[jm-rivera]

On dev, "What share of ODA to Kenya came from the United States each year since 2015?" sometimes asked the user the same single-option question about an "updated or revised dataset definition" twice and reworked its draft four times. Two of seven recorded turns in Langfuse (environment dfo-dev-ai) repeat it. The planner writes both questions through adk_request_input, prompted by user_decision_dropped and user_decision_contradicted refusals, and the four reworks are those two refusals, an invalid_plan (rolling_needs_open_period, no_source_on_record) and a second user_decision_contradicted.

Two faults start it. First, compiler.evidence._tokens split Search's provenance label "OECD DAC2A" into dac2, a, while a requirement stating "OECD DAC2a" gives dac2a. source_requirement_matches then failed, so _pin_source (translator/discovery.py) treated the orchestrator's pin as an unconfirmed source and searched again, and the pin came back "not found". A turn that pinned the same series with no stated source translated with no search. _CAMEL_BOUNDARY now splits a digit from a capital only where the capital starts a word: "OECD DAC2A" and "OECD DAC2a" both give oecd, dac2a, while "Covid19Deaths", "UNICEFData" and "WorldDevelopmentIndicators" tokenize as before. Over every provenance label and ref in evals/, services/ and packages/, tokens change only for the DAC2A and DAC2B names. The matches gained are all DAC2a against DAC2A and DAC2b against DAC2B. The matches lost are a bare "DAC2", "DAC2 A", a single "A" and "G20" against "G20GDP", none of which occurs in the recordings.

Second, DiscoveryLog.recorded_alternative answered the identities of a logged source's sibling measures, never of the source itself. The candidate a pinned_source_not_found referral offered therefore had no identity on record, so after the user accepted it identity_on_record refused its open period, the model closed the period, discovery minted another spec_id, and the user's decision read as contradicted, which asked the question again. The method is now recorded_definition and also answers each logged source's own verified identity. An identity must hash to the spec_id it is looked up by, so it cannot answer a stale one. admit passes it to identity_on_record, and _recorded_source and _first_year_identity read it, so a pin held only in the log gets its first-year read. Over 2015 to 2024, windowed_identity copies the accepted candidate to exactly the spec_id discovery had re-minted, so _window_pins reaches it with no search and the user's decision stays on the accepted id.

A pin of a source only the translator found now moves to its copy over the requirement's years, the rule a pin of a searched source already followed. The user-decision guard is unchanged, so a different dataset is still refused as a contradiction.

admit gains a one-token rename that touches a line #80 also edits, so whichever merges second needs a one-line rebase.

Testing: new rows cover a label with a digit before a capital, a shown block labelled "OECD DAC2A" honouring a requirement stating "OECD DAC2a" with no search, a referral candidate held only in the log carrying an open period, and a first-year read for a log-held pin with no history saved. Each fails without the change. The full harness suite passes (7169 passed, 1 skipped, 1 xfailed). ruff and ty are clean.