Explore GRIP and the proposal for Agentic Language Models. Both preprints are free to read, cite and question. Published on Zenodo on 24 August 2026; not yet peer reviewed.
Preprint · ZenodoPaper 01
GRIP: The Governed Runtime Intent Protocol
Madhav D. Parashar · Arnav Parashar
24 Aug 2026 · Version 1.0-draft.2
Abstract
AI agents act: they call tools, spend money and message people. GRIP makes that work authorised before it runs, bounded at a gate that enforces the limits, and attributable afterwards.
1 · Introduction
DOI10.5281/zenodo.22078457
Preprint · ZenodoPaper 02
Agentic Language Models: A Proposal
Madhav D. Parashar · Arnav Parashar
24 Aug 2026 · 2026-08 proposal
Abstract
Proposes a class of model whose native interface is the rules of the runtime it serves: a checked package of context goes in, an act the rules can check comes out. And a public test.
1 · Introduction
DOI10.5281/zenodo.22078465
A drawing of the two filed papers, abstracts shortened · a DOI is a permanent address: it always leads to the filed record
Madhav D. Parashar and Arnav Parashar
AI agents now act. They call tools, spend money and message people. The protocols that connect them say what an agent can reach, but not who allowed the work, how far it may go, or who answers for it. GRIP proposes a small protocol that makes agent work authorised before it runs, bounded at a gate that enforces the limits, and attributable afterwards, whatever the model, framework or vendor.
Six signed objects on one chain. Every act leaves a receipt, linked to the one before it.
A gate before every act. Authority only narrows as it is handed down, and every refusal says why.
A person’s yes, tied to one act. An approval covers the act it names, and nothing else.
Four levels to adopt it. Start by keeping receipts, then grow into full, checked sign-off.
Inside the paper · six signed objects, one chain
Each object is a small, signed record. Together they answer who allowed the work, how far it could go, and who said done.
01
02
03
04
05
06
Madhav D. Parashar and Arnav Parashar
Most agent systems borrow their judgement from a general-purpose model, then add the rules from outside, as prompts and wrappers. This paper proposes a class of model whose native interface is the rules of the runtime it serves: a checked package of context goes in, an act the rules can check comes out. It also proposes a public test, with pass marks set before any model is measured, so that, once the test code is out, anyone can check whether a model belongs. It is the companion to the GRIP paper.
A definition. Membership criteria for the class, written so someone outside Foxora can test them.
An interface. A budgeted, checked package of context goes in. A typed GRIP act comes out. Gates outside the model hold its authority.
A public test. Pass marks fixed before any model is measured, ours included.
A worked example. Fennec appears as one. How it is trained is not disclosed.
Research is where Foxora is going. These are the rules we keep on the way there.
01
02
03
Foxora Studio keeps the same habits today. Agents ask before going past your brief, stop at your budget, and hand back work you can check. Nothing is done until you accept it.
Free plan · 200 credits a month · No card required
The decision point
Sample scenario
A clear receipt
Sample scenario