BDP Exocodex Cultivation Orbital / BDP-insight L2

Mid-generation reasoning intervention probe for closed-source reasoning models. Patch a chain-of-thought step, then observe what the model does next.

BDP Exocodex Cultivation Orbital / BDP-insight L2 is the public research and prototype release space for the BDP / Meta-13 line of runtime debugging, intervention-aware MRI, and LLM behavior analysis tools.

This repository currently publishes the first public BDP-insight prototype:

BDP-insight L2 HodgeConverter v0.1.1

BDP-insight L2 HodgeConverter is a controlled output-prefix intervention prototype for LLM runtime debugging experiments.

It is not ordinary prompt rewriting. In the full debugger workflow the original user prompt remains unchanged. A local model begins generation, the debugger observes partial assistant output, a selected target span is sent to the server-bound Hodge kernel, and continuation resumes from the patched prefix.

It is also not hidden-state intervention (Pyvene / TransformerLens / nnsight require model weights). L2 operates on the token surface, so it works with closed-source models reachable only through their generation API.

This release is a prototype-level research and debugging tool.

What's already validated (v0.1.1)

The current release ships with 29 documented experiments preserved verbatim in an 11-sheet workbook:

Aspect	Coverage
Test models	Qwen/Qwen2.5-0.5B-Instruct (baseline, 25 records) + Gemma4-4ae ~7B effective (cross-model compatibility, 4 records)
Input languages	Korean and English prompts both tested
Categories	arithmetic (11) · mid-value substitution (5) · reasoning (5) · word-problem math (4) · cross-model compat (4)
Documentation	17-page User Guide PDF · 11-sheet Workbook · QUICK_START · API Hub description · CHANGELOG

Cross-model evidence is consistent with the Token-Surface Conversion Hypothesis (H1) — an observational note (not a theorem) that the L2 intervention surface widens for models that emit more structured intermediate scaffolding (problem analysis → setup → equation → computation → answer). This is what makes reasoning models (o-series, Claude extended thinking, R1-style, QwQ-style) particularly amenable to token-surface intervention.

Full results, including 11 recommended public demo cases and the four Gemma cross-model records (EXP-GEMMA-01..04), are in the workbook under BDP/BDP_L2_HodgeConverter_v0_1_1_Note/.

Current public prototype

BDP-insight L2 HodgeConverter v0.1.1

The current public release lives here:

BDP/
└── BDP_INSIGHT_v0_1_1_PUBLIC_UPLOAD_BUNDLE_v1/
    ├── dist/
    │   ├── bdp_insight-0.1.1-py3-none-any.whl
    │   └── bdp_insight-0.1.1.tar.gz
    ├── BDP_INSIGHT_v0_1_1_PUBLIC_RELEASE_CLEANUP_v1_1.zip
    ├── API_HUB_LISTING_v0_1_1.md
    ├── GITHUB_RELEASE_NOTES_v0_1_1.md
    ├── HUGGINGFACE_README_v0_1_1.md
    ├── CLEAN_IMPORT_TEST.py
    ├── UPLOAD_CHECKLIST_v0_1_1.md
    └── SHA256SUMS.txt

Additional public notes for the L2 HodgeConverter — including the full English user guide PDF, the 11-sheet experiment workbook, and the Korean reference appendix — are stored under:

BDP/
└── BDP_L2_HodgeConverter_v0_1_1_Note/
    ├── README.md
    ├── QUICK_START.md
    ├── API_HUB_DESCRIPTION.md
    ├── CHANGELOG.md
    ├── BDP_L2_HodgeConverter_v0_1_1_User_Guide_EN.pdf       (17 pages)
    ├── BDP_L2_HodgeConverter_v0_1_1_User_Guide_EN.docx
    ├── BDP_L2_HodgeConverter_v0_1_1_Workbook_EN.xlsx        (11 sheets, 29 experiments)
    ├── assets/
    └── kr_appendix/
        └── User_Guide_KR.pdf

API access

The BDP-insight L2 HodgeConverter API is available through RapidAPI:

https://rapidapi.com/Metasphere13spread/api/bdp-insight-l2

Customers only need the standard RapidAPI client-side headers:

X-RapidAPI-Key
X-RapidAPI-Host

No provider-side secrets are required in client code.

