Add dbn-legal-agent targeted check step to BVJ Analyzer (Step 6b)

Probe testing revealed the fine-tune loops when asked to check a brief
directly (tool-planning architecture conflict) but answers focused legal
Q&A reliably in ~55s. New step 6b asks one targeted question per document
type (akuttvedtak → § 4-25 klar nødvendighet, adopsjon → Strand Lobben,
undersøkelse → fvl § 17/§ 41) and merges the finding into
procedural_red_flags with check_model provenance. Silent on timeout/error.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-05-18 22:38:29 +02:00
parent 04555a96b1
commit 47aa35e946
2 changed files with 141 additions and 0 deletions
+19
View File
@@ -953,6 +953,25 @@ PROMPT;
];
}
// Step 6b: dbn-legal-agent targeted legal Q&A check (azure engines only; silent on failure)
// Asks one focused question about the document's statutory basis to surface domain knowledge
// that Azure reliably misses (klar nødvendighet threshold, Strand Lobben, fvl §17/§41).
if (in_array($engine, ['azure_mini', 'azure_full'], true)) {
$checkFindings = dbnToolsRunLegalCheck(
(string)($json['advocacy_brief'] ?? ''),
$docType
);
if (!empty($checkFindings)) {
if (!is_array($json['procedural_red_flags'] ?? null)) {
$json['procedural_red_flags'] = [];
}
foreach ($checkFindings as $cf) {
$json['procedural_red_flags'][] = $cf;
}
$json['check_model'] = 'dbn-legal-agent';
}
}
return ['json' => $json, 'deploy_label' => $deployLabel];
}