Models that prove themselves
An agentic factory for training models, an eval gate that cannot be overridden, and a signed audit log that makes every claim checkable.
AINFERA RESEARCH·H. RAZ·SIGNED AUDIT LOG · EVENT #000198Agents are becoming the main consumers of models, and agents cannot read marketing. We describe the Neptune factory: a training line whose stages — SFT, DPO, GRPO, MTP, QUANT — are separated by a single eval gate with no override path. Every gate verdict is appended to a signed audit log with an event number; eval certificates are rendered from those entries; and every model's name carries its true parameter count. The result is a release process in which competence is demonstrated, not claimed.
Agents buy models
When a human picks a model, a launch post and a demo can close the sale. An agent selecting a model for a task — or an engineer provisioning a fleet of them — needs something machine-checkable: what was evaluated, on which suite version, with what result, and whether anyone could have overridden the verdict. Today's answer is usually a README. Ours is a signed record.
One line, six stages
Neptune models are produced, not crafted. A checkpoint moves through supervised fine-tuning, preference optimization, and reinforcement over full agent episodes; after the gate, it picks up multi-token prediction and quantized variants. The line is the same for every model — flagship, compact, or domain-tuned.
Two verdicts, no override
The gate runs the eval suites registered for a stage and issues exactly one of two verdicts. There is no promote-anyway flag: downstream stages can only consume checkpoints the gate has passed, the way a linker can only consume objects that compiled.
checkpoint → run(suites[stage]) → scores
scores → sign() → audit_log.append(event#)
verdict ∈ { PROMOTE, STOP } · no third value
A signed, append-only record
Each verdict is appended with an event number, the checkpoint hash, suite versions, and scores. Eval certificates — the documents that ship with releases — are rendered from these entries and cite their event #, so a certificate can always be traced back to the record that produced it.
The name is the spec
Neptune-1.0-27B-Agent has 27 billion parameters — true count, no rounding games. Mixture-of-experts releases will state both totals: parameters held and parameters active. A name an agent can parse is a spec it can trust.
Records, not promises
We publish no roadmap dates. A date is a promise about the future; the gate only issues records about the past. A model listed as proof-gated ships when its evals clear — and you will be able to read the entry that says so.
title = {Models That Prove Themselves},
author = {Raz, Hizrian},
institution = {Ainfera Inc.},
number = {WP-001}, year = {2026}, note = {v0.9 draft}
}