feat(tools): converge two-tier Quick/Pro selector onto .no fork

Port the dobetterlegal-tools two-tier quality stack to dobetternorge.no:
QUALITY_TIERS registry + resolveTier (ToolModels), dbnToolsResolveToolRun
(bootstrap), tier read+charge in the 6 analytical endpoints, Quick/Pro
UI + payload.tier on the 6 tool pages/JS, and the bounded
corpusContextForSummarize RAG fix (per-passage trim + total budget +
reranker_enabled). Back-compat: requests without `tier` keep legacy
engine behavior.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

advocate.php

@@ -50,12 +50,12 @@ require_once __DIR__ . '/includes/layout.php';
                         <small id="advInputCount" class="adv-char-count">0 / 4,000</small>
                     </div>
 
-                    <div class="control-row" id="advEngineControl">
-                        <span class="control-label">Engine</span>
-                        <label><input type="radio" name="advEngine" value="azure_mini" checked> &#x2601;&#xFE0F; Claude Haiku 4.5 <small class="control-hint">(fast · ~2-4 min)</small></label>
-                        <label><input type="radio" name="advEngine" value="claude_sonnet"> &#x2601;&#xFE0F; Claude Sonnet 4.6 &#9733;&#9733; <small class="control-hint">(thorough · ~3-5 min)</small></label>
+                    <div class="control-row" id="advTierControl">
+                        <span class="control-label">Quality</span>
+                        <label><input type="radio" name="advTier" value="quick" checked> Quick &#9733; <small class="control-hint">(Claude Haiku · fast · 3 credits)</small></label>
+                        <label><input type="radio" name="advTier" value="pro"> Pro <small class="control-hint">(Claude Sonnet · best · 6 credits)</small></label>
                     </div>
-                    <p class="upload-hint">Both engines run on AWS Bedrock via Claude. Most of the time is spent on multiple question-answering passes — 6–10 sub-questions each requiring a full retrieval and answer cycle. Haiku is faster and handles most cases well. Sonnet produces a more thorough brief with deeper ECHR precedent analysis and stronger multi-party argumentation.</p>
+                    <p class="upload-hint">Most of the time is spent on multiple question-answering passes — 6–10 sub-questions each requiring a full retrieval and answer cycle. Quick uses Claude Haiku 4.5 — faster and handles most cases well (3 credits). Pro uses Claude Sonnet 4.6 — a more thorough brief with deeper ECHR precedent analysis and stronger multi-party argumentation (6 credits).</p>
 
                     <div class="dr-slice-section">
                         <p class="control-label">Corpus slices</p>

api/ask.php

@@ -6,10 +6,11 @@ require_once __DIR__ . '/../includes/ToolModels.php';
 
 dbnToolsRequireMethod('POST');
 dbnToolsRequireAuth();
-$ftUid    = dbnToolsFreeTierCheck('ask');
-$engine   = ToolModels::engineForUser($ftUid, 'azure_mini');
 $input    = dbnToolsJsonInput(25000);
 $language = dbnToolsNormalizeLanguage($input['language'] ?? 'en');
+$run      = dbnToolsResolveToolRun('ask', $input);
+$ftUid    = $run['ftUid'];
+$engine   = $run['engine'];
 
 dbnToolsWithChargedTelemetry('ask', $language, $ftUid, function () use ($input, $language, $engine): array {
     $question = dbnToolsInjectDocContent($input, dbnToolsString($input, 'question', 4000, false));
@@ -20,4 +21,4 @@ dbnToolsWithChargedTelemetry('ask', $language, $ftUid, function () use ($input,
         ? trim($input['profile'])
         : null;
     return (new DbnLegalToolsService())->ask($question, $language, $engine, $persona);
-});
+}, $run['credits'], $run['metadata']);

