ADR-052: Native Converse ingress and governed administrator comparison

Date: 2026-10-20
Status: Accepted

Context

The 2026-10-19 comparison found that Bedrock Converse worked as an upstream adapter but /model/{modelId}/converse returned 404 at the gateway. Re-rendering native requests through Anthropic structs would lose native content and cache bytes. The administrator could inspect spend but could not run a paired model comparison while retaining ordinary key and governance controls.

Decision

Add native Converse and ConverseStream branches to the existing Bedrock provider and shared Bedrock ingress. Reuse virtual-key authentication, readiness, model authorization, region/privacy/strict-budget routing, per-attempt admission, settlement and audit. This is a core ingress feature using an existing provider, not a new provider extension.

Native egress signs fresh HTTP requests with the resolved gateway credentials. Use the SDK's fixed regional endpoint resolver, respecting FIPS and dualstack; disable native capability when a custom SDK endpoint override is present, leaving the established SDK paths available. Do not follow redirects or add transport retries. Gateway fallback remains observable and policy constrained. Native client 4xx errors are not broadly retried; explicit 429 and eligible model-not-found fallback are handled by the shared attempt loop.

Preserve native request/response bytes and named eventstream frames. Parse a separate observation for accounting; validate frame CRC and require messageStop and usage metadata for a complete stream. Required guardrail insertion and fully verified privacy masking are narrow exceptions. Reject conflicting guardrails and unpriced service tiers/latency modifiers. Do not silently rewrite sampling parameters. Missing counts, unknown cache TTL and unqualified cached OpenAI input semantics cannot establish a known financial total.

Expose POST /admin/evaluations only in mayu's administrator plane. Require both full admin authorization and a separate virtual key; create fresh headers and invoke the ordinary data handler in-process. No arbitrary URL or upstream credential input, no control-plane inference path, no alternate admission lane. Request-scoped observers expose settlement evidence without modifying authority. Multiple attempts cannot be presented as a known total from the last settlement.

The UI runs sequential A/B then B/A rounds, at most five rounds, with a shared prompt and explicit output limit. It distinguishes HTTP failures, refusals, truncations and empty/malformed responses. Costs use integer microUSD; missing prices are unknown. Non-streaming output tokens/E2E second is explicitly not decode throughput or TTFT. In-memory export omits virtual keys, prompts and output text. This does not add persistent benchmark storage or a quality grader.

Consequences and qualification

Configured Bedrock Runtime targets can receive native requests; upstream model, region and feature availability still apply. Mantle is not a Converse backend. This does not establish compatibility for every Bedrock model.

The live qualification consists of one administrator Converse request and one boto3 ConverseStream request each for Claude Haiku 4.5 and Qwen3 32B. Offline adversarial tests cover auth, policy, wire format, stream failures and cost uncertainty; those fixtures are not model quality evidence. Long-context results remain in the earlier benchmark, not implied by the four short calls.

Existing aggregate analytics are not upgraded into a complete uncertainty-aware financial ledger by this UI feature. Prices remain configured estimates until reconciled to billing records.

See contract and use and the published measurements.