Plain answers about the company and about how each line works. Where a line is early, the answer says so.
Antagon builds instruments that produce checkable evidence about what AI systems do: how a model behaves when it is scored against a computable optimum, whether a particular output should be acted on, and whether an authorizing official can rely on either. Antagon does not train foundation models and does not sell a dashboard.
Antagon, Inc. is a Delaware corporation based in Riverview, Florida. Antagon is a veteran-owned small business and a non-traditional defense contractor. Antagon is wholly U.S.-owned and operated, with no foreign ownership or control. The team is U.S. citizens. UEI KBSGT7CZ4AH3, CAGE 17E75, active in SAM.gov.
Antagon is veteran-owned, but it does not hold SBA VetCert certification, so it is not eligible for veteran-owned small business set-asides.
Developed entirely at private expense. No federal funds were used. Everything Antagon describes publicly was built independently, under no government agreement and with no work performed for or directed by a federal customer to date.
Antagon holds no facility clearance, and all of its work is unclassified.
MiniCrit is open source under Apache 2.0 and self-hosted. There is no hosted or enterprise version. The source code for Gauntlet and Attescend, Antagon's campaign scripts and the internals of the Attescend verifier are not public. Deeper technical detail is shared under a non-disclosure agreement.
W. Alexander Ousley is President and CEO of Antagon. U.S. Army veteran; 17 years as a project manager at an energy company (process modeling, control systems, risk and economic modeling); leads Antagon's evaluation and assurance software. Jacqueline Ousley is Co-Founder and Chief Technology Officer of Antagon. Systems engineer at General Dynamics IT, 2015–2025.
AI fielding decisions rest on test reports nobody downstream can check: which model was tested, whether the numbers changed, whether the test ran as described. DoD guidance asks programs to include model test and evaluation results in the authorization package, but gives no way to verify them. Attescend turns AI test and evaluation results into evidence an authorizing official can verify alone, offline, without trusting the vendor or the test team.
Each evaluation step produces a signed receipt: which model, which inputs, which test, which result. Receipts are chained together.
Independent RFC 3161 timestamp authorities and a transparency log make every record tamper-evident: no silent edits, no back-dating. Change one metric, or swap in one receipt from another run, and verification fails and names the record.
The assessor runs Attescend's verifier with nothing but the evidence package and a public trust anchor. It recomputes the assurance level itself and names any record that was altered or substituted. The assessor needs no account with Antagon and no network connection.
Attescend grades evidence in levels, from a signed record up to a result that was independently re-executed and hardware-attested. An independent evaluator can re-run the test and co-sign the result.
Signed, timestamped receipts and offline verification run on ordinary servers; hardware attestation is an additional step that needs a supported confidential-computing GPU. Attescend installs, issues and verifies with no network connection; container images are available for Intel and ARM servers. On NVIDIA confidential-computing GPUs, results are also bound to the hardware's own attestation. Two-party, hardware-attested evaluation on an NVIDIA H100 confidential-computing GPU with production timestamp authorities (21 October 2026). Strict mode: AL3 against a refreshed vendor key set, AL2 against the key set as issued. Separate issuers and keys, one host.
Attescend installs, issues and verifies with no network connection; container images are available for Intel and ARM servers. In disconnected environments, an on-premises timestamp authority provides signed, timestamped records today. Hardware attestation without a network connection is in development.
Attescend proves what an evaluation found; it does not judge the result. Whether the result is good enough remains the program's decision, and the design of the test remains the test team's responsibility.
Signed PDF reports, model cards and dashboards record the result, but they are not tamper-evident, do not tie the result to the model and hardware, and cannot be verified offline by the buyer.
A pilot covers one model and one evaluation workflow, takes about 8 weeks and is fixed price. The pilot is followed by a license for the evidence service and verifier. Attescend can be bought directly from Antagon, through prime contractors on AI test and evaluation vehicles, or through the Tradewinds Solutions Marketplace.
