Skip to content

I3297 — A waste finding names its evidence and confidence and is never presented as a saving

Derive findings (low cache-read share, oversized tool results, repeated tool calls, an expensive model on routine work, subagent cost without an accepted outcome) from attributed usage, each with the events behind it, a confidence and an estimated effect labelled counterfactual.

BLOCKED wave 3 · p3 · implementation profile · parallel safe

Part of every-sessions-usage-is-observed — What every AI session consumed is observed from the provider, attributed to the work it served, and never estimated into a fact.

Blocked. This issue cannot start until I3293, I0501 are done. The status is derived from that, not declared.

Objective

Derive findings (low cache-read share, oversized tool results, repeated tool calls, an expensive model on routine work, subagent cost without an accepted outcome) from attributed usage, each with the events behind it, a confidence and an estimated effect labelled counterfactual.

Why

Recommendations without evidence are opinions with a command around them (routing-recommendations problem statement); ECONOMICS already reserves the counterfactual class for exactly this. The prose-only refusal already holds on master (case 431); this plan's proof is the refusal across generated and UI surfaces. It extends milestone routing-recommendations. Proposed in ADR 0213 (Usage is observed through each provider's official telemetry into one provider-neutral record, and reported, estimated and billed amounts are never merged) (proposed; it carries no authority until the owner accepts it). Sources: chatgpt-majordomus:c236#4, chatgpt-majordomus:c236#7.

Current state

Nothing derives findings.

Desired state

economics findings lists them; economics check refuses prose that turns one into a saving.

Scope

  • apps/majordomus-cli/src/economics
  • docs/ECONOMICS.md
  • apps/majordomus-cli/src/economics/claims.rs
  • share/commands.yaml

Dependencies

Acceptance criteria

  • Each finding cites its events and threshold
  • An estimated effect is classed counterfactual everywhere it is shown
  • economics check refuses a waste finding's counterfactual effect, or a subscription session's price-list estimate, when either is presented as a saving or as spend, reading generated and UI surfaces as well as prose (case)
  • The economics findings capability is declared canonically and its share/commands.yaml and MCP/HTTP/OpenAPI projections are generated (majordomus generate --check)

Validation

  • cargo test --manifest-path apps/majordomus-cli/Cargo.toml --lib economics::claims
  • bash test/run.sh 431_economics_claims_gated
  • majordomus economics check

Evidence required

  • test_waste_findings_cite_evidence

Evidence

None recorded. Every token above needs a command or an artifact behind it before this issue can be completed; narrative is refused.

Timeline

started
—
verified
—
completed
—

Those three fields, the evidence above and the state of the dependencies are all the status is made of. There is no status field to disagree with them.

Canonical record: .ai/repo/project/issues/I3297.yaml. Read it back with majordomus plan show I3297.

Searches the titles and descriptions of every page this site publishes.