Reports
Writeups & infographics — the project story, the harness, and the numbers. Newest first; click any card to open the full report.
July 2026
2026-07-06
The catch-up — a week that rebuilt the bench
What shipped in a short stretch: the v2 platform, the run queue, harness A/B, data-integrity guards, and the two-box cluster experiment.
open report →
2026-07-03
When more scaffolding made it worse
The harness A/B view — and the real v1.7 regression it caught on qwen3.6-35b-a3b.
open report →
2026-07-03
A good coder isn't automatically a good agent
Reasoning vs basic agency, and a new axis: how hard we had to fight the model — the compliance metric.
open report →
2026-07-03
Rebuilding the bench on solid ground
The Postgres + FastAPI + Next.js v2 rebuild, and the data-integrity guards that keep the board honest.
open report →
2026-07-03
Benching on demand
The request-a-run queue — approve, schedule, ETA, reorder, and drill into every result.
open report →
June 2026
2026-06-29
From a flat board to a climb
The gauntlet journey — turning a flat leaderboard into a climb, with names and coverage.
open report →
2026-06-29
mdskills — DB-driven injections
Context injections sourced from the database, with the harness left completely pure.
open report →
2026-06-29
The GX10 cluster — two boxes, one pooled brain
Linking two GX10s into one pooled inference brain for frontier-size models.
open report →
2026-06-27
The catch-up — the bench grows up
A progress catch-up as the bench matured — numbers, coverage, and what changed.
open report →
2026-06-24
Project journal — teaching small models senior-dev work
The origin journal: getting small local models to do real senior-developer work.
open report →
2026-06-24
Hermes × Paperclip adapter
The adapter wiring local models into the Paperclip agent company.
open report →
2026-06-24
Comparison matrix
Where the local stack sits against the alternatives, side by side.
open report →
2026-06-24
localharness — the pitch
The case for localharness — scaffolding that lifts weak local models.
open report →
2026-06-24
localharness — capabilities
What the localharness scaffolding actually does, capability by capability.
open report →
2026-06-24
Cost of the bench — local vs OpenRouter
What it costs to run the bench locally versus paying for frontier APIs.
open report →