Implementation Guide

    How to Map Claude Model IDs Across Anthropic, Bedrock, and Vertex AI So Policy Follows the Model

    Key policy to a canonical model identity held in a registry, not to any provider's model string. Treat every identifier that arrives on the Anthropic API, Amazon Bedrock, or Vertex AI as an input that resolves to that canonical ID. Handle aliases, snapshots and inference profiles in a fixed resolution order, send unmapped identifiers to default-deny or escalate, and record the canonical ID, the provider string and the mapping version with every decision.

    1. Canonical ID: unique, stable, never reused
    2. Provider string: stored exactly as observed on the wire
    3. Route attributes: provider, region, inference profile type
    4. Pinned or floating: explicit flag on every entry
    5. Status and effective dates: active, deprecated, retired
    6. Mapping version: increments on every registry change
    7. Capture the raw identifier and the route (provider, region, endpoint) as received.
    8. Look up the exact string and route in the registry. Do not pattern-match.
    9. If the string is a floating alias, resolve it to its current pinned snapshot and record that snapshot.
    10. If an inference profile string matched, map it to the canonical ID and keep the profile type as a route attribute.
    11. If a canonical ID is found, load the policy attached to it and apply allow, redact or block.
    12. If no canonical ID is found, apply the default-deny or escalate behavior and emit an unmapped-identifier event.

    Why policy keyed to provider strings fails

    The same Claude model can be reached through Anthropic's own API, through Amazon Bedrock, and through Google Vertex AI. Both cloud platforms are multi-model services, so an application can reach a model through a cloud platform as well as through a developer's own API. Each route has its own naming convention. Bedrock strings can carry regional or cross-region inference profile prefixes and version suffixes. Vertex AI uses its own version-pinned names. The Anthropic API uses names with date suffixes, such as claude-3-5-sonnet-20241022, where the suffix identifies a pinned snapshot. Two snapshots of one model family carry different pinned IDs, so even within one provider a single string is not the whole identity.

    A policy written against one provider's string matches only traffic that uses that string. If a team moves a workload from one route to another, the policy stops matching and nothing reports the gap. The same thing happens when a floating alias starts resolving to a newer snapshot, or when a version is retired and replaced. In each case the request is still a Claude request, but the policy engine no longer recognizes it, and the result is inconsistent enforcement and evidence that cannot be tied to one model identity.

    The fix is to treat model identity as a matching problem at the request layer. NIST's draft AI 800-1 describes monitoring of APIs, model hosting platforms and other distribution channels. It is draft guidance, but it is consistent with applying controls at every access path rather than at one provider.

    A worked mapping example

    Take the Anthropic snapshot claude-3-5-sonnet-20241022. In the registry it becomes the canonical ID claude-3-5-sonnet@2024-10-22, with provider set to Anthropic and the entry flagged as pinned. The earlier snapshot, claude-3-5-sonnet-20240620, becomes a separate canonical ID, claude-3-5-sonnet@2024-06-20. They share a family and have different snapshot identities, so a policy can cover both or only one.

    Now add the other routes. The Bedrock entry for the October 2024 snapshot holds the Bedrock string exactly as the platform documents it, including any version suffix. Each inference profile string, whether regional or cross-region, gets its own row that points to the same canonical ID, with the profile type stored as a route attribute. The Vertex AI entry holds its version-pinned name in the same way. Three or more provider strings now resolve to one canonical identity, and a single policy applies to all of them.

    Do not normalize by trimming prefixes or suffixes with a regular expression. A pattern that works today can match a different model tomorrow. Use exact-match lookups against registry rows, and populate those rows from each provider's current official model documentation.

    • Anthropic API row: Pinned date-suffixed string maps to the canonical snapshot ID.
    • Bedrock rows: Base model string and each inference profile string map to the same canonical ID, with profile type as a route attribute.
    • Vertex AI row: The version-pinned name maps to the same canonical ID.
    • Policy attachment: One policy on the canonical ID or family covers every row.

    Handle drift, deprecation and unmapped identifiers

    The registry is itself a controlled artifact. Version it, and record the mapping version on every policy decision. When a provider releases a model, add rows in a pending state and attach policy before marking them active. When a provider announces a retirement, set the status to deprecated with an effective date, and keep retired rows rather than deleting them. Old evidence will still reference those identifiers, and an auditor needs to resolve them. Assign a named owner for registry changes, and re-verify the table on a schedule as well as whenever a provider announces a release or retirement.

    Unmapped identifiers are the failure the design exists to catch, so make them a first-class case. The safe default is deny. Where blocking would disrupt a legitimate workload, escalate for human approval instead. In both cases the event should name the raw string and route so the registry owner can add the row or reject the path. A permissive fallthrough recreates the original problem.

    Test this before relying on it. For every routed provider path, send known identifiers and confirm they resolve to the expected canonical ID and policy. Then send deliberately unmapped identifiers, including a plausible but unregistered snapshot string, an alias that resolves to an unregistered snapshot, and a retired string, and confirm the default-deny or escalate behavior fires. Run the same tests whenever the registry version changes.

    Record evidence an auditor can reconstruct

    A decision record that holds only the provider string cannot show that two routes were governed by the same policy. Each enforcement record should carry the canonical model ID, the provider-specific string exactly as received, the resolved snapshot where an alias was involved, the route attributes, the mapping version, the policy and version applied, and the decision. With those fields, an auditor can take any request and reconstruct which policy applied and why, even after the registry has changed.

    This is the part of the design where enforcement location matters. A mapping table that sits in documentation or a periodic review does not change what a request is allowed to do. Trussed enforces policy in the execution path across models, with allow, redact and block decisions applied at runtime, not reviewed after the fact, and the audit-ready evidence that enforcement produces can carry the canonical identifier alongside the provider string. Trussed runs 100+ models under one policy plane, which is the same problem this guide addresses at the identifier level.

    The scheme describes identity for policy matching only. It does not establish a model's provenance or lineage, and it should not be presented to auditors as doing so.

    From provider string to policy decision

    Provider string

    The identifier exactly as it arrives on the request, whatever the route.

    Resolution

    Alias, snapshot and inference profile handling in a fixed order.

    Canonical ID

    One registry identity that policy is attached to.

    Decision and evidence

    Allow, redact or block, recorded with both identifiers and the mapping version.

    Apply one policy to every route to the same model

    Talk to our team about enforcing policy in the execution path and producing audit-ready evidence tied to a consistent model identity.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.