Skip to content
Pramen

Dispatch policy

Dispatch policy (E2.1)

Section titled “Dispatch policy (E2.1)”

Label: mock/stub-measured analytical frontier — not live Bedrock. Reopen when S2.2 live provider numbers exist.

Research question RQ1 asks: given per-record work, provider online and batch pricing, batch completion windows, and a pipeline deadline, when does batch dominate online?

Cost model

Section titled “Cost model”

pramen_ai::dispatch estimates USD cost and wall-clock latency for both modes, then recommends the cheaper mode that still meets the deadline. execution: auto uses the same planner when a semantic transform declares:

execution: auto
dispatch:
  expectedRecords: 10000
  deadlineSeconds: 3600
  # optional:
  # inputTokensPerRecord: 800
  # outputTokensPerRecord: 200
  # rateCard: mock   # or openai-compat-stub, bedrock-illustrative

Without those hints, auto stays online (safe for unbounded work).

CLI

Section titled “CLI”
Terminal window
# One workload
pramen ai dispatch-plan --rate-card mock --records 10000 --deadline-seconds 3600


# Published frontier sweep
pramen ai dispatch-plan --sweep --out docs/research/e2-1-dispatch-frontier.md

Frontier

Section titled “Frontier”

The checked-in table is regenerated from the analytical model over volumes {100, 1k, 10k, 100k} × deadlines {5m, 1h, 24h} × rate cards mock and openai-compat-stub. See docs/research/e2-1-dispatch-frontier.md.