eval: publish Inkling Small max-effort sweep

[antoinezambelli]

Summary

• add a managed inkling-small-rpc batch configuration for Inkling Small
• generalize RPC topology attachment from the DeepSeek-specific path to configurations explicitly marked as requiring RPC
• publish the 26-scenario, N=50 Inkling Small max-effort sweep
• append 2,600 rows to the canonical v0.9.0 corpus
• update the publication contract and regenerate the Markdown and HTML reports

Evaluation setup

• Model: Inkling-Small-UD-IQ4_XS
• Backend: llama-server, native tool calling
• Deployment: two-node Vulkan RPC with a 50/50 layer split
• Context: 262,144 tokens
• KV cache: F16
• Reasoning effort: max
• Reasoning replay: none
• Generation: 3
• Coverage: 26 scenarios × 50 runs, Reforged and Bare

Results

Configuration	Score	Validated accuracy	Completion	Avg time
Reforged	1187/1300 (91.3%)	94.6%	96.5%	124.4s
Bare	749/1300 (57.6%)	93.3%	61.8%	113.6s

The run completed without backend connection failures or context compactions.

Publication

The canonical v0.9.0 source now contains 28,600 unique rows. Publication view counts, metric counts, source hashes, and snapshot digests were recomputed from the finished corpus. Generated reports include the new Inkling Reforged/Bare comparison.

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

Validation

• python -m tests.eval.publication --source-root .
• focused publication/report suite: 91 passed
• full unit suite: 1,537 passed, 42 deselected
• 28,600 canonical rows with 28,600 unique evaluation identities

🤖 Generated with Codex and human-reviewed.