ENGINEERING VERSION LOG · 2026.07

ASR rollout rules and isolated runtimes

Before a model appears in the client, it must be deployed, healthy, and pass the same acceptance suite. This report defines candidate groups and the rollout contract; it does not claim every candidate is already serving production traffic.

InkTyper Engineering2026.07.29

5 planned groupsTarget baseline and isolated candidate runtimes

4 required endpointsServing contract before rollout

1/2/4/8Concurrency acceptance ladder

30 minutesDefault model idle unload

ScopeRollout policyModel existence is not user availability

EvidenceCommon contract plus isolated candidatesObserved results, targets, and estimates are kept separate.

StatusBaseline selected; candidates remain under evaluationThis log records a dated engineering state, not an evergreen promise.

01 · DECISION

Keep verified baselines and isolate candidate groups

SenseVoice is the verified Chinese baseline and Whisper remains a compatibility baseline. Paraformer, Qwen3-ASR, Fun-ASR Nano, GLM-ASR Nano, and FireRedASR2S are staged candidates for isolated validation; they do not enter the client before they are ready.

02 · RUNTIME CONTRACT

Different runtimes, one rollout contract

Before rollout, every candidate runtime must expose /health, /v1/models, /v1/warmup, and /v1/transcribe. Production must record requested_model and actual_model, forbidding silent fallback.

Client model ID→Scheduler→Ready runtime→GPU

03 · ACCEPTANCE

Every model runs the same samples and metrics

Chinese, English, code-switching, noise, and supported dialects.
3–10 seconds, 60 seconds, 5 minutes, and 30 minutes.
Cold load, warm inference, end-to-end, RTF, and peak memory.
Throughput, queue time, and p50/p95/p99 at 1/2/4/8 concurrency.
Chinese CER, English WER, and repetition rate.

04 · CLIENT POLICY

Unready models never enter the selector

Only models backed by a serving replica reporting ready=true are visible. Incomplete, untested, or resource-starved models stay hidden and may not masquerade as another model.

EVIDENCE · DIAGNOSIS

What the evidence establishes—and what it does not

The model catalog is a production capability list, not a research wish list. A model appears in the client only after weights, runtime startup, health, representative Chinese/English acceptance, and resource budgets pass. Both requested_model and actual_model must be recorded.

Reading ruleProduction observations describe the observed path. Controlled benchmarks isolate a component. Targets remain targets until a sustained distribution proves them.

OPERATIONS · OPEN WORK

Failure behavior and the next verification gate

SenseVoice remains the Chinese/English primary baseline and Whisper the compatibility baseline. Paraformer left the production route because mixed-language and English quality were inadequate. Other candidates stay in shadow or isolated runtimes and may not masquerade as supported through silent fallback.

Failure visibility: Every fallback needs a reason code and stage timing. A successful final transcript must not erase evidence that real-time, AI, or paste failed first.
Release evidence: Keep request ID, actual model, node, queue, upload, ASR, AI, and delivery timing together so a regression can be reconstructed end to end.