When the model knows the answer but forgets the assignment

What it does This repository is a benchmark report and engineering diagnosis arguing that DeepSeek V4-Flash and V4-Pro often fail to deploy their full agent abilities due to “capability-realization loss” — a gap between what the model can represent and what it actually delivers under specific tool schemas and first-turn conditions. It documents J-Space Cognition Suite V3.6, a separate plugin that attempts to close this gap by managing workspace state, routing, checkpoints, and verification loops without touching model weights. The included tables show score increases on HLE and Terminal Bench 2.1 when the suite is loaded.

The interesting bit The report coins “chain-of-thought diode” to describe a quirk where tiny changes in persona or tool schema can snap DeepSeek V4 onto an entirely different reasoning trajectory, rather than shifting behavior proportionally. J-Space responds with an almost bureaucratic level of process control: functional first-person pronoun contracts (I for judgment, we for action), a three-tier task gate (fast, full, loop), and a persistent five-field ledger (Goal / Core / Verified / Open / Next) to keep long-running tasks from derailing after tool calls.

Key highlights

• Claims HLE (with tools) jumps from 51.5 to 60.6 for V4-Flash and 60.0 to 67.7 for V4-Pro with J-Space loaded; Terminal Bench 2.1 also rises.
• Treats the model as a capable but context-sensitive employee who needs the right desk setup: it argues DeepSeek V4’s agent behavior is fingerprinted to specific Minimal tool schemas and first-turn formatting.
• Uses “dense tracks” for internal reasoning compression while requiring clear external language for user-facing and tool-facing output.
• Explicitly positions itself as broader than single-point anchoring but not a universal replacement — acknowledges dsh-anchored-standard and dsh-routing-suite handle narrower jobs more directly.

Caveats

• All benchmark figures are single-run snapshots without confidence intervals or multiple-run means, and competitor scores are copied from vendor public results rather than a shared harness.
• The report is licensed CC BY-ND 4.0, meaning you can cite it but cannot legally modify or redistribute derivative versions of the document itself.
• J-Space itself lives in a different repository; this repo is the report and evidence folder, not the plugin source code.

Verdict Worth reading if you are building agent harnesses for DeepSeek V4 and have noticed confusing variance between “official Minimal” and “Standard” tool setups. Skip it if you are looking for a drop-in code library or reproducible open-source benchmark infrastructure — the plugin is elsewhere and the raw data is thin.