professor-x-AGI

Read the actual failing trajectory before fixing an agent/harness failure on Professor X — never guess. Use when a HIRO or repo-fix task fails, pass@1 stalls or won't rise, the agent loops/thrashes, an edit doesn't land, or you're about to 'improve' a prompt or tool. Guessed fixes failed; trajectory-diagnosed fixes lifted repo-fix 0.50→0.77. Ships a Tier-3 helper: scripts/pull_trajectory.py.

Author a new repo-fix benchmark fixture for Professor X correctly. Use when adding/expanding coding tasks for --repo-fix-bench, growing the benchmark, or making it harder/more representative. Enforces the eval-trust rule: a fixture is only valid once you've verified it goes red→green (broken→fixed) with stdlib checks (NOT pytest — pytest is not installed).

Two-thread adversarial review of your OWN work before you commit it on Professor X: steelman it, then attack it. Use before any commit or PR, before recording a result in docs/memory, before declaring a milestone done, or when you're about to claim a 'win' or improvement. Adapted from ARIS's kill-argument; the same pattern lives in Prof X's agent.critic.

Operational interface for running and measuring Professor X. Use when running or interpreting --repo-fix-bench, --hiro-null, --hiro-smoke, --evolve-on-repofix, or --evolve-skill-on-repofix; when picking the model/data-dir; when a run needs Ollama; or when committing/PRing work on this project. Knows the K=2/MDE conventions, the trustworthy-vs-untrustworthy metrics, and the branch/PR policy.

The safe workflow for changing Professor X's Rust harness code (react.rs, executor.rs, hiro.rs, main.rs, proposer/loop_runner). Use when editing the harness/agent loop, tools, evaluation, or evolution code; before committing a Rust change; or when a change touches behavior a test or benchmark covers. Enforces: build → FULL test suite → measure on the trustworthy benchmark → commit on harness-gaps → PR.

Design principles for making a weak local model (qwen3:8b) into a capable agent on Professor X — the project's core thesis. Use when designing/changing the agent loop, tools, prompts, or evolution; when deciding how to fix a capability gap; or when tempted to blame the model. The recurring finding: the gap is almost always the HARNESS, not the 8B's reasoning.

Mechanism-check a metric before trusting it on Professor X. Use when about to report/record/act on a pass-rate, pass@1, pass@3, φ, or any score; when a number jumps or looks too good; before writing a result to docs/research or memory; before calling a milestone done; or when a metric reads a suspicious constant (0.000, always-same). This project has shipped fabricated 'confirmed' metrics — M0 exists to stop that.