Tiger3807861189 / J-Space-Cognition-Suite-V3.6

J-Space Cognition Suite V3.6

简体中文

J-Space Cognition Suite is a model-agnostic inference-time control system for deep reasoning, long-horizon work, tool use, verification, and recovery. It is packaged as a Skill for cross-platform use, selective loading, and low-friction integration.

The suite organizes an agent's accessible working representations into a deliberately managed workspace. It operates through a single entry, nine selectively loaded modules, three supporting references, and an optional standard-library controller for durable task state.

J-Space operates at inference time. Model weights and training remain unchanged.

Quick start

Option A — manual installation

1. Download or clone this repository.

2. Locate the user-level Skills directory used by your AI host.

3. Copy the complete j-space/ directory into it so that the installed entry is <skills-directory>/j-space/SKILL.md.

4. Run the integrity check with an available Python 3 interpreter:

   <python-command> <skills-directory>/j-space/scripts/verify_suite.py
   

   Replace <python-command> with the Python 3 command available on the host, commonly python, python3, or py -3.

5. Reload the host if it discovers Skills at startup.

The directory must remain intact because SKILL.md routes to relative paths under modules/, references/, and scripts/.

The repository-level LICENSE and THIRD_PARTY_NOTICES.md remain part of the distribution. Include copies of both when redistributing j-space/ as a standalone package.

Option B — ask an AI agent to install it

Copy the following prompt into an agent that can access files and this repository:

Install J-Space Cognition Suite from
https://github.com/Tiger3807861189/J-Space-Cognition-Suite-V3.6
into this environment's user-level Skills directory.

First inspect the host configuration or documentation to locate the correct Skills directory.
Install the complete j-space/ directory as j-space/, preserving SKILL.md, modules/, references/,
and scripts/. If a j-space target already exists, compare it and ask before replacing anything.
Run scripts/verify_suite.py with an available Python 3 interpreter after installation.

When finished, report the installed path and verification result, then tell me how this host
invokes the Skill. Briefly explain fast, full, and loop, and explain that the optional controller
records long-task state rather than choosing solutions. If this host has no native Skill loader,
explain the selective system/developer-instruction integration instead of reporting an
installation.

Use it

Invoke the Skill through the mechanism provided by your host—such as its Skill picker, /j-space, $j-space, or a direct request:

Use j-space for this task. Audit this repository, preserve its architecture,
verify every finding, and keep the work consistent across all affected files.

The entry gate selects the lightest suitable pass automatically.

Operating modes

Pass	Suitable work	What loads
fast	One step, or a result checkable in one glance	Nothing extra
full	Several dependent steps and one bounded deliverable	One or two relevant modules; ship before delivery
loop	Multiple stages, files, turns, tools, or persistent state	Ledger, seams, checkpoints, register audit, and recovery

A request for brevity changes the outer response length while verification remains aligned with the task's floor. Short work stays light; long work receives durable state only when it needs it.

Core mechanisms

Mechanism	Function
Selective workspace loading	Keeps one or two load-bearing ideas active and externalizes the rest
Broadcast hub	Gives dependent branches one shared source for names, values, constraints, and style anchors
Dense Track	Carries long internal chains in compact, decodable notation before returning to clean outer language
Bridge-before-conclusion reasoning	Makes required intermediates explicit before a conclusion consumes them
Metacognitive control	Routes confidence, inconsistency, and failure signals into a concrete next action
Empirical escape and verification	Converts stalled derivation into bounded tests with a named verifier and coverage
First-person agency and functional echo	Uses I, we, let's, and we need to bind workspace state to later actions and checks

The mechanisms are selectively loaded. They are not a fixed checklist for every request.

Optional controller

j-space/scripts/jspace.py externalizes loop state into .jspace/ in the current task workspace. Invoke it by its resolved Skill path while keeping the task workspace as the current directory.

