OpenPond model card · Harness Refiner v1

Across the same twenty tasks, one Harness refinement reduced foreground tokens on thirteen tasks and cut total usage by 41.38%, without changing the model, runtime, or tools.

Token result

13 / 20

tasks used fewer tokens

7 tasks used more tokens

All 20 tasks

−41.38%

5,599,785 → 3,282,811

Adaptation

5/10 lower

−28.35% in aggregate

Held out

8/10 lower

−57.87% in aggregate

Per-task token direction

Each square is one paired task under the baseline and refined Harness.

Adaptation

5 lower·5 higher

Held out

8 lower·2 higher

Foreground tokens

Provider-reported totals across the same paired tasks.

All 20 tasks

−2,316,974 raw change

Baseline5,599,785
Refined3,282,811

Adaptation

−887,153 raw change

Baseline3,128,871
Refined2,241,718

Held out

−1,429,821 raw change

Baseline2,470,914
Refined1,041,093

What the run showed

Refiner cut total foreground usage by 41.38%, but the reduction was not consistent across the suite: thirteen tasks used fewer tokens and seven used more. This run did not reach the intended 20/20 per-task efficiency target.

The refinement loop

Refiner waits until work settles, learns once from the adaptation cohort, then proves the release on replay and unseen tasks.

  1. 01

    Work

    Run the adaptation tasks with one pinned Harness release and capture the result, tools, artifacts, grades, and usage.

  2. 02

    Refine once

    Review the completed cohort together and propose one small, reusable Harness change—or no change at all.

  3. 03

    Validate

    Independently critique the proposal, enforce bounded schemas and permissions, then create an immutable release.

  4. 04

    Replay

    Run adaptation and held-out tasks again with the same model, runtime, tools, seed, and sampling policy.

Evidence in, bounded update out

Refiner sees the requests, visible answers, tool outcomes, recoveries, artifacts, grades, usage, and the pinned Harness. The model decides what is reusable; deterministic validation owns schemas, permissions, safe paths, and release identity.

InstructionsSkillsAgentsApproved memoryTool capabilities

A controlled comparison

  • Same model revision and high reasoning effort
  • Same local Work runtime and ten tools
  • Same seed, temperature, top-p, and output limit
  • Frozen research evidence replayed across each pair

Inspect the complete run

The public Taskset includes all twenty tasks, fixtures, graders, and the deterministic runner. The technical note covers the evidence contract, paired protocol, and measured result.