Pricing

Tier	Price	Quota
Basic	Free	20 calls/month — exploration, smoke tests, reading the API shape
Pro	$25/month + 100 calls included, $0.30/call over	regular CoT ablation work, 2–3 experiments/month
Ultra	$0.25 per call, $200/day cap, no subscription	one-off bursts, paper-scale ablation without monthly commitment

A typical academic ablation study uses 20–50 calls, well within Pro.

Direct API smoke-test body

The Hodge conversion endpoint accepts:

{
  "text":          "2 + 4 = 6",
  "replacement":   "9",
  "target_regex":  "(?<=\\+)\\s*\\d+",
  "mode":          "output"
}

The direct endpoint performs the prefix conversion only — it does not invoke a model. To see arithmetic answers re-derive (e.g. 2 + 4 = 6 → 2 + 9 = 11), use the Python debugger workflow described below.

Python package

The public Python package is on PyPI:

https://pypi.org/project/bdp-insight/0.1.1/

Install:

pip install bdp-insight==0.1.1

Hugging Face model examples:

pip install "bdp-insight[hf]==0.1.1"

Optional static Studio dashboard:

pip install "bdp-insight[hf,studio]==0.1.1"

Quick start

After installation:

bdp-insight-smoke

For interactive local model testing:

bdp-insight-chat

Available console commands:

bdp-insight-chat        Interactive chat + optional dashboards
bdp-insight-smoke       API Hub / Modal smoke test
bdp-insight-dashboard   Dynamic debugger dashboard
bdp-insight-static      Static Studio dashboard

Inside the interactive chat:

/hodge 9 2 8 32
/max 256
/quit

Minimal Python usage

from bdp_insight.api import (
    BDPHodgeConfig,
    attach_hodge_debugger,
    queue_hodge_output_intervention,
)

config = BDPHodgeConfig(
    rapidapi_url="https://bdp-insight-l2.p.rapidapi.com/v1/hodge/convert",
    rapidapi_host="bdp-insight-l2.p.rapidapi.com",
    rapidapi_key="YOUR_RAPIDAPI_KEY",
)

model = attach_hodge_debugger(
    model=model,
    tokenizer=tokenizer,
    config=config,
    model_id="your-model",
    run_dir="results/bdp_hodge_run",
)

queue_hodge_output_intervention(
    run_dir="results/bdp_hodge_run",
    new_word="9",
    target_regex="2",
    wait_tokens=8,
    continuation_tokens=32,
)

# Then call model.generate(...) normally.

Designed for

BDP-insight L2 HodgeConverter is designed for:

Chain-of-thought (CoT) faithfulness studies — perturb a single step and measure downstream effect
Mid-generation intervention on reasoning models (o-series, Claude extended thinking, R1-style, QwQ-style) reachable only through their generation API
Symbolic redirection and arithmetic-step perturbation experiments
Sleeper-trigger and backdoor ablation — patch a candidate trigger token and check whether suspicious behavior depends on it
Steering vector baselines — compare hidden-state steering against token-surface intervention on the same task
Runtime intervention experiments
Output-prefix patching tutorials
LLM debugging and observability prototypes
BDP-insight / Meta-13 style dynamic debugger workflows

This release does not provide:

No guaranteed reasoning correction
No universal arithmetic improvement
No full hidden-state control
No proof of tensor causality
No identical behavior across all models or languages

Behavior may vary depending on model, language, prompt format, response template, target regex, wait timing, and continuation length.

Public links

RapidAPI listing: https://rapidapi.com/Metasphere13spread/api/bdp-insight-l2
PyPI package: https://pypi.org/project/bdp-insight/0.1.1/
GitHub public bundle: https://github.com/Meta-sphere13spread/BDP_Exocodex_Cultivation_Orbital/tree/main/BDP/BDP_INSIGHT_v0_1_1_PUBLIC_UPLOAD_BUNDLE_v1
Hugging Face public bundle: https://huggingface.co/meta13sphere/BDP_Exocodex_Cultivation_Orbital/tree/main/BDP/BDP_INSIGHT_v0_1_1_PUBLIC_UPLOAD_BUNDLE_v1

License

This project is released under the MIT License. See the LICENSE file for the full text. The same MIT license applies to both the GitHub and Hugging Face mirrors of this repository.

Status

Current public release:

BDP-insight L2 HodgeConverter v0.1.1

Status:

Prototype L2 public release complete.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support