AI Server Home — technical evidence appendix, compiled October 20, 2026

Data retained at original measurement precision. Historical results are September 29–30;
current application results are October 20. The original compilation used preserved measurements. The expanded lesson-generation run adds 12 new real-model requests; its summary, per-request CSV and full responses are included as expanded-lesson-* files.
embedding-requests.csv: every request, actual input, hash, status, latency, dimensions, norm.
embedding-vectors.json: all 400 returned 1024-dimensional vectors and request details.
embedding-telemetry.json: all 838 captured host/resource observations, with infrastructure IDs omitted.
model-comparison.csv and model-requests-and-answers.json: all eight completed matched educator
responses, prompts, token counts, machine-check failures and answer text.
failed-trials.json: separate failed first model comparison and embedding preflight evidence;
not a complete reproduction of infrastructure/controller logs, never pooled with completed data.
application-requests.csv and application-requests-and-answers.json: all 16 submissions from
October 20, including eight busy rejections, measured requests, generated outputs and constraints.
summary.json: computed timing summaries plus explicitly documented historical resource findings.
manifest.json: original source hashes and hashes for this sanitized export.

Scope: relevant recovered studies, not every experiment ever run. Original immutable archives
remain the audit authority. These exports intentionally exclude SSH/private network addresses,
credentials, operator identities, raw infrastructure logs, and private student/database records.
The source-operation names are retained; whole standards payloads are omitted. Standards
provenance: Learning Commons; Kansas State Department of Education; Achievement Network;
source release v1.13.1, CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

4B speed is not proof of equal quality: 9B passed 2/4 full check sets and 4B 0/4.
Machine checks are diagnostics, not educator scores. One response per case/model.
No 29B result was located. No model is represented as Intel-certified.
