Skip to content

ontocast.tool.facts_validation.critic_findings

Convert critic-proposed fixes into the finding shape the repair pass reads.

The loop had two repair channels that never met. Deterministic findings went to _run_finding_driven_repair, a bounded rewrite-in-place pass costing one patch render. Critic fixes went somewhere else entirely: they were stashed on state.suggestions and consumed by the next full render, which re-extracted the unit from scratch under a prompt that also invited it to "proactively identify and fix additional problems not mentioned in the critique".

So the cheap, contract-bound channel carried the reliable evidence and the expensive, open-ended one carried the model's opinion. Routing both through the same findings pipeline is what makes a rejection cost a repair rather than a re-extraction.

critic_fixes_to_findings(fixes, policy=None)

Render critic fixes as findings, blocking ones marked mandatory.

mandatory follows the policy's severity cut, with one exception that is not configurable: an action="REMOVE" fix is never mandatory. The findings block it would be rendered into states that a finding is never resolved by deleting the statement, so a mandatory REMOVE would contradict the instruction printed directly above it -- the same contradiction shape shacl_catalog_contradictions exists to catch, and one that has already caused repair renders to delete valid values wholesale.

Parameters:

Name Type Description Default
fixes Sequence[TripleFix]

Fixes from the critique report, in the order proposed.

 required
policy FactsAcceptancePolicy | None

The deployment's severity cut. None uses the defaults.

 None

Returns:

Type Description
list[FactsUnitFinding]

One finding per fix, advisory unless the policy blocks on it.
Source code in ontocast/tool/facts_validation/critic_findings.py 
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
def critic_fixes_to_findings(
    fixes: Sequence[TripleFix],
    policy: FactsAcceptancePolicy | None = None,
) -> list[FactsUnitFinding]:
    """Render critic fixes as findings, blocking ones marked mandatory.

    ``mandatory`` follows the policy's severity cut, with one exception that is
    not configurable: an ``action="REMOVE"`` fix is **never** mandatory. The
    findings block it would be rendered into states that a finding is never
    resolved by deleting the statement, so a mandatory REMOVE would contradict
    the instruction printed directly above it -- the same contradiction shape
    ``shacl_catalog_contradictions`` exists to catch, and one that has already
    caused repair renders to delete valid values wholesale.

    Args:
        fixes: Fixes from the critique report, in the order proposed.
        policy: The deployment's severity cut. ``None`` uses the defaults.

    Returns:
        One finding per fix, advisory unless the policy blocks on it.
    """
    active = policy if policy is not None else FactsAcceptancePolicy()
    findings: list[FactsUnitFinding] = []
    for fix in fixes:
        message = f"{fix.action}: {fix.explanation}".strip()
        if fix.incorrect_value:
            message = f"{message} (currently: {fix.incorrect_value})"
        findings.append(
            FactsUnitFinding(
                kind=FactsUnitFindingKind.CRITIC_FIX,
                mandatory=active.blocks_fix(fix),
                message=message,
                value=fix.incorrect_value or "",
                suggestions=[fix.correct_value] if fix.correct_value else [],
            )
        )
    return findings