eval: publish Ornith 1.5 9B and 35B sweeps

[antoinezambelli]

Summary

• publish complete Ornith 1.5 9B and 35B evaluation sweeps;
• append 5,200 generation-3 rows to the canonical v0.9.0 corpus;
• add the corresponding sampling defaults and managed batch configurations;
• normalize both Ornith families in reports;
• update the deterministic publication contract; and
• regenerate the Markdown reports and HTML dashboard.

Evaluation setup

• Models:
  • Ornith-1.5-9B-Q4_K_M
  • Ornith-1.5-35B-Q4_K_M
• Backend: llama-server with native tool calling
• Quantization: Q4_K_M
• Reasoning replay: none
• Reasoning level: default
• Generation: 3
• Coverage: 26 scenarios ? 50 runs in both Reforged and Bare modes
• Total new rows: 5,200

Results

Model	Configuration	Score	Validated accuracy	Completion	Avg time
Ornith 1.5 35B	Reforged	97.8%	97.8%	100.0%	7.1s
Ornith 1.5 35B	Bare	60.3%	90.8%	66.4%	9.0s
Ornith 1.5 9B	Reforged	94.2%	94.2%	99.9%	11.1s
Ornith 1.5 9B	Bare	62.2%	91.2%	68.2%	12.5s

Publication

The canonical v0.9.0 source grows from 28,600 to 33,800 unique rows. The new rows preserve the existing source byte-for-byte and append four complete 1,300-row configurations.

The refreshed publication contract contains:

• history: 552,500 rows;
• snapshot: 412,100 rows; and
• latest: 293,800 rows.

Reports group the new models as ornith-1.5-9b and ornith-1.5-35b, with Q4_K_M recognized as the quantization. The Markdown reports and HTML dashboard were regenerated from all six canonical JSONLs.

The Hugging Face Parquet bundle will be rebuilt and published separately from merged main.

Validation

• strict six-source publication audit: passed
• focused publication, reporting, and configuration suite: 106 passed
• full suite: 1,538 passed, 1 skipped, 42 deselected
• v0.9.0 corpus: 33,800 rows with 33,800 unique evaluation identities
• LFS pointer and attributes verified
• git diff --check: clean
• generated dashboard: human-reviewed and approved

Release

No package-version bump is included. This is an evaluation corpus and reporting update.

?? Generated with Codex and human-reviewed.