Skip to content

$ oracle: ok_tool_use_tokens@fixture bills toolUsePromptTokenCount as input, 290 cents (rule 21, ledger x rate card; 1.5.5 billed 250) - #1023

Merged
MattJackson merged 1 commit into
predevfrom
lane-gaps-money-tool-use
Oct 15, 2026
Merged

MattJackson merged 1 commit into
predevfrom
lane-gaps-money-tool-use

Conversation

@MattJackson

Copy link
Copy Markdown
Collaborator

$ DRAFT (rules 15 and 21). Stacked on #999, which is stacked on #926 (rule 19). It lands alone, after #999.

Cell. llm|gemini|gemini|request|ok_tool_use_tokens@fixture. Its 1.5.5 golden lands in #999 and was recorded from the published 1.5.5 by oracle-rerecord run 38549469444. The upstream answer reports:

  • promptTokenCount 11
  • toolUsePromptTokenCount 4
  • candidatesTokenCount 7
  • totalTokenCount 18

Ledger x rate card. Entry 576 (owner 2026-10-02) sets the ledger classes: tool-use prompt tokens are billed input, and a total below the itemized sum (18 < 22) adds no line. So the ledger is 15 in / 7 out.

  • The oracle rate card is input_utok 100000 and output_utok 200000 on every lane (busbar-release recorder/config_emit.rs).
  • 15 × 100000 + 7 × 200000 = 2,900,000 utok = 290 cents, exact.

1.5.5 ledgered 11 / 7 and spent 250 cents, dropping the 4 tool-use tokens. So 1.5.5 was wrong (rule 21).

Entry. RULE21-576 ...:

  • kind improvement, using the CHANGELOG line entry 576 already ships.
  • classes effects.usage and effects.metrics.
  • expected_cells 1.
  • The premise requires the candidate to answer 200 with /effects/usage/spend_cents == 290. The difference is forgiven only when 1.6.0 is right.

The other eight new cells.

  • ok_cachepoint_document@fixture and ok_citation@fixture: 11 / 7 = 250 cents. That equals ledger x rate card, so 1.6.0 stays byte-identical and no entry is needed.
  • stream_upstream_error@fixture x6: 1.5.5 bills nothing (usage {requests: 1}), consistent with the owner's mid-stream-failure ruling (M-2: a failed stream is not billed as a completion). No entry is needed. If the candidate bills these, that is a code defect, not a difference to register.

…fixture bills toolUsePromptTokenCount as input, 290 cents (rule 21: ledger x rate card; 1.5.5 billed 250)

The mock reports promptTokenCount 11, toolUsePromptTokenCount 4, candidatesTokenCount 7,
totalTokenCount 18. Under entry 576's ledger classes the ledger is 15 in / 7 out, and the oracle card
(100000 / 200000 utok) prices that at 2,900,000 utok = 290 cents exactly. 1.5.5 ledgered 11 / 7 =
250 cents. The premise forgives the effects.usage / effects.metrics difference only while the
candidate spends exactly 290 cents.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant