[KongchangAI]
Productphocinae

dsh-phocinae

A local decision-making plugin for the DeepSeek Harness (dsh) agent. It embeds a small model that answers structured decision questions in a single forward pass, and uses a confidence threshold to determine whether a query should be escalated to a hosted model.

Timeline (last 90 days)

Oct 13

dsh-phocinae is a local decision plugin for the DeepSeek Harness (dsh) Agent, providing two tools, phocinae_ask and phocinae_gate, backed by a small model that can answer structured decision questions in a single forward pass.

Unverified50%
Oct 13

With the same plugin and the same 400-case workload, the τ=0.4 threshold saves 86.1% of billed Tokens, while at the τ=0.6 threshold only 8.5% remains saved.

Unverified50%
Oct 13

A single local call has 99 tokens of arguments and 135 tokens of rendered result, i.e. 234 tokens per decision in steady state; if this is the first decision in a session, the fixed 685 is also added.

Unverified50%
Oct 13

On the pure local path, 1 decision saves 69.9%, 10 decisions save 90.1%, and 1000 decisions save 92.3%.

Unverified50%
Oct 13

In the hybrid mode at τ=0.4, 27 of the 400 cases are escalated (6.8%), billed Tokens drop from 1,222,211 to 82,499, and the saving including plugin overhead is 86.1%.

Unverified50%
Oct 13

τ=0.6 is the plugin's default factory value.

Unverified50%
Oct 13

On accuracy: τ=0.4 is 0.7305, the hosted single run is 0.7085, the local single run is 0.7320, and τ=0.6 is 0.7115.

Unverified50%
Oct 13

At τ=0.6 the confidence distribution is concentrated around a median of 0.6294, and 851 of the 2000 decisions (42.6%) fall below 0.6, which by case count is 90.8%.

Unverified50%

All Facts (8)

Source Articles