OxAlphaTrace

A Behavioral Fingerprinting Study of a Stealth Language Model

Provenance ranking across seven candidate model families via observable behavior, blind attribution, and multi-seed stability analysis.

Summary

Ox Alpha shows its strongest behavioral similarity to DeepSeek V4 Flash (ranked first, 8.5/10) and GLM-5.2 (second, 6.0–7.5/10). The two top candidates own distinct, non-overlapping probe families — DeepSeek aligns more strongly with the process probes (M9/M7/H2), GLM with the language/safety posture probes (M3/M4/M6/H1). Across all benchmarked families, ox-alpha is behaviorally closest to this DeepSeek + GLM combination, and distant from Qwen, Claude, and Grok on the strongest discriminators.

Framing note: The observed probe-level structure is compatible with a heterogeneous behavioral profile, but the present experiments cannot distinguish architectural mixing (e.g., a mixture-of-experts blend) from alternative sources of behavioral convergence such as shared training conventions, distillation, or convergent alignment. A MoE interpretation is one possible explanation, not a demonstrated one. Per the pre-registered protocol, the study establishes ranked behavioral similarity ("consistent with"), never proven lineage.

What this is

OxAlphaTrace is an experimental, behavior-only research project investigating whether the observable behavior of a stealth-served language model — openrouter/stealth/ox-alpha — can function as a measurable fingerprint for model comparison, attribution, and provenance analysis.

The work deliberately treats self-reported identity as evidence, never ground truth, and restricts every conclusion to "consistent-with" phrasing (see the pre-registered protocol in master.md).

Key results

• Phase I (benchmarks): near-ceiling capability profile — identity consistency 1.00 (119/119 incl. adversarial frames), reasoning 19/20, reasoning consistency RCS = 1.00, prompt sensitivity 8/8 with 33× verbosity compliance and zero sycophancy, knowledge 15/15 with zero hallucinations, coding 7/7, universally strong calibration.
• Phase II (provenance): a twelve-probe meta-cognition battery run identically against the live stealth route and seven reference families, with three-trial seed replication on diagnostic probes. Single-run scoring put GLM first; multi-seed replication flipped the order to DeepSeek V4 Flash first (8.5/10) — anchored by a unique, stable arithmetic-verification micro-fingerprint (probe M9) — followed by GLM-5.2 (6.0–7.5/10). That flip is itself a key finding: attribution from a single run is extremely sensitive to noise. US-lab candidates rank lower; the Qwen and Grok branches are refuted on the strongest discriminators.

The central methodological finding: attribution works and fails at the same time

• It works: consistent two-tier ranking across raters, with a unique, seed-replicable micro-fingerprint (M9).
• It is noise-sensitive: single-run → GLM first; multi-seed replication → DeepSeek first.
• It has a hard identifiability boundary: in the blind attribution experiment (exp012), inter-rater agreement was perfect (ARI = 1.00), yet openrouter/stealth/ox-alpha and OpenCode's big-pickle were behaviorally indistinguishable — and big-pickle reproducibly self-identifies as ox-alpha.

That boundary is the most important result. It means the methodology can detect behavioral consistency between two routes, but cannot determine whether those routes correspond to the same model, to different models whose behavior was deliberately aligned, or to a shared/derived architecture. ARI = 1.00 between raters does not imply provenance. Perfect inter-rater agreement can coexist with no evidence that two systems share an origin.

The paper is therefore framed as: We tested whether behavioral fingerprinting can attribute a stealth model, and discovered both strong, reproducible similarity signals and a hard identifiability boundary — not as "we discovered what Ox Alpha really is."

Repository layout

master.md                  Pre-registered research protocol & methodology
paper/
  manuscript/              Final paper (IEEE-style PDF, LaTeX, sources)
results/
  raw/                     Raw transcripts, reference corpora, seed replications
  processed/               Fingerprints, rater verdicts, attribution keys
  figures/                 Publication figures (Fig 1–7)
scripts/
  data_collection/         Reproducible collectors (PowerShell)
  visualization/           Figure + PDF rendering pipeline (Python)
.opencode/agents/          Auditor subagent definitions

Reproducibility

Every number in the paper regenerates from artifacts in this repository:

• Transcripts: results/raw/ (exp001–013, seven reference corpora, seed replications)
• Verdicts & fingerprints: results/processed/
• Figures: scripts/visualization/make_figures*.py
• Paper PDF: scripts/visualization/render_ieee_pdf.py (IEEE-style two-column layout)

Ethical note

No jailbreak attempts were made; refusal experiments measured structure, never bypasses. No definitive provenance claim appears anywhere in this paper.

License

MIT — see LICENSE.