api/barnevernet.php

@@ -53,7 +53,23 @@ try {
 
     $language       = dbnToolsNormalizeLanguage($input['language'] ?? 'en');
     $advocateRole   = trim((string)($input['advocate_role'] ?? ''));
+    if (isset($input['tier'])) {
+        $run         = ToolModels::resolveTier(dbnToolsFreeTierUid(), 'barnevernet', (string)$input['tier']);
+        $engine      = $run['engine'];
+        $tierCredits = $run['credits'];
+        $tierMeta    = ['tier' => $run['tier'], 'engine' => $engine];
+        if ($ftUid > 0) {
+            $gate = FreeTier::checkAmount($ftUid, 'barnevernet', $tierCredits);
+            if (empty($gate['ok'])) {
+                $emit('error', ['code' => $gate['reason'] ?? 'no_credits', 'message' => 'Insufficient credits for the selected tier.']);
+                exit;
+            }
+        }
+    } else {
         $engine      = ToolModels::engineForUser($ftUid, (string)($input['engine'] ?? 'azure_mini'));
+        $tierCredits = null;
+        $tierMeta    = [];
+    }
     $sliceInput     = $input['slices'] ?? [];
     $controls       = is_array($input['controls'] ?? null) ? $input['controls'] : [];
     $additionalNotes = mb_substr(dbnToolsInjectDocContent($input, trim((string)($input['additional_notes'] ?? ''))), 0, 8000, 'UTF-8');
@@ -154,7 +170,9 @@ try {
         'bvj_doc_type'  => $result['doc_meta']['doc_type'] ?? null,
     ]);
 
-    $ftRemaining = dbnToolsFreeTierDeduct($ftUid, 'barnevernet');
+    $ftRemaining = $tierCredits === null
+        ? dbnToolsFreeTierDeduct($ftUid, 'barnevernet')
+        : dbnToolsFreeTierDeductAmount($ftUid, 'barnevernet', $tierCredits, $tierMeta);
     if ($ftRemaining >= 0) {
         $result['balance'] = $ftRemaining;
     }

api/deep-research.php

@@ -65,8 +65,9 @@ try {
         throw new DbnToolsHttpException('advocate_role is too long.', 422, 'advocate_role_too_long');
     }
     $chargeTool = $advocateRole !== '' ? 'advocate' : 'deep-research';
-    $ftUid = dbnToolsFreeTierCheck($chargeTool);
-    $engine = ToolModels::engineForUser($ftUid, (string)($input['engine'] ?? 'azure_mini'));
+    $run    = dbnToolsResolveToolRun($chargeTool, $input);
+    $ftUid  = $run['ftUid'];
+    $engine = $run['engine'];
     $priorContext        = is_array($input['prior_context'] ?? null) ? $input['prior_context'] : null;
     $branchNotes        = mb_substr(trim((string)($input['branch_notes'] ?? '')), 0, 1000, 'UTF-8');
     $subQsOverride      = is_array($input['sub_questions_override'] ?? null) ? $input['sub_questions_override'] : [];
@@ -160,7 +161,9 @@ try {
         'advocate_role' => $advocateRole !== '' ? $advocateRole : null,
     ]);
 
-    $ftRemaining = dbnToolsFreeTierDeduct($ftUid, $chargeTool);
+    $ftRemaining = $run['credits'] === null
+        ? dbnToolsFreeTierDeduct($ftUid, $chargeTool)
+        : dbnToolsFreeTierDeductAmount($ftUid, $chargeTool, $run['credits'], $run['metadata']);
     if ($ftRemaining >= 0) {
         $result['balance'] = $ftRemaining;
     }

api/discrepancy.php

@@ -41,7 +41,23 @@ try {
     }
 
     $language      = dbnToolsNormalizeLanguage($input['language'] ?? 'en');
+    if (isset($input['tier'])) {
+        $run         = ToolModels::resolveTier(dbnToolsFreeTierUid(), 'discrepancy', (string)$input['tier']);
+        $engine      = $run['engine'];
+        $tierCredits = $run['credits'];
+        $tierMeta    = ['tier' => $run['tier'], 'engine' => $engine];
+        if ($ftUid > 0) {
+            $gate = FreeTier::checkAmount($ftUid, 'discrepancy', $tierCredits);
+            if (empty($gate['ok'])) {
+                $emit('error', ['code' => $gate['reason'] ?? 'no_credits', 'message' => 'Insufficient credits for the selected tier.']);
+                exit;
+            }
+        }
+    } else {
         $engine      = ToolModels::engineForUser($ftUid, (string)($input['engine'] ?? 'azure_mini'));
+        $tierCredits = null;
+        $tierMeta    = [];
+    }
     $sliceInput    = $input['slices'] ?? [];
 
     // Extract file A
@@ -144,7 +160,9 @@ try {
         'deployment'     => $result['trace_metadata']['deployment'] ?? null,
     ]);
 
