Ox Alpha Model Forensics

Independent model-forensics investigation into the possible identity of OpenRouter's stealth ox-alpha model.

Read the English report | Türkçe raporu oku | Review the findings

Important

This repository documents an independent attribution assessment. It is not an official confirmation by OpenRouter, Zhipu AI, or any other model provider.

Final assessment

The report assigns its highest confidence to:

• Likely developer: Zhipu AI (Z.ai)
• Likely model family: GLM-5.3
• Variant hypothesis: probable multimodal variant
• Assessment confidence: approximately 93%

The confidence figure is the report's evidence-aggregation result. It is not a benchmark accuracy score, a frequentist confidence interval, or a calibrated posterior validated against a large set of known stealth models.

Status and scope

This project studies attribution signals across multiple layers rather than treating any single behavioral resemblance as proof. The report describes:

1. Infrastructure forensics
2. Tokenizer fingerprinting
3. Vision and video behavior analysis
4. Knowledge-cutoff probing
5. Self-knowledge asymmetry testing
6. A/B behavioral comparison
7. Bayesian-style evidence aggregation

Evidence summary

Layer	Reported observation	Evidentiary role
Infrastructure	A Java service-layer trace containing com.wd.paas.api.domain.v4.chat.ChatCompletionRequest and error-code behavior	Narrows likely serving-stack relationships; not model-weight proof
Tokenizer	A reported 30/30 match against the GLM-5.3-family tokenizer probe set	High-value family fingerprint if independently reproduced
Vision/video	Similar video-token usage and encoder behavior to GLM-5V-Turbo	Supports a multimodal-family hypothesis
Behavioral probes	Sixteen probes covering reasoning, knowledge boundaries, self-knowledge, and response behavior	Supporting attribution evidence
A/B comparison	Output-style and behavioral similarities between ox-alpha and GLM-family reference models	Corroborative, not decisive on its own
Aggregation	An eight-candidate baseline updated across evidence layers to a reported ~93% assessment	Makes judgment assumptions explicit

The report's update chain is 12.5% -> 42% -> 68% -> 78% -> 83% -> 87% -> 88% -> 93%. See Research findings, the evidence framework, and the Bayesian-style analysis.

Research methodology

The investigation reports using:

• OpenCode
• LM Arena Agent Mode
• Codex
• Custom behavioral probes
• Session JSON analysis
• Automated analysis combined with manual verification and independent evidence review

The repository separates observations, interpretations, and conclusions. Reproduction claims should be evaluated only against the artifacts actually published here.

Repository contents

• reports/ - final English and Turkish reports
• methodology/ - research design, evidence framework, and confidence aggregation
• evidence/ - evidence categories and artifact manifests
• sessions/ - locations for sanitized sessions and analysis notes
• tools/ - analysis scripts and usage notes
• docs/ - timeline, FAQ, limitations, and source provenance
• CONTRIBUTING.md - evidence-submission and review requirements

Reproduction

1. Record the tested model identifier, provider route, timestamp, parameters, and raw response.
2. Run the same probe set against ox-alpha and declared reference models.
3. Preserve tokenizer counts, error payloads, modality metadata, and complete outputs.
4. Sanitize secrets and personal data without changing model-produced content.
5. Score each evidence item under the published framework.
6. Recompute the aggregation with documented assumptions and sensitivity checks.

Raw sessions are not considered available until files are present under sessions/sanitized-session-data/. Empty manifests must not be interpreted as evidence.

Report integrity

Both final reports are preserved byte-for-byte from the supplied source files. Their SHA-256 checksums, page counts, and visual-QA notes are recorded in reports/README.md.

Limitations

• No official confirmation from OpenRouter or Zhipu AI is documented in the reports.
• The exact model variant remains unresolved.
• Behavioral similarity is supporting evidence, not standalone proof.
• Shared infrastructure can identify a provider or serving path without identifying the underlying weights.
• A 30/30 probe match is meaningful only to the extent that probe selection, controls, and raw outputs are available for audit.
• The ~93% figure depends on priors, likelihood assignments, and evidence-independence assumptions.

See Limitations for the complete discussion and Sources for provenance.

Citation

Citation guidance is available in CITATION.md and machine-readable form in CITATION.cff.

License

Repository text and original code are released under the MIT License. Third-party material and the included reports may carry separate rights; consult the report files and source notices before redistribution.