KKingy AILOCAL LABObserved Aug 18
CHANGELOG · MATERIAL UPDATES ONLY

What changed—and why it matters.

A new timestamp is not an update. This log records source changes, formula revisions, reviewed evidence and benchmark-invalidating changes.

DATASET 2026.08.18.3

Receipt contract and claim-scope correction

  • Made every public receipt valid under the published Draft 2020-12 JSON Schema.
  • Added standards-compliant schema execution plus repetition, statistics, memory, runtime-argument and hash consistency checks.
  • Reframed 124,928/125,952 as an allocated-context VRAM-fit boundary; measured requests used 512 input and 128 output tokens.
  • Qualified the pinned upstream base revision because exact GGUF conversion lineage was not independently established.

Re-benchmark required: No for the corrected allocation-fit claim. Yes before claiming near-full-context prompt operation.

DATASET 2026.08.18.2

First reviewed RTX 4090 benchmark series

  • Added ten observed Qwen3.8-27B Q4_K_M success and failure receipts.
  • Verified the 17,106,775,008-byte artifact and SHA-256 against the pinned GGUF revision.
  • Observed successful short requests with 8K, 32K and 64K allocations; fixed 128K allocation failed with CUDA OOM.
  • Bracketed the FP16 KV allocation boundary: 124,928 succeeded and 125,952 failed.
  • Published raw evidence archive, receipt JSON, versioned exports and explicit limitations.

Re-benchmark required: Yes after any artifact, llama.cpp commit, KV-cache precision, batch, offload or GPU change.

DATASET 2026.08.18.1

Lab foundation and Qwen3.8-27B record

  • Added immutable Qwen3.8-27B revision and architecture inputs.
  • Added official RTX 4090, RTX 5090 and Mac Studio memory records.
  • Published estimator v0.1.0 with separate weights, KV, runtime and reserve terms.
  • Opened four high-priority evidence tasks.

Re-benchmark required: No—there were no prior Kingy runs.