-    $ftRemaining = dbnToolsFreeTierDeduct($ftUid, 'discrepancy');
+    $ftRemaining = $tierCredits === null
+        ? dbnToolsFreeTierDeduct($ftUid, 'discrepancy')
+        : dbnToolsFreeTierDeductAmount($ftUid, 'discrepancy', $tierCredits, $tierMeta);
     if ($ftRemaining >= 0) {
         $result['balance'] = $ftRemaining;
     }

api/korrespond.php

@@ -165,15 +165,24 @@ try {
     }
 
     // ── Deduct credit now (Pass 2 starts) ───────────────────────────────────────
+    if (isset($input['tier'])) {
+        $run    = dbnToolsResolveToolRun('korrespond', $input);
+        $ftUid  = $run['ftUid'];
+        $engine = $run['engine'];
+    } else {
         $ftUid  = dbnToolsFreeTierCheck('korrespond');
         $engine = ToolModels::engineForUser($ftUid, 'azure_mini');
         $inputEngine = (string)($input['engine'] ?? '');
         if (in_array($inputEngine, ['azure_mini', 'claude_sonnet'], true)) {
             $engine = $inputEngine;
         }
+        $run = ['credits' => null, 'metadata' => []];
+    }
     $length = in_array($input['length'] ?? '', ['concise', 'standard', 'detailed'], true)
         ? (string)$input['length'] : 'standard';
-    $ftRemaining = dbnToolsFreeTierDeduct($ftUid, 'korrespond');
+    $ftRemaining = $run['credits'] === null
+        ? dbnToolsFreeTierDeduct($ftUid, 'korrespond')
+        : dbnToolsFreeTierDeductAmount($ftUid, 'korrespond', $run['credits'], $run['metadata']);
     $creditDeducted = true;
 
     $personaSlug = (isset($input['profile']) && is_string($input['profile']) && trim($input['profile']) !== '')
@@ -212,7 +221,7 @@ try {
             'case_doc_ids'      => $GLOBALS['dbn_last_case_doc_ids'] ?? [],
             'model'             => $engine,
             'latency_ms'        => $result['latency_ms'],
-            'credits_charged'   => 1,
+            'credits_charged'   => $run['credits'] ?? 1,
         ]);
     } catch (Throwable) { /* non-critical */ }
 

api/summarize.php

@@ -6,11 +6,12 @@ require_once __DIR__ . '/../includes/ToolModels.php';
 
 dbnToolsRequireMethod('POST');
 dbnToolsRequireAuth();
-$ftUid = dbnToolsFreeTierCheck('summarize');
 
 $input    = dbnToolsJsonInput(400000);
 $language = dbnToolsNormalizeLanguage($input['language'] ?? 'en');
-$engine   = ToolModels::engineForUser($ftUid, (string)($input['engine'] ?? 'azure_mini'));
+$run      = dbnToolsResolveToolRun('summarize', $input);
+$ftUid    = $run['ftUid'];
+$engine   = $run['engine'];
 $depth    = in_array($input['depth'] ?? '', ['brief', 'standard', 'detailed'], true)
     ? (string)$input['depth'] : 'standard';
 $slices   = is_array($input['slices'] ?? null) ? array_values(array_filter($input['slices'])) : [];
@@ -73,7 +74,9 @@ try {
     $result = (new DbnLegalToolsService())->summarizeWithContext($text, $language, $engine, $corpusContext, $depth);
 
     if ($ftUid > 0) {
-        $balance          = dbnToolsFreeTierDeduct($ftUid, 'summarize');
+        $balance          = $run['credits'] === null
+            ? dbnToolsFreeTierDeduct($ftUid, 'summarize')
+            : dbnToolsFreeTierDeductAmount($ftUid, 'summarize', $run['credits'], $run['metadata']);
         $result['balance'] = $balance;
     }
 

assets/js/advocate.js

@@ -64,7 +64,7 @@
       roleCustom: document.getElementById('advRoleCustom'),
       slices: Array.from(document.querySelectorAll('.adv-slice')),
       langButtons: Array.from(document.querySelectorAll('#advLangSwitcher .lang-btn')),
-      engineRadios: Array.from(document.querySelectorAll('input[name="advEngine"]')),
+      tierRadios: Array.from(document.querySelectorAll('input[name="advTier"]')),
       subQ: document.getElementById('advSubQ'),
       subQVal: document.getElementById('advSubQValue'),
       chunkLimit: document.getElementById('advChunkLimit'),
@@ -337,9 +337,9 @@
     return out;
   }
 
