hqeric/maka-eval-trajectories

Maka Eval Trajectories

Private trajectory snapshots for evaluating Maka, Codex, Claude Code, and future agent harnesses under shared benchmark conditions.

Available runs

• Terminal-Bench 2.1:
  • model: DeepSeek V4 Flash
  • reasoning effort: max
  • arms: Maka, Codex, Claude Code
  • run ID: deepseek-three-way-v2
  • latest snapshot: latest.json
• DeepSWE 1.1 full:
  • model: DeepSeek V4 Flash
  • reasoning effort: max
  • arms: Maka, OpenCode
  • run ID: deepseek-v4-flash-maka-vs-opencode-deepswe-full-v1
  • latest snapshot: latest-deepswe.json

Each benchmark has an independent latest pointer so uploading an older completed run does not replace a newer benchmark snapshot.

Two layers are available:

• snapshots/: compact normalized trajectories with benchmark test/solution outputs removed
• raw-private/: per-cell archives containing raw sessions, thinking, tool calls, tool outputs, configs, verifier files, and telemetry

Privacy and redaction

• the normalized layer excludes raw session configuration and shell snapshots
• the raw-private layer preserves sessions and benchmark leakage evidence
• both layers redact secret-shaped values and runtime credential fields
• duplicate binary runtime.sqlite* stores are excluded because they contained credential-shaped fixtures already represented in JSONL
• DeepSWE raw archives exclude submitted workspace snapshots, dependency trees, Git object stores, and compiler caches; trajectories, sessions, patches, telemetry, verifier artifacts, and final results remain available

This repository is private and intended only for explicitly invited collaborators. Do not make it public without a separate review.