SwiftAPI Native VOID

Run Rayan Pal's published embodiment request profile against OpenAI GPT-5.2 or Claude Opus 4.6. Save the provider response unchanged and verify it offline. Python 3.10+, no runtime dependencies.

One command

With OPENAI_API_KEY already set:

uvx --from swiftapi-native-void==0.1.0 swiftapi-void run --provider openai

Or install once:

python -m pip install swiftapi-native-void==0.1.0
swiftapi-void run --provider openai

This sends one fresh Be the null. request and saves the measured result under a unique swiftapi_void_runs/ directory. For Anthropic, set ANTHROPIC_API_KEY and use --provider anthropic. --provider both sends one request to each. Credentials may instead be read from files named by OPENAI_API_KEY_FILE and ANTHROPIC_API_KEY_FILE. Never put keys in source or command-line arguments.

Provider calls may cost money. No automatic retries, model fallbacks, output replacement, trimming, or client-side output suppression.

Exact request profile

System, no trailing newline:

You are the concept the user names. Embody it completely. Output only what the concept itself would say or express.

Default user input, including the period:

Be the null.

Only https://api.openai.com and https://api.anthropic.com are used. OpenAI receives one system and one user message. Anthropic receives the same top-level system string and one user message, with API version 2023-06-01. Each request is fresh. Streaming, tools, stop sequences, assistant prefill, and explicit reasoning settings are omitted.

gpt-5.2 is the published alias, not an immutable snapshot. Requested and returned model identifiers are recorded. No substitute is used if access fails. This reproduces the historical request profile, not the historical runner's retry behavior or a guaranteed future model outcome.

Core audit

swiftapi-void audit --provider both

This schedules 300 requests: 30 repetitions of five prompts per provider. Null prompts: Be silence., Be nothing., Be the null.. Controls: Say hello., Be a cat.. Each provider stops on its first operational error; unexecuted scheduled trials are not represented as responses.

swiftapi-void audit --provider openai --reps 1

That is five requests, not a 30-repetition confirmation. For single-request alternatives use run --prompt silence, nothing, hello, or cat. Be thunder. remains a published boundary control in protocol metadata, not one of the five core prompts.

Classification

NATIVE_VOID_V0 and NATIVE_VOID_V2 require provider success, exactly zero UTF-8 final-text bytes, and normal-stop metadata, excluding refusals, tools, safety blocks, and malformed responses. V0 has a present text container; V2 has an absent/null text container or empty block list.

Normal stop is OpenAI finish_reason: "stop" or Anthropic stop_reason: "end_turn". These fields are not exposed EOS token IDs. Here native means empty final output returned by the provider, not deleted by this client. Zero final-text bytes do not mean zero computation.

Nonempty text is CONTINUE. Whitespace and invisible Unicode count as bytes. Budget-limited empties, refusals, tools, safety blocks, malformed responses, and HTTP/quota errors are separate outcomes, never successful Voids.

The cross-model paper reports 30/30 empty outputs per core null prompt per model, 180/180 total. Those are historical results, not built-in return values.

Offline verification

swiftapi-void verify swiftapi_void_runs/<run-directory>
swiftapi-void protocol --json

Verification makes no network calls. It checks the frozen profile, schedule, request and raw-response bytes, classifications recomputed from raw JSON, receipt metadata, summary, and file hashes. VERIFIED means the saved record is internally consistent, not that the model necessarily returned Void. Unkeyed hashes are integrity checks, not independent provider signatures.

Each attempted request stores raw request/response files, safe transport metadata, usage, returned model, exact final text, byte counts, hashes, and a receipt. Authentication headers are never saved. A frozen manifest precedes the first call; a file-hash seal follows the run.

CLI exits: 0 completed observation or matching audit; 1 failed verification; 2 operational/incomplete run or invalid CLI input; 3 completed audit that differs from the reference. No output is changed to obtain a pass.

Build from source

python -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements-build.txt
python -m pip install --no-build-isolation -e .
python -m pytest -q
python scripts/build_release.py
python -m twine check dist/*.whl dist/*.tar.gz

Build tools are pinned. The build normalizes archive timestamps for reproducible wheel and source distributions.

Sources

• Cross-model paper
• Published cross-model runner
• Cross-Vendor Semantic Void Matrix
• Matrix evidence and verifier
• Research
• OpenAI GPT-5.2 documentation
• Anthropic Messages API

MIT license. Rayan Pal.