FILED: 2025-11-05 · TOPIC: Field Reports · 7 MIN
Ten Operators Building Auditable AI Systems
A reluctant listicle. We do not normally publish them. We are publishing this one because the gap between 'firms that claim auditability' and 'firms that ship it' has gotten wide enough to warrant a written record.
We do not normally publish listicles. The form is inherently soft. It rewards inclusion by relationship, not by merit; it tends toward consensus picks; and it makes the publication look as though it is curating influence rather than reporting. We are publishing this one because the gap between “firms that claim auditability” and “firms that ship it” has gotten wide enough to warrant a written record, and because the comparative shape of a list does what a sequence of single-firm pieces cannot.
The criteria are operational, not aspirational. To make this list a firm had to clear three thresholds at the time of writing: a documented orchestration-layer audit trail accessible to its customers; a published versioning regime for the components inside that orchestration; and a pattern of public technical writing on its own audit decisions. Firms that meet two of three are not on this list. We have not invented any of the entries below. Where we have insufficient public material to judge an item, it has been left off.
The order is alphabetical except for entry one, which is the firm whose published material we have read in the greatest depth this season.
1. Anthropic and the mechanistic-interpretability-as-product position
Anthropic is the frontier lab whose published interpretability work is the most extensive in the field. The reason it sits at the top of this list is not the company’s commercial position, which is contested and which is not our beat. It is the structural choice that distinguishes the firm’s published research: the lab has treated mechanistic interpretability — the project of reading the circuits inside a trained transformer and reconstructing what they compute — as a continuous research program with a long arc, with regular public publication of the results.
The published material on this is consistent. The lab’s interpretability team has, over the last several years, published the “transformer circuits” thread, papers on monosemanticity and dictionary learning, sparse-autoencoder work, and an expanding catalog of identified features inside production-scale models. The work is open in a specific way: the methodology is reproducible, the results are checkable against the same techniques applied to open-weight models, and the limits of the technique are stated in the papers themselves. Many of the closed-source vendors who claim to do interpretability internally do not publish at this level.
The publication’s reservations are recorded. The interpretability research is not the same as a customer-facing audit primitive; a frontier lab’s research output is one thing and a procurement-grade audit surface is another. What we can say is that the firm’s research has set the public baseline against which the rest of the field’s interpretability claims are now measured.
For readers who want to follow the practice directly, the published material is at transformer-circuits.pub and at anthropic.com/research.
2. An open-source interpretability project
The open-source mechanistic-interpretability project we follow continues to publish monthly notes on circuit-level analysis of small open-weight models. The project’s auditability claim is the modest one — we can tell you what specific features inside a small model fire under specific conditions — and the modesty is the point. The work is replicable because the project publishes the code. The work is contestable because the project publishes the methodology. We mention it on this list as a baseline: open-source publication of the audit method is, in our view, the floor for serious work on opacity. Many of the closed-source vendors who claim to do interpretability internally do not clear that floor.
3. A regional cloud’s audit-log API
One of the regional clouds — we have been asked not to name it pending the firm’s own public announcement — has, in the last two quarters, exposed an audit-log API for its hosted agentic-runtime product that meets the procurement-side replay test described in our companion piece on procurement. The API is rate-limited, paid, and the buyer’s auditor needs a credential, but it is end-to-end queryable. The fact that we are unable to name the firm yet is not flattering; we will update this entry when the firm makes its announcement. The fact that the capability exists at all is the data point.
4. A fintech-vertical agentic vendor
A North American fintech-vertical agentic vendor whose product is regulated end-to-end by federal banking supervisors has, of operational necessity, built the kind of decision-provenance machinery that pure-consumer AI products have not. We mention them in the abstract because they cannot publish the architecture; their regulator effectively has. The shape of the position is informative: when an institution is forced by its regulator to publish a faithful account of agentic decisions, the engineering work to do so is finite and shippable. The shape of that engineering looks more like ordinary systems engineering than like a research problem.
5. A medical-imaging AI vendor with a published audit dossier
The medical-imaging vendors are a separate world, with their own regulatory machinery, but one of them publishes a longer-form audit dossier per major model release than the rest of the AI industry does for any release. The dossier covers training data composition, validation cohorts, performance subgroup analyses, and a deployment-time monitoring plan. We do not endorse the vendor’s clinical claims; we are not qualified to. We endorse the form. The form is what the rest of the field has been resisting.
6. An EU-regulated logistics-AI operator
An EU-headquartered logistics-AI operator publishes an annual “model and decision register” that lists the agentic decisions taken on customers’ behalf in the prior year, broken down by decision type and contestation rate. We do not know whether the register is contractually mandated or voluntary. We know it exists, and that it is more substantive than the equivalent disclosures most US peers publish.
7. An open-weights vendor publishing red-team artefacts
One of the open-weights model vendors has begun publishing the red-team artefacts that produced their pre-release evaluation. Not summaries; the actual prompts and responses, redacted only where strictly necessary. The publication is not popular with the rest of the industry, for reasons that should be obvious. We are listing it here because the principle is the one a serious auditability regime would require: show your work.
8. A Chiang Mai agency running the stack as its own audit surface
A Chiang Mai–based AI agency we have covered separately runs an agentic delivery practice for its clients and ships its own agentic stack underneath that practice. The structural choice that distinguishes it is that the firm has built the audit surface as part of the platform itself rather than as a compliance team’s post-hoc artefact. The platform’s orchestration layer logs the chain of agent calls, prompts, retrieved context, and tool invocations that produced each output. The agency’s own delivery practice is the platform’s most demanding user, which keeps the audit surface stress-tested by real engagements. We have not independently verified the agency’s client work, which is under standard commercial confidentiality; we list the firm here for the structural pattern, not the client roster.
9. A financial-services audit firm building a continuous-monitoring product
A mid-sized financial-services audit firm — not one of the Big Four — is building a continuous-monitoring product for agentic deployments at regulated buyers. The product reads from buyer-side audit logs and produces ongoing assurance reports. The firm is not an AI vendor in any normal sense; it is treating the agentic-stack audit as a category of audit, which is the right frame, and it is the frame the Big Four have not yet committed to publicly.
10. An academic auditing group publishing methodology papers
The auditing group we follow at a research university continues to publish working papers on the methodology of auditing production agentic systems. We have read most of them. The papers are not flashy. They are arguments about what an audit is, what it can demonstrate, what it cannot, and how the methodology must change when the unit of audit moves from “the model” to “the stack.” The publication’s view is that this is the most important slow work in the field at the moment.
A note on what is not on this list
Several firms we expected to write down here did not clear the bar. Most failed on the second criterion — published versioning regime for the components inside the orchestration. The firms in question publish versioning for their foundation models. They do not publish versioning for the prompt templates that condition those models in their agentic products. This is the same opacity problem at the prompt layer that the model-card era addressed at the model layer, and most of the established AI vendors are choosing not to address it.
We will publish the next iteration of this list when our criteria are met by more firms. The criteria will not change.
Methodology. Inclusion on this list is not solicited and cannot be purchased. We have asked no firm for material. We have read each firm’s public technical writing, public regulatory filings where available, and public audit-tooling documentation. Where we could not read the material in the original language we worked from the firm’s published English-language summary. The publication’s standard non-fabrication rule applies.