# Kolibri placement and response investigation

## Observed behavior

Read-only runner placement, broker journals, and request metadata on October 7,
2026 established that `robit/kolibri-1:q4_k_m` loaded 81/81 layers across the two
operator-selected A100s at PCI `01:00.0` and `82:00.0`. Its native context was
262,144 tokens. Both cards showed the same runner, roughly 37 GB allocated per
card, and active utilization during generation. Timings reported approximately
50 generated tokens per second. The third A100 and display GPU were not used.

The observed quiet periods included avoidable model replacement. Optional
semantic-memory embeddings evicted a freshly warmed exclusive Kolibri group;
Kolibri then had to load again. Later SNR relevance/noise evaluators requested
the previous Qwen model because their engine retained the configuration object
from before the parent's model selection changed. This was distinct from a
failed tensor split, GPU fault, or Telegram model override.

At the initial placement observation, the installed daemon and TUI had adopted
version 1.0.813. A later automatic update adopted 1.0.814 with the same older
CLI bundle; neither version contained the repairs described here. The affected
news request originated in the TUI; the earlier daemon GUI session was not its
request owner.

## Repairs and acceptance boundaries

- Resolve the SNR parent's configuration at each evaluation batch and share
  that snapshot between its two evaluators. Model, endpoint, provider protocol,
  credentials, and timeouts must follow the same parent selection. Existing
  callers supplying a fixed configuration remain supported.
- Mark optional memory embeddings as broker `background`/`yield` work. The
  broker permits warm reuse and genuinely free capacity, but returns a
  retryable `background_capacity_deferred` response before any resident lane
  would need reclamation. Durable memory and text retrieval remain available
  when embeddings are unavailable. Foreground and explicit background/wait
  requests retain their existing admission behavior.
- Preserve the actual six-empty-response terminal cause in a typed,
  run/epoch-bound failure receipt. Successful read-only observations are not
  command receipts, and their absence must not replace a known model-response
  failure with a generic verification diagnosis. Failed checks remain failed;
  an empty response never constitutes an accepted answer or completion.

The broker change preserves exclusive peer groups, GPU health checks, scoped
leases, context accounting, and reservations until the entire owning process
group exits. Neither repair lowers context to avoid a failure or redirects
inference to another GPU.

## Remaining causal uncertainty

The news run executed successful web tools and subsequently received six
responses without usable visible text or tool calls. Existing journals retain
token counts and timings but not the raw response envelopes. They cannot
establish whether the model emitted hidden reasoning, control-only output, or
output lost during parsing/normalization. A positive generated-token count is
not evidence of a visible answer.

Regression tests establish the repaired host behavior using isolated CPU
fixtures and mocked transports. They do not certify a successful live news
answer or completion of the earlier Telegram PDF task. A broker restart and
supported client update are separate adoption boundaries; source commits and a
fresh publication artifact alone do not update already running processes.

## Later Telegram PDF findings

The subsequent admin DM used the verified 262,144-token context with roughly
20,000 estimated input tokens. Its budget nevertheless declared compaction
eligible: a roughly 217,000-token optional generation ceiling was counted as
occupied context, and the compiler then tried to reduce total occupancy to a
fixed fraction of the window. The fixed output allowance alone made that
target impossible. Generation allowance is not accumulated input; compaction
pressure and its target must instead use input, the existing minimum reply
allowance, admission safety, and the verified context capacity. The final
transport admission guard remains independently authoritative.

Two further host boundaries were implicated. Capability discovery satisfied a
generic observation criterion before a PDF was read, so the objective
controller blocked the actual extractor as delivery due. A deictic native reply
also lost its exact referenced upload batch when the durable objective was
created, leaving a coarse contract instead of per-source work. Preparation
receipts must remain auditable without standing in for requested source
evidence; native attachment lineage must reach both the task contract and the
runner request.

A later numerical reply was delivered despite an incomplete-verification
terminal record, failed extraction evidence, and no PDF extraction in that
run. Independent CPU extraction of all 23 hash-matched source PDFs established
that the answer was wrong, including internal arithmetic contradictions.
Private document values and conversation identities remain outside this
repository. Provisional provider-delivery authority must reject failed work
evidence and blockers, while allowing a genuinely ready answer whose only
remaining obligation is the provider send. Action drafts are not accepted
answers merely because the model produced them.

## Passive Telegram history acceptance boundary

Under genuine input pressure, passive Telegram chronology should be the first
context externalized. The current native request, selected reply ancestors,
task contract, attachment workset, and observed task evidence remain in the
request. Ordinary history remains available when input fits; a large optional
reply ceiling alone must not trigger history removal.

The distinction must originate in the bridge's typed context composition,
with exact scope, source hash, and authored offsets. Quoted headings or model
text cannot declare a historical segment disposable. Before replacing a
segment, the runner must durably save its exact bytes, flush storage, verify
readback, and provide an exact retrieval reference. Storage failure must retain
the original data. Externalization changes the request working set, never the
durable conversation, task ownership, evidence ledger, or completion status.

Acceptance requires production-boundary controls for preserved unpressured
history, actual input reduction before task compaction, unchanged current-task
and source bytes, valid retrieval, storage failure, stale/foreign manifests,
and recovery of full history in a later unpressured request.
