harnessie.com / docs / Brains

Brains run under Harnessie

Brain-agnostic is a testable claim, not a slogan. This page is the receipt: the models that have actually produced work under the harness, each linked to the on-disk record that proves it, plus the set you can swap to by editing one file.

Scope: the table below tracks models run under the harness at runtime, each backed by a decision record. The models that built Harnessie's own code are a separate and honest story, credited under Built with at the end.

Proven under the harness

Each of these formed an independent position in an adversarial contested phase, recorded verbatim with its model and provider in a hash-chained decision record. Different task classes route to different tiers, so a single contest draws genuinely different brains, which is what earns a record its independent-positions claim.

Four providers, four model families, from a frontier closed model down to 20B-parameter local open-weights models, all producing arbitrated positions under the same harness. The decision the operator made in each case was a human's; every position is preserved, including the ones that did not win.

Configured and swappable

Declared in config/models.yaml. Point any task class at any tier; the gates, jails, budgets, and audit stay identical. API keys come from environment variables, never the file.

Any OpenAI-compatible endpoint works with no code change: vLLM, Ollama, llama.cpp, Together, OpenRouter, Fireworks, DeepSeek, Mistral, xAI, and others. Swapping a provider is a model_id and base_url edit.

Built with

Development provenance, distinct from the runtime table above: these models built, reviewed, and fact-checked Harnessie during construction rather than running under it, so they are not each backed by a single decision record. The trail is in the git history, source-verification.json, and the session handoffs.

• Claude Fable 5 (Anthropic): the primary implementation and review model across the 0.1 to 0.4 line, and the frontier orchestrator in config.
• Claude Opus 4.8 (Anthropic): also a recorded significant development co-author in the git history.
• Claude Sonnet 5 (Anthropic): was used for several smaller cleanup tasks, especially those leveraging pre-existing skills.
• GPT-5.5 (OpenAI, via Codex): the secondary implementation and review model, cycling and verifying independently in alternating sessions.
• Gemini 3.5 Flash (Google): reviewed the 0.4.0 patch, verified the trust-bundle manifest and live scorecard skip policies, and ran the live scorecard against a local OpenAI-compatible endpoint.
• Perplexity Sonar Pro (Perplexity): research and independent source verification at the outset; the trail is source-verification.json, where twenty prior-art sources were checked and several confirmed via sonar-pro.

Four providers building, reviewing, and fact-checking the harness itself is the thesis applied to its own construction.

Coverage we would like to add

The claim gets stronger as it spans more of the capability curve and more independent providers. The proven table already covers a frontier closed model (claude-fable-5) plus three ~20-35B local open-weights families (Qwen, Google, OpenAI). The gaps below are the honest missing coverage, each closable by running the named model through a contested phase so it earns a record. Model names are Ollama pull identifiers as of the 2026-07-06 library scan; verify tags at pull time, and note that any tag reading cloud runs on Ollama Cloud rather than fully on-box.

More of the capability curve:

• The small-model floor, where structured tool-calling breaks first and the harness's tool contract is under the most stress. Everything proven so far is 20B or larger; nothing below that has been run. Strong candidates: granite4.1 (3B, 8B, 30B; IBM, tuned for disciplined function-calling), functiongemma (270M; a sub-1B model whose entire job is function-calling), nemotron-3-nano (4B; tools + thinking), or phi4-mini (3.8B; tools). If the registry's contract survives a 270M brain, it survives anything.
• A coding-specialist worker, directly relevant to the harnessie-worker role. Candidates: qwen3-coder (30B; tools) or devstral (24B; Mistral's agentic code-editing model, built for exactly this job).

More independent providers:

• A frontier closed model from another provider with a runtime receipt. GPT-5.5 already built the harness (see Built with) but has not yet run under it; a frontier non-Anthropic brain producing an arbitrated position (a GPT-5-class model, Google Gemini, or Grok) would put a second frontier provider in the proven table, not only in the build.
• Meta Llama (open weights): the most widely deployed open family, currently unrepresented. Candidates: llama3.3 (70B; tools) or llama3.1 (8B/70B; tools).
• Mistral: a European provider adding architectural and geographic diversity. Candidates: magistral (24B; tools + thinking) or mistral-small3.2 (24B; vision + tools).
• DeepSeek: a distinct strong-reasoning provider, already namechecked in the config comments. Candidates: deepseek-r1 (7B-70B local; tools + thinking) or deepseek-v3.2 (cloud).

How a brain becomes "proven"

Run it in a contested-decision workflow so it forms an independent position:

python3 -m harness.cli run workflows/contested-decision.yaml --goal "<a real decision>"

Route one of the positions at a tier pointed at the new model. When the panel resolves, the model and provider are written verbatim into runs/<id>/decisions/DR-<phase>.md. Promote that record into decisions/ and add a row above. Configured is a claim; proven is a record.