Command	Purpose
note --goal "..." --next "..."	Open the ledger and define done plus the first action
note --next "..."	Replace the single next action after a checkpoint or seam
note --core "..."	Record a hub entry
note --core "..." --core-slot 1	Swap a selected live hub entry
note --check "..." --by "..."	Append a checkpoint with verifier and coverage
note --open "..." --settled-by "..."	Record a question and what would settle it
note --close N --check "..." --by "..."	Close question N against a new recorded checkpoint
seam	Re-read current state and report recent movement
ship FILE	Inspect outgoing text for register leakage and failure signatures
resume	Reload the premise, invariants, and full ledger after a long gap

<python-command> <skill-root>/scripts/jspace.py note --goal "what done means" --next "first action"
<python-command> <skill-root>/scripts/jspace.py note --close 1 --check "what now holds" --by "verifier and coverage"
<python-command> <skill-root>/scripts/jspace.py seam
<python-command> <skill-root>/scripts/jspace.py ship OUTPUT_FILE
<python-command> <skill-root>/scripts/jspace.py resume

The controller records and reports state. Solution choice remains with the model. It uses the Python standard library and writes working state only under the task's .jspace/ directory.

Generic model integration

An environment with a native Skill loader can install j-space/ directly. For a chat or API environment, provide j-space/SKILL.md as a system- or developer-level instruction and expose modules/ and references/ through file or retrieval tools.

Selected files should be retrieved on demand. Selective loading is part of the operating design.

Benchmarks

All values use the native score of the corresponding benchmark; higher is better. — means that no result is reported. HLE is separated into no-tool and tool-enabled conditions.

Evaluation context

The J-Space evaluations on DeepSeek were configured with reference to the official DeepSeek Harness minimal-mode setup, with max reasoning effort, temperature = 1.0, and top_p = 0.95. J-Space participated across the inference-time workflow through workspace routing, state continuity, verification, and recovery.

Results were collected within the project's available evaluation environment. Hardware conditions, process isolation, tool availability, and information-access boundaries form part of that context. J-Space tends to encourage more initiative and goal-directed exploration, making accessible artifacts and execution traces relevant to observed outcomes.

The table presents project-level benchmark records under these conditions. Comparator values retain the evaluation contexts published by their respective providers, and score variation across environments and harness configurations is expected. Source records include the DeepSeek V4-Flash-0731 model card, Z.ai's GLM-5.3 release-evaluation record, the Kimi-K3 model card, and Anthropic's Claude Fable 5 & Claude Mythos 5 System Card, which also reports its named comparator conditions. The GLM record is identified at provider level because no stable model-card URL accompanies the source record used here.

Model comparison

Benchmark	DeepSeek V4-Flash-0731	DeepSeek V4-Flash-0731 + J-Space V3.6	GLM-5.3	Kimi-K3	Opus-4.8	Fable 5 (w/ fallback)
HLE (w/o tools)	37.8	45.5	—	43.5	49.8	53.3
HLE (w/ tools)	51.5	60.6	62.5	56.0	57.9	63.0
Terminal Bench 2.1	82.7	87.1	88.2	88.3	85.0	88.0
NL2Repo	54.2	70.2	58.0	58.0	69.7	—
CyberGym	76.7	81.7	84.5	80.0	78.3	83.1
DeepSWE	54.4	67.4	66.9	67.5	58.0	70.0
Toolathlon-Verified	70.3	77.7	73.0	76.5	76.2	77.9
Agents' Last Exam	25.2	30.1	28.5	27.6	25.7	23.8
AutomationBench (Public)	25.1	31.7	48.2	30.8	27.2	29.1

Efficiency

These task-level indices retain the same task and model conditions and each records one evaluation run. Control is the matched baseline; J-Space is the corresponding suite-assisted condition. Speed is benchmark score divided by elapsed time, where higher is better. Token cost is consumed tokens divided by benchmark score, where lower is better. Elapsed time and token count use fixed, uniform scaling coefficients across both conditions. The coefficients affect the displayed scale while the within-metric improvement ratio remains comparable.

