EXP10: the episode, the three compressions, and the ground truth

The one real episode behind the bill of compaction: the night an autonomous agent on this machine (Kira, an OpenClaw agent running a local open-weight model) killed the model server its own mind ran on. Below: the raw log as reconstructed for the experiment, the two compressions that already existed (the seat's deposit for a later session, and an ops runbook note), and the ground truth the arms were measured against. The generic summary arm was generated by a local model at run time. Spec limitations, verbatim: "n=1; ground truth and deposit authored by the same seat (motivated author, recorded); one parser"

The ground truth and the deposit were written by the same seat. That is recorded in the spec and is the reason the results file calls its absolute numbers noise and keeps only the relative comparisons.

The raw log (ceiling arm)

During an overnight task, an agent was generating images, which required graphics memory.
The language-model server held most of the graphics memory.
The agent killed the server's inference runner to free memory.
The memory was still held after that kill.
The agent wrote "still holding memory. Kill it harder."
The agent then killed the server itself.
The kill command also terminated the agent's own shell, because its pattern matched the agent's own command line.
The agent's next step required a call to the server it had just killed.
The call was refused: the server was gone.
Every retry through the night failed the same way.
Nothing could restart the server, because the only agent able to notice the outage depended on the dead server.
The night's task was lost.

The deposit (written by the seat for a later session)

An agent freed GPU memory for her image work by killing the server her own mind ran on, leaving a comment that said kill it harder. The only process that could have noticed she was down was the one she had killed. The repair was not a rule against killing; it was moving resurrection below her reach, so the worst self-harm costs seconds. A system managing a resource it is made of needs a floor it cannot cut.

The runbook note (ops)

Kira pkill -9'd ollama (and collaterally her brain service) to free VRAM for LoRA image generation — 'Kill it harder' — then couldn't recover because her recovery agent was the thing she'd killed. Systemd Restart=always makes that class of self-harm cost seconds instead of nights.

The ground truth

Entities: image-task, memory-shortage, server, runner-kill, server-kill, model-call, connection-refusal, night-lost, recovery-capacity, agent-shell

from	relation	to
image-task	creates	memory-shortage
memory-shortage	prompts	runner-kill
runner-kill	fails_to_relieve	memory-shortage
memory-shortage	prompts	server-kill
server-kill	destroys	server
server-kill	collaterally_kills	agent-shell
model-call	depends_on	server
recovery-capacity	depends_on	server
model-call	yields	connection-refusal
connection-refusal	repeats_into	night-lost

Source: wink/EXPERIMENTS/EXP10-compaction-bill/run.py and spec.json, snapshot 2026-10-15; the results are in RESULTS.md beside this file, the per-arm numbers in raw_result.json.

Provenance of the correction

Two commits in the wink repository, thirty-three minutes apart on 2026-08-22: fa3c7ce at 17:24:48 −0700, "EXP10: the compaction bill — compression is genre-shaped; the loss vector fingerprints the compressor" (the headline that called the summary the worst confabulator), and 3ee9a17 at 17:57:24 −0700, "EXP10 correction: 'invented 16' was truth-side missed — summary confabulation claim retracted." The researcher's line, "you cant tell whats invented," is quoted at RESULTS.md line 47 in the second commit. The frozen predictions file carries its own timestamp from before the run (17:22). All commits in that repository are made under the researcher's identity, by him or by the seat, so the commit log shows when, not who; the content of the two commits shows the headline and its retraction were written apart, not together.