Joseph Magly’s AIWG adds optional Jev decisions to existing coding assistants. It handles yes/no questions, choices from named options and bounded scores, with a fallback that returns the decision to the assistant.

Sampling adds calls before it saves work

The inspected askDecision implementation defaults to five requests, escalating to ten total when needed. These are configured request counts, not measured performance results.

  • Requests run in bounded waves under one deadline.
  • Missing configuration, errors and unresolved uncertainty return an explicit fallback.
  • Fallback tells the caller to reason normally; ordinary ask does not itself invoke a second model.
  • The deployed instructions keep destructive or outward-action authorization outside Jev’s role.

An optional agreement policy can accept consistent answers below the ordinary confidence cutoff. The source explicitly labels its aggregate a stability signal. Repeated calls to one model can share the same mistake. Typed output is not a correct decision.

Read the savings ledger as a model

AIWG’s default savings helper credits an answered call with assumed caller reasoning and output tokens, plus estimated input. An offline check of the actual helper credited 1,152 avoided tokens with zero estimated input and zero Jev tokens supplied.

That result demonstrates the formula’s assumptions. No caller reasoning was observed or saved in that check.

  • Known Jev usage includes unsuccessful offload attempts.
  • Fallback receives no avoided-caller credit and still carries the supplied Jev cost.
  • Configured rates and assumed caller work determine the net estimate.

Measure a whole routing policy

The separate benchmark implementation can collect paired Jev and baseline responses, retain usage and served models, and select policy on tuning cases. Held-out cases assess that choice. Its modeled routing cost includes the baseline expense on fallbacks.

This gives operators a way to test whether extra sampling earns its cost. A standalone baseline request still differs from a small decision inside an already-running conversation.

Verification: inspected commit 4bc80292f1faace6b670c759c1fec85da04ca053; three offline savings-helper checks passed with synthetic arguments. The full CLI, sampling path and live benchmark were not run; no security audit was performed.