Experiment reports

Local measurements from September 18, 2026, with inputs, outputs, timings, and execution settings retained for each run. Start with the shared-state workload comparison for the current native JSON scorer.

initial-smoke.json: original uncached bfloat16 integration smoke test.
cache-bfloat16.json, cache-float32.json, cached-smoke.json: historical precision and cache checks.
component-profile.json, cache-trimmed.json: earlier answer-tail optimization.
full-detail-comparison.json, cache-gathered.json, smoke-gathered.json: full-depth answer-position gathering.
two-field-video.json: separate requests, shared serial fields, and shared batched fields.
catalog-comparison.json, smoke-catalog.json: experimental shared-question catalog.
head-pilots/: supervised text pilot configuration, dataset hashes, and failed quality results. No weights are included; reproduce using scripts/train_head.py and examples/head-pilot/.
head-runtime-*.json, encoder-check*.json: experimental head runtime and feature extraction checks. Runtime probes with untrained heads do not establish useful predictions.
validation.json, web-validation.json: historical local checks, not hosted CI results.
native-json-validation.json: native JSON scoring versus compact normal generation on text, images, speech, video, soundtracks, and the reported counting regression. Contains disagreements with expected answers and invalid normal outputs; see interpretation.
native-json-web-validation.json: checks through the restarted local web server for the full skunk video with its soundtrack and speech counts/negation. Both paths match the expected answers; the video expectation is baseline agreement, not independent annotation.
demo-workload-benchmark.json: warmed repeated comparisons on shared-state support, security, inbox, independent-label, and large-choice workloads. Includes serial scoring controls, all raw responses, and predeclared development expectations; see interpretation.
demo-workload-host-pressure.json: interrupted earlier attempt while the idle web model was also resident and host memory pressure/latency variation were observed. Retained separately and excluded from the subsequent run's medians.
demo-workload-web-validation.json: browser checks of all four workload presets and the live 28-field comparison, including the corrected default yes/no wording mismatch.

Absolute workstation model paths were replaced with models/gemma-4-e2b-it-4bit. Local command paths were normalized for publication; measurements and predictions were preserved. The exact model repository and revision are recorded in model-source.json.

All reproduction commands in the experiment notes run from the repository root. Write fresh outputs to the ignored work/ directory so archived measurements remain intact.