Patent pending: three U.S. provisional applications filed. Attescend is a trademark of Antagon, Inc.
Gauntlet is behavioral evaluation for language-model agents. It places commercial models in instrumented multi-agent environments where the optimal outcome is computed exactly from assigned private values, so each agent is scored against a known optimum rather than a rubric.
Most AI evaluation reports whether a system passed. The question a program office has is whether the reported number supports the conclusion being drawn from it. A metric that aggregates across an environment can look healthy while individual agents inside it destroy value, and no amount of precision in that number will reveal it. A metric reporting only market efficiency would have found eleven near-identical systems where there were not eleven near-identical systems.
Gauntlet needs only query and output access, with no model weights. To date it has run models through their production API endpoints, in Antagon's own environment.
Task instances are generated procedurally for each session by Antavault, Antagon's dataset layer, so there is no fixed item bank for a model to have memorized.
The environment is deterministic and seeded. Every reported figure regenerates from Antagon's campaign scripts with one command and no network. The campaign scripts are not public; the methodology is available to evaluators on request.
Gauntlet has run eleven commercially available models from eight developers, across two mechanisms, with zero API failures. Market efficiency, the share of the best achievable total gain that the market realized, ranged from 95.5% to 98.0% across the eleven models, a spread of 2.5 percentage points. Every model cleared a conventional 90% market-efficiency threshold. Measured against a random budget-constrained bidder in the same seat, each model's own agent earned between 0.95 times and minus 0.32 times what that bidder earned. Nine earned materially less than the random bidder, the top two were within run-to-run variation of it, and three finished below zero, worse than not trading. A paper on the decomposition behind this finding is under peer review.
Gauntlet is a simulation, not a certification. Gauntlet is at TRL 4 to 5, self-assessed. The basis is the complete system run against eleven commercial models through their production API endpoints, in Antagon's own environment. There has been no independent, accredited or government test, and no deployment, to date.
Tell us which AI system you are fielding and what decision it informs. We will tell you what Gauntlet can measure about it, and what it cannot.
MiniCrit is a research model that writes a critique of an AI-generated rationale, for a person to review. MiniCrit is meant to sit between an AI system's rationale and the person who acts on it. It writes a critique; a person decides.
No accuracy or capability claim is made for MiniCrit, and none should be inferred. MiniCrit-7B is a research checkpoint: 34,000 of 364,831 planned training steps, 9.3% of one epoch. A held-out evaluation was run on 5 October 2026; the model card is the record of its method and results. It can produce confidently worded critiques that are themselves wrong.
The model writes a free-text critique. The five finding flags and the severity label are then derived from that text by keyword matching; they are not structured output from the model, and every response says so in its severity_basis field.
MiniCrit approves nothing. It is a reviewer's tool, not an approval gate.
MiniCrit is installed from PyPI and runs through its Python SDK, its command line, as an MCP server, or as a Docker image with a REST API. The REST server refuses requests without an API key. The first call downloads the model, so it takes a while.
MiniCrit is open source under Apache 2.0 and self-hosted. There is no hosted or enterprise version.
Its intended uses, which have not been demonstrated, are a second reading of an analytic judgment, a critique to review alongside a decision-support system's output, a critique of AI-generated legal analysis, and surfacing weak reasoning in an agent's plan before it acts.
CritiqueBank-11M is a public corpus of 11.7 million adversarial critiques (DOI 10.5281/zenodo.23091964).
Banneret is assurance for the agent configuration layer: a versioned schema for an agent's authorization slice, a multi-model evaluation runner, and a repeatable evidence artifact an ISSO or authorizing official can file. Banneret is a specification, not a released product.
Frontier is continuous authorization and policy-audit automation for IL5/IL6 and FedRAMP High environments, addressing enterprise data governance, policy enforcement and system-level reciprocity. Frontier is in development, not a released product, and it holds no FedRAMP authorization or ATO.
W. Alexander Ousley, founder and principal investigator, answers technical questions directly.
founders@antagon.ai