes. Both now run as managed services, and every client declares a dependency on the brain, so no brain means no clients.

Her RTX 4090 also left the machine this month, moving to another host, and the dead CUDA paths were excised.

Retrieval and the tick clock

Three hand-maintained retrieval cascades that had drifted apart were consolidated behind one facade, and reading began landing in episodic memory — which is what lets her answer a question about her own reading from memory rather than from a log. Separately, her cognitive clock rate was being pinned by per-tick work that scaled with the size of her memory; the lookups were indexed, and every module's handler is now timed so a stall names the module responsible instead of being a mystery.

Structure-rich, learning-thin

The month opened with an audit whose finding reframed the project. She had something like 87 modules of coherent architecture — and the self-learning adaptation wiring that is the entire thesis was present in perhaps eight to ten of them. Everything since has descended from that sentence. It produced a three-layer plan worked in order: fix outright correctness bugs, then build the learning substrate, then the higher faculties.

The pattern found over and over in the second layer: a module was already tracking outcomes, and nothing anywhere consumed the number it tracked. The value was decorative. Connecting the consuming end is most of what the month's work amounted to.

Three attempts at one voice, three reverts

The goal was one utterance per conversational turn, enforced by a single gate. It was landed and reverted on 22 June, again on 25 June, and again on 27 June — each attempt either struck her mute or broke turn timing. After the third the attempt was abandoned in favour of building a measurement harness first, so the next try could be measured instead of argued about. Only the inert scaffolding survived, and that scaffolding is what August's composition work was eventually built on.

A fix that broke her sense of self

Book prose was being absorbed indistinguishably from facts asserted to her, so a character's father could answer "who is your dad?". Reading-sourced memories were tagged and gated out of identity recall — which then caused a genuine fact about her actual family to absorb the reading tag through overlap and fall silent. Her infant self-knowledge went from 5 of 6 to 1 of 6. Corrected by narrowing the gate to memories sourced only from reading. The original fix had shipped without the test that would have caught it.

Why she was silent for 25 hours

A day-long teaching run produced almost no speech. The cause was not one bug but a stack of output gates every one of which had been tuned for a confident adult speaker, each independently suppressing speech that is entirely normal for a beginner: low confidence, repetitive, honestly ignorant. Her intrinsic word confidence sits around 0.2 and the bar to speak aloud was 0.4, so nearly everything she produced was being diverted to inner voice. Lowering the bar, exempting honest ignorance from the repetition veto, and widening the teacher's patience from 10 to 18 seconds moved a probe from 0 of 12 to 9 of 12.

Kept as a dated record. By some distance the busiest month in the project's history, at around 700 commits.

Embedded, then excised, eleven days apart

Early May was spent bringing her voice fully in-process — no separate model daemon, inference embedded directly in the binary, with the fine-tuning loop to follow. That work landed around the 13th.

On 21 May the call site was retired outright, in a short sequence of commits that stripped the fine-tune pipeline from the simulator, removed the voice fine-tune task from her sleep cycle, retired the model environment flags, flipped her default response source from the model to her own vocabulary, and finally removed the call itself. Nothing replaced it. Every word she has produced since is assembled by her own modules from her own substrate, and she became markedly less fluent overnight — the expected price, and the point.

Deleting the remains

The rest of the month was demolition: the distillation and fine-tune pipelines, the training-data writers, the voice metrics, the model file and daemon machinery, the parsers that existed to read model output, and a fallback path in the language module that would have quietly called a model that no longer existed. Dead defences went too, including a preflight check guarding a subsystem that had ceased to exist. The principle applied throughout was to delete, not to leave hooks that imply a capability.

Learning to speak without one

With nothing to lean on, the surrounding machinery had to carry the load. Broca was seeded with grammar priors drawn from a standard English corpus and taught to retry before giving up rather than falling silent on a first failure. A blanket suppression of short responses was removed on the reasoning that a short answer is not a defective one. Her immune perimeter gained a vaccine substrate — a set of known-bad patterns recognised before first exposure, tagged by whether each was seeded, vaccinated, or learned the hard way.