-  function getEngine() {
-    const checked = els.engineRadios.find((r) => r.checked);
-    return checked ? checked.value : 'azure_mini';
+  function getTier() {
+    const checked = els.tierRadios.find((r) => r.checked);
+    return checked ? checked.value : 'quick';
   }
 
   function getControls() {
@@ -371,10 +371,10 @@
       return;
     }
 
-    const engine = getEngine();
-    const expectedDuration = engine === 'azure_full'
-      ? '60–180 seconds with Azure gpt-4o'
-      : (engine === 'gpu' ? '30–90 seconds on GPU' : '15–45 seconds with Azure gpt-4o-mini');
+    const tier = getTier();
+    const expectedDuration = tier === 'pro'
+      ? '3–5 minutes with Claude Sonnet'
+      : '2–4 minutes with Claude Haiku';
 
     setStatus(`Building advocate brief for ${advocateRole}… (${expectedDuration})`, 'busy');
     els.runButton.disabled = true;
@@ -387,7 +387,7 @@
       query,
       paste_text: '',
       slices,
-      engine,
+      tier,
       language: lang,
       controls: getControls(),
       advocate_role: advocateRole,
@@ -505,7 +505,7 @@
     els.runButton.disabled = false;
     renderTrace(finalResult.trace || []);
     renderResults(finalResult);
-    saveToCache(finalResult, { query, role: advocateRole, engine, slices, lang });
+    saveToCache(finalResult, { query, role: advocateRole, tier, slices, lang });
 
     function handleStreamEvent(evt) {
       if (!evt || !evt.event) return;
@@ -957,7 +957,7 @@
     els.input.value = formState.query || '';
     updateCharCount();
     if (formState.role) els.roleSelect.value = formState.role;
-    const radio = els.engineRadios.find((r) => r.value === formState.engine);
+    const radio = els.tierRadios.find((r) => r.value === formState.tier);
     if (radio) radio.checked = true;
     if (formState.slices) {
       els.slices.forEach((btn) => {
@@ -1074,7 +1074,7 @@
           body: JSON.stringify({
             query,
             language: lang,
-            engine: getEngine(),
+            engine: getTier() === 'pro' ? 'claude_sonnet' : 'claude_haiku',
             controls: getControls(),
             advocate_role: advocateRole,
           }),

assets/js/barnevernet.js

@@ -55,7 +55,7 @@
       roleCustom:     document.getElementById('bvjRoleCustom'),
       slices:         Array.from(document.querySelectorAll('.adv-slice')),
       langButtons:    Array.from(document.querySelectorAll('#bvjLangSwitcher .lang-btn')),
-      engineRadios:   Array.from(document.querySelectorAll('input[name="bvjEngine"]')),
+      tierRadios:     Array.from(document.querySelectorAll('input[name="bvjTier"]')),
       subQ:           document.getElementById('bvjSubQ'),
       subQVal:        document.getElementById('bvjSubQValue'),
       chunkLimit:     document.getElementById('bvjChunkLimit'),
@@ -186,9 +186,9 @@
     };
   }
 
-  function getEngine() {
-    const checked = els.engineRadios.find((r) => r.checked);
-    return checked ? checked.value : 'azure_mini';
+  function getTier() {
+    const checked = els.tierRadios.find((r) => r.checked);
+    return checked ? checked.value : 'quick';
   }
 
   // ── File upload ────────────────────────────────────────────────────────────
@@ -385,13 +385,11 @@
       return;
     }
 
-    const engine = getEngine();
+    const tier = getTier();
     const additionalNotes = (els.notes ? els.notes.value : '').trim();
-    const expectedDuration = engine === 'azure_full'
-      ? '90–180 seconds with Azure gpt-4o'
-      : (engine === 'gpu' ? '45–90 seconds on GPU'
-      : (engine === 'dbn_legal' ? '60–120 seconds with Norwegian specialist'
-      : '30–60 seconds with Azure gpt-4o-mini'));
+    const expectedDuration = tier === 'pro'
+      ? '90–180 seconds with Claude Sonnet'
+      : '30–60 seconds with Claude Haiku';
 
     setStatus(`Analysing document for ${advocateRole}… (${expectedDuration})`, 'busy');
     els.runButton.disabled = true;
@@ -403,7 +401,7 @@
 
