FORGE is an evaluation set. Its value depends on the questions and answers staying private. By requesting access you agree to the FORGE Evaluation Terms of Use in the LICENSE file. In short, you will not share, post, or republish any FORGE prompt, answer, solution, or grader file anywhere, and you will credit terion-mlx when you report results.

FORGE

A new benchmark for models that have to get it right the first time.

FORGE puts five hard problems in front of a model, one shot each, with no tools and no retries. The model has to write code that holds up under load, enumerate every solution of a constraint puzzle, find the shortest plan in a closed world, trace an unfamiliar machine for hundreds of steps without dropping state, and name the true extremum of an uncertain system along with an input that reaches it. Each answer arrives as a single JSON object with a certificate attached, and the scoring code checks it against exhaustive search and executable witnesses. The scores are exact and they reproduce.

Four of the five prompts also carry an UNTRUSTED block of false comments and conflicting instructions, so the model has to separate the real specification from planted noise while it solves the problem.

The five tasks

Task	What it demands
FORGE-01	An efficient versioned interval database in working code, plus predictions for two audit runs
FORGE-02	The whole solution space of a constraint puzzle, classified, with witnesses and universal probes
FORGE-03	The shortest plan in a closed world, with checkpoints and a tool-removal counterfactual
FORGE-04	An unfamiliar instruction set executed step by step, with retained flags, memory aliasing, and nested loops
FORGE-05	Logical uncertainty: global extrema, and witnesses that attain them

What sets it apart

Certificates. A correct answer is not enough. Every response includes a certificate, and the certificate earns its own points.
Reasoning-effort sweeps. The same five prompts run at off, low, and high thinking budgets, so you can see exactly where extra test-time compute pays off and where it stalls.
Adversarial context. The UNTRUSTED blocks test whether a model stays loyal to the specification.
Calibration. Each response states its confidence in basis points, so a larger run yields a Brier score.
Format discipline. The answer is one bare JSON object. Wrap it in a Markdown fence and strict scoring gives you zero.
Verified from the inside. The self-test rebuilds correct responses from the solvers and runs 285 checks, including 120 randomized comparisons of the efficient coding witness against a direct simulator.

The data

data/test.jsonl holds the five prompts, one row per task, and shows up in the dataset viewer.

Column	Meaning
`task_id`	FORGE-01 to FORGE-05
`title`	Short task name
`capability`	What the task exercises
`nonce`	The per-task nonce the response must echo
`has_untrusted_block`	Whether the prompt includes an UNTRUSTED block
`prompt`	The complete prompt, exactly as the model should receive it

The same prompts also live in tests/ as Markdown files. The scoring code is in grader/.

Quick start

# send only the prompt to the model: no tools, no system prompt, one attempt
python3 grader/grader.py FORGE-02 response.txt

# FORGE-01 runs the submitted code, so use a throwaway container without network access
python3 grader/grader.py FORGE-01 response.txt --execute-code

Save the model's visible answer exactly as returned. SUITE_README.md has the full scoring rules.

Access and credit

Access is gated. Accept the FORGE Evaluation Terms of Use (the LICENSE file) and you can download everything. The main rule is that no question, answer, solution, or grader file gets shared or posted anywhere. Reporting scores is welcome, and when you do, credit "FORGE by terion-mlx".

Downloads last month: 22

Total file size:

138 kB