Real LLMs (observed with anthropic/claude-haiku-4-5 on a 23-agent run)
sometimes return Likert values as JSON strings ('3' not 3). The 4 subagent
validators rejected this with isinstance(v, int), losing ~30% of agents at
N=23. Added a shared coerce_int helper in base.py that accepts ints and
numeric strings, rejects bools/floats/garbage, and is now used by:
- Longitudinal: response values 1-5
- Diversity: Q-sort placements -3..+3 and 6 Likert axes 1-7
- Delphi: R2 and R3 importance/plausibility 1-5
- Scenario: 4 dimensions 1-7
Validators now coerce in place so downstream code sees ints regardless of
the wire format. Added 8 tests (4 unit on coerce_int + 4 per-subagent
contract tests showing stringified values are accepted).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds SchemaValidationFailure exception carrying both retry attempts' raw
output, so audit.jsonl preserves what the model actually said when an
agent's response can't be coerced into the instrument schema. Lets us
diagnose persona-vs-format failures without re-running. Two new tests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Christian Moellmann