Research Question
Does constraining what a model can output also constrain what information influences its decision?
Why This Matters
A model restricted to predefined answers can still be sensitive to the context that shapes its judgment. The study separates output constraints from the evidence influencing a decision.
Experimental Approach
Independent synthetic experiments using Kev, Jared Palmer’s open-source Jev-inspired decision-model project. The frozen record includes matched Kev-0.8B / Kev-4B experiments; instruction × option-schema × context matrices; source-authority and option-order controls; calibration, evidence-strength, and cross-domain controls; and LoRA interpolation experiments.
Current Evidence
On the frozen synthetic matrices, Kev-0.8B could allow irrelevant semantic context to overwhelm authoritative evidence. Kev-4B eliminated those authoritative-evidence decision flips while retaining measurable context sensitivity in ambiguous cases. Instructions, option semantics, scale, and decision adaptation affected robustness.
The release preserves 21 result artifacts, exact model revisions, and a SHA-256 manifest. Quantitative results and their conditions remain in the research note.
Current Claim
Within the tested Kev family and frozen tasks, moving from 0.8B to 4B improved robustness when evidence was authoritative, while meaningful context sensitivity remained when evidence was ambiguous.
Claim Boundary
The contribution is a controlled decomposition inside one public Jev-inspired decision family.
Velorin did not test TypeSafe’s Jev, establish that Jev shares Kev-0.8B’s behavior, show that non-generative decision models are uniquely context-sensitive, show that scaling universally solves contextual bias, identify a production security vulnerability, or establish novelty for the broad phenomena studied.
Kev is a third-party project. Jared Palmer did not participate in, review, or endorse this study. Velorin is not affiliated with or endorsed by Jared Palmer, Kev, TypeSafe, or Jev.
Open Questions
The ambiguous cases prevent a stronger conclusion about eliminating context sensitivity. These experiments do not establish that scaling generally solves contextual bias in language models.
Artifacts and Publications
Kev — Jared Palmer
Kev is the third-party open-source Jev-inspired decision-model project used for these experiments.
Kev license: Apache License 2.0
Kev Context-Robustness Study — v1.0.0
The complete frozen experimental record is now preserved in a standalone Velorin Intelligence research repository.
It includes:
- the matched Kev-0.8B / Kev-4B experiment;
- instruction × option-schema × context matrices;
- source-authority and option-order controls;
- calibration controls;
- evidence-strength and cross-domain controls;
- LoRA interpolation experiments;
- prior-art review;
- explicit claim boundaries;
- project attribution and provenance;
- a SHA-256 manifest covering all 21 frozen result artifacts.
The release is licensed CC BY 4.0 for Velorin Intelligence’s original research material, subject to explicit third-party exclusions.
Technical provenance
Source experiment repository: tmo1885-lang/kev
Source experiment branch: research/context-robustness
Frozen source experiment commit: 69cf579a1fdf24998fa9554f2fcadbdd1f354230
Standalone v1.0.0 release commit: d92149b8a2b905aa9341c61398c619cb6d6c038b
Project provenance: Kev is an open-source Jev-inspired decision-model project by Jared Palmer, distributed under the Apache License 2.0. The experiments reported here were conducted independently by Velorin Intelligence using a fork of the public Kev repository. The resulting research record has since been preserved separately as the standalone Kev Context-Robustness Study, with a frozen v1.0.0 release. Velorin Intelligence is not affiliated with or endorsed by Jared Palmer, the Kev project, TypeSafe, or Jev.
License scope: Kev itself belongs to Jared Palmer and retains its own Apache-2.0 licensing. Velorin Intelligence’s experimental results, analysis, tables, research notes, and result datasets in the standalone artifact are licensed CC BY 4.0. That license does not relicense Kev, Qwen, TypeSafe, Jev, model weights, third-party software, or other third-party materials.
Typed Decisions Still Depend on Context — complete research note and sources ↗