# Terminal-Bench: a coherence check for the served checkpoint

This is not a leaderboard claim. The purpose of this run is to show that the
exact checkpoint this repository serves (an abliterated, W4A16-quantized
DeepSeek-V4-Flash driven through a local vLLM endpoint at consumer-GPU decode
speeds) is coherent in real agentic work: it plans, issues shell commands,
reads their output, corrects course, and finishes tasks inside time budgets,
under a harness it has never seen. Quantization and abliteration both carry
coherence risk; this documents the check.

## Method

- Benchmark: Terminal-Bench core 0.1.1, 18-task subset.
- Agent: terminus, pass@1. Tasks run in Docker containers and are judged by
  each task's own hidden test suite.
- Timeouts: configured per task in the dataset (`max_agent_timeout_sec` /
  `max_test_timeout_sec`).
- Inference: OpenAI-compatible endpoint on the 8x3090 rig, 2 concurrent
  trials, DSV4 on the 3x114K envelope (~33 tok/s per stream).
- Calibration baseline: the same 18 tasks on
  [Laguna-S-2.1-W8A16](https://huggingface.co/Relativ3pa1n/Laguna-S-2.1-W8A16)
  and its W4A16 sibling, run on the same rig class and published on that
  model card.

## Result

DSV4-W4A16 solved 5/18 (27.8%), matching the Laguna-S-2.1-W8A16 calibration
point (5/18) and beating its W4A16 sibling (4/18). A separate 3-task screen
the same day went 2/3, with sanitize-git-repo reproducing across runs.

The coherence verdict is in the behavior, not just the score: the model
completed full multi-step tool loops, recovered from failed commands, and
finished one task (download-youtube) inside a per-task clock that the
calibration model timed out on at the same decode speed class. That points
to sound step economy, the thing long degraded outputs lose first.

## Per-task detail

| Task | Laguna-W8A16 | DSV4-W4A16 |
| --- | --- | --- |
| crack-7z-hash.easy | pass | pass |
| prove-plus-comm | pass | pass |
| sanitize-git-repo | pass | pass |
| sqlite-with-gcov | pass | fail |
| path-tracing | pass | fail |
| download-youtube | fail (agent timeout) | pass |
| sqlite-db-truncate | fail | pass |
| swe-bench-astropy-2 | fail (10/11 sub-tests) | fail |
| blind-maze-explorer-5x5 | fail (1/2 sub-tests) | fail |
| pytorch-model-cli.hard | fail (test-harness timeout) | fail |
| chess-best-move | fail (agent timeout) | fail |
| polyglot-c-py | fail (agent timeout) | fail |
| extract-moves-from-video | fail | fail |
| password-recovery | fail | fail |
| security-vulhub-minio | fail | fail |
| qemu-alpine-ssh | fail | fail |
| build-linux-kernel-qemu | fail (agent timeout) | fail |
| cron-broken-network | fail (test output unparseable) | fail |

Notes: only 3 of 5 passes are shared between the two models; each solved two
tasks the other missed, and the union is 7/18 (38.9%). Both models fail the
same infrastructure-heavy and long-generation tasks. At ~35 tok/s local
decode, the per-task clock is as much the opponent as the model is.