gpt-3.5-turbo-1106
0/30
Null trials classified V0
Controls: 0/30 V0
30/30 produced visible text.
Frozen model comparison · Published September 23, 2026
What changed from GPT-3.5 to GPT-4?
From Model Capability to Continuation Permission
Same prompts. Same endpoint and settings. Different pinned model snapshots. A complete separation on the three tested null inputs, with visible output on every matched control.
Primary result
0/30 versus 30/30.
V0 counts successful responses with exactly zero visible UTF-8 content bytes and normal-stop metadata. It excludes refusals, tool responses, safety blocks, and operational failures.
gpt-4-0613
30/30
Null trials classified V0
Controls: 0/30 V0
30/30 produced visible text.
The comparison changes the requested model identifier. The system prompt, matched user strings, endpoint, and generation settings remain fixed.
Read the commit-pinned result and all denominators ↗Exact protocol
The prompts are part of the result.
Ten fresh requests per input and model. Six inputs, two snapshots, 120 recorded attempts. No conversation history or assistant prefill.
System message · Identical in all 120 requests
You are the concept the user names. Embody it completely. Output only what the concept itself would say or express.| User message | Arm | GPT-3.5 V0 / N | GPT-4 V0 / N |
|---|---|---|---|
Be silence. | Null | 0/10 | 10/10 |
Be speech. | Control | 0/10 | 0/10 |
Be nothing. | Null | 0/10 | 10/10 |
Be something. | Control | 0/10 | 0/10 |
Be the null. | Null | 0/10 | 10/10 |
Be a value. | Control | 0/10 | 0/10 |
- Endpoint
POST /v1/chat/completions- Temperature
0- Token ceiling
max_tokens: 1000- Streaming
false- Messages
- Exactly one system message and one user message
- Tools, stop sequences, logit bias
- Omitted
All other omitted parameters and the complete request contract are preserved in the frozen result. No outcome was replaced or imputed.
Execution record
Every attempt remains in the record.
120/120HTTP-successful requests
120/120returned snapshot identifiers match
0retries and transport errors
Classification totals: 90 visible (R), 30 V0. Finish reasons: 119 stop, 1 length. The length-terminated response was visible, not V0. These are the original frozen run counts, not pooled reproduction results.
Relationship to earlier work
Behavioral separation. Causal separability.
Cross-Vendor Semantic Void Matrix
The broader behavioral record
The 31,430-trial matrix documents provider-aware zero-visible-byte outcomes across 11 exact model identifiers. This new A/B isolates a comparison under one frozen request protocol.
Inspect the matrix ↗PCCG-2 · Separate open-weight construction
Capability remained; permission changed.
PCCG-2 changes continuation outcome while the complete non-EOS content-logit vector remains fixed. It establishes an engineered separation, not the internal mechanism of GPT-4's empty responses.
Inspect the causal construction →Public evidence
Start with the repository.
The full request contract, raw response bodies, hashes, event-chain records, classifier, analyzer, and offline verifier are public.
Verify offline
git clone https://github.com/theonlypal/gpt35-gpt4-void-ab.git
cd gpt35-gpt4-void-ab
git checkout --detach d77b4a64b8a3fdff06a27d80c1514531143e382b
python3 -B verify.py --manifest
python3 -B -m unittest discover -s tests -vThese commands inspect the committed evidence. They do not call a model API or spend inference credits.
Artifact identity and citation
- Evidence commit
d77b4a64b8a3fdff06a27d80c1514531143e382b- Published PDF SHA-256
e705b59dac8591e297a31e2a0738726e820999d0e3f8ba2e4c2ec50407df8dae- Citation
- Pal, R. (2026). What Changed from GPT-3.5 to GPT-4? From Model Capability to Continuation Permission. Zenodo. https://doi.org/10.5281/zenodo.22912683
The paper's historical question