inc-0008
data-integrity
citable URL: https://halobench.com/records/inc-0008/ — this address never moves; the anchor /records/#inc-0008 keeps resolving
HG-003 nominal depth prompt exceeded its server context before placement
detected
2026-08-19 · window 2026-08-19T13:23:15Z → 2026-08-19T13:23:17Z · by server-log-and-prompt-metadata-review · lag immediate at failed-attempt ingestion
signature: The nominal depth-2048 prompt recorded 6691 prompt tokens; the server counted 6779 request tokens against n_ctx=6144 and returned HTTP 400 before placement/generation.
should have caught it
The runner should have tokenized the complete rendered request, including template and tool overhead, and rejected or resized it before starting the scientific arm.
diagnosis
initial: LFM2 failed retrieval at depth 2048
actual: Repeated filler text expanded to far more model tokens than the nominal depth and the configured context could hold. The server rejected the request with zero tokens processed, so no retrieval observation or safe depth exists.
false-path cost: model load plus one rejected request; stage B was correctly skipped
falsified by: prompt metadata, server request-admission error and header-only retrieval.tsv
actual: Repeated filler text expanded to far more model tokens than the nominal depth and the configured context could hold. The server rejected the request with zero tokens processed, so no retrieval observation or safe depth exists.
false-path cost: model load plus one rejected request; stage B was correctly skipped
falsified by: prompt metadata, server request-admission error and header-only retrieval.tsv
lesson
Retrieval depth must be defined and gated in model-token space after full request rendering. HTTP admission failures are runner invalidity, never capability failures.
state
resolved · class config