Tiny-runtime rejection Report artifact 2026-09-01

Needle 2 Tool-Selection Boundary

The evaluation deliberately tested the most favorable remaining systems hypothesis: reduce the tool catalog before considering any fine-tune. Overall exactness rose four points, but the Pace slice regressed, file operations stayed at zero, and out-of-scope actions remained. The result closes the current integration path without pretending that packaging efficiency is capability.

Headline Numbers

Full-catalog exact

32/94 34.0% public gate

Routed-catalog exact

36/94 38.3% with oracle task family

Pace slice

10/28 → 6/28 target regression

Competitive Context

System Metric Score Size / Class Comparable? Readout
Needle 2 full catalog exact tool selection 34.0% 94 public fixtures Direct Included eight out-of-scope false calls and two destructive-action calls.
Needle 2 oracle task catalog exact tool selection 38.3% same 94 fixtures Direct The upstream task family was supplied; this does not prove Needle can discover it.

Direct rows share this artifact's eval setup. Directional rows are useful market context but should not be read as leaderboard claims.

Catalog ablation

GateFullTask catalogReadout
Overall exact32/9436/94+4, below usefulness
Pace exact10/286/28regressed
File operations0/60/6no capability
OOS false actions83still unsafe

Release Blockers

Capability and safety gate failed

The favorable catalog ablation regressed Pace and retained unsafe false actions.

Unblock: Do not unblock historically; any future work needs a materially different model/training hypothesis and a new gate.

Evidence

Next Release Action

Closed and rejected for Pace. Preserve the 149.7ms mean latency and 27.7MB RAM result as systems evidence only.