Metric	Control	J-Space	Improvement
Speed (score/time; higher is better)	0.43	1.09	2.53×
Token cost (tokens/score; lower is better)	2.63	1.19	2.21×

Related evaluation material: DeepSeek V4 × J-Space Capability Realization Report.

Cross-model compatibility

The operating effects have been reproduced across the DeepSeek, Qwen, GLM, GPT, and Claude model families. Effect size varies with base capability, context policy, tool harness, sampling configuration, and benchmark implementation.

The portable unit is the protocol: workspace loading, selective routing, state externalization, verification, and recovery. It is independent of a vendor-specific tokenizer or model API.

Project structure

J-Space-Cognition-Suite-V3.6/
├── .github/workflows/verify.yml    # three-platform integrity and regression checks
├── CITATION.cff                    # machine-readable citation metadata
├── CONTRIBUTING.md                 # contribution and provenance requirements
├── LICENSE                         # Apache License 2.0
├── README.md                       # English engineering guide
├── README.zh-CN.md                 # Chinese engineering guide
├── THIRD_PARTY_NOTICES.md          # attribution and license boundaries for source material
├── tests/test_jspace.py            # standard-library controller regression tests
└── j-space/
    ├── SKILL.md                    # single entry, gate, routing, and invariants
    ├── modules/                    # nine selectively loaded protocols
    ├── references/                 # evidence, induction, and worked exemplars
    └── scripts/
        ├── jspace.py               # optional loop controller
        ├── workspace-ledger.md     # ledger template and contract
        └── verify_suite.py         # authoring-time integrity check

SKILL.md is the only registered entry. Modules and references are loaded on demand so the control system does not become its own source of context pressure.

Maintainers can verify the package from its root:

<python-command> j-space/scripts/verify_suite.py
<python-command> -m unittest discover -s tests -v

Technical basis and scope

J-Space uses the operational workspace terminology established by Anthropic's related interpretability research. Within this suite, first-person language is treated as control grammar: accessible state descriptions are bound to explicit actions, checks, and settles.

The suite focuses on observable functional properties—reportability, deliberate maintenance, intermediate computation, broadcast, monitoring, and causal sensitivity. Detailed research interpretation, terminology, evidence boundaries, and sources are maintained in j-space/references/j-space-science.md.

Design principle:

Dense on the inside, decodable on demand, clean on the outside.

Use only the machinery the task earns.

Release history

J-Space has progressed through:

V1 → V1.5 → V1.8 → V2 → V2.5 → V2.6 → V3 → V3.1 → V3.2 → V3.5 → V3.5Turbo → V3.6

The V3.6 package contains one entry, nine focused modules, three supporting references, an optional runtime controller, an authoring-time verifier, standard-library regression tests, three-platform CI, Apache-2.0 licensing, and machine-readable citation metadata.

Citation

If you use J-Space in research, please cite the accompanying paper when it becomes available. For engineering use, cite this repository:

Tiger3807861189. (2026). J-Space Cognition Suite V3.6 (Version 3.6). Zenodo. https://doi.org/10.5281/zenodo.21977271

@software{j-space-cognition-suite,
  author  = {Tiger3807861189},
  title   = {{J-Space} Cognition Suite V3.6},
  year    = {2026},
  version = {3.6},
  doi     = {10.5281/zenodo.21977271},
  url     = {https://github.com/Tiger3807861189/J-Space-Cognition-Suite-V3.6}
}

GitHub-compatible metadata is available in CITATION.cff. The version DOI above identifies an immutable Zenodo snapshot; the all-releases concept DOI is 10.5281/zenodo.21971181. A repository commit identifies the exact maintained file set between archival deposits, including its current licensing and third-party notices; the fixed snapshot is not a live mirror of those files.

License

J-Space Cognition Suite is released under the Apache License 2.0. It permits use, modification, redistribution, and commercial integration under its notice and patent terms. See LICENSE for the complete terms. Quoted or summarized external source material remains subject to its source terms and is identified in THIRD_PARTY_NOTICES.md. When redistributing only the runtime j-space/ directory, carry both root files with it.