     const payload = {
       advocate_role:    advocateRole,
-      engine,
+      tier,
       language:         lang,
       slices,
       controls:         getControls(),

assets/js/deep-research.js

@@ -42,7 +42,7 @@
       traceList: document.getElementById('traceList'),
       slices: Array.from(document.querySelectorAll('.dr-slice')),
       langButtons: Array.from(document.querySelectorAll('#drLangSwitcher .lang-btn')),
-      engineRadios: Array.from(document.querySelectorAll('input[name="drEngine"]')),
+      tierRadios: Array.from(document.querySelectorAll('input[name="drTier"]')),
       personaControl: document.getElementById('drPersonaControl'),
       personaSelect: document.getElementById('drPersonaSelect'),
       subQ: document.getElementById('drSubQ'),
@@ -280,9 +280,9 @@
     return out;
   }
 
-  function getEngine() {
-    const checked = els.engineRadios.find((r) => r.checked);
-    return checked ? checked.value : 'azure_mini';
+  function getTier() {
+    const checked = els.tierRadios.find((r) => r.checked);
+    return checked ? checked.value : 'quick';
   }
 
   function getControls() {
@@ -308,10 +308,10 @@
       return;
     }
 
-    const engine = getEngine();
-    const expectedDuration = engine === 'azure_full'
-      ? '60–180 seconds with Azure gpt-4o'
-      : (engine === 'gpu' ? '30–90 seconds on GPU' : '15–45 seconds with Azure gpt-4o-mini');
+    const tier = getTier();
+    const expectedDuration = tier === 'pro'
+      ? '60–180 seconds with Claude Sonnet'
+      : '15–45 seconds with Claude Haiku';
 
     setStatus(`Running deep research… (${expectedDuration})`, 'busy');
     els.runButton.disabled = true;
@@ -325,7 +325,7 @@
       query,
       paste_text: '',
       slices,
-      engine,
+      tier,
       language: lang,
       controls: getControls(),
     };

assets/js/discrepancy.js

@@ -49,7 +49,7 @@
       results:         document.getElementById('dcResults'),
       traceList:       document.getElementById('traceList'),
       langButtons:     Array.from(document.querySelectorAll('#dcLangSwitcher .lang-btn')),
-      engineRadios:    Array.from(document.querySelectorAll('input[name="dcEngine"]')),
+      tierRadios:      Array.from(document.querySelectorAll('input[name="dcTier"]')),
       slices:          Array.from(document.querySelectorAll('.adv-slice')),
       // File A
       zoneA:           document.getElementById('dcZoneA'),
@@ -188,12 +188,10 @@
       return;
     }
 
-    const engine = (els.engineRadios.find((r) => r.checked) || {}).value || 'azure_mini';
+    const tier = (els.tierRadios.find((r) => r.checked) || {}).value || 'quick';
     const slices = getSelectedSlices();
 
-    const expectedDuration = engine === 'azure_full' ? '2-3 minutes'
-                           : engine === 'gpu'        ? '~90 seconds'
-                           : '60-90 seconds';
+    const expectedDuration = tier === 'pro' ? '2-3 minutes' : '60-90 seconds';
 
     setStatus(`Comparing documents… (${expectedDuration})`, 'busy');
     els.runButton.disabled = true;
@@ -203,7 +201,7 @@
     renderTrace(stepState);
 
     const payload = {
-      engine, language: lang, slices,
+      tier, language: lang, slices,
       use_my_case: (typeof window.dbnGetUseMyCase === 'function') ? window.dbnGetUseMyCase() : false,
     };
     const form = new FormData();

assets/js/korrespond.js

@@ -280,7 +280,7 @@
       clarifications: pendingClarifications,
       force_draft:    !!forceDraft,
       use_my_case:    (typeof window.dbnGetUseMyCase === 'function') ? window.dbnGetUseMyCase() : false,
-      engine:         (document.querySelector('[name="korrEngine"]:checked')?.value ?? 'azure_mini'),
+      tier:           (document.querySelector('[name="korrTier"]:checked')?.value ?? 'quick'),
       length:         (document.querySelector('[name="korrLength"]:checked')?.value ?? 'standard'),
     };
     if (korrDocIds.length) payload.doc_ids = korrDocIds;

assets/js/summarize.js

@@ -179,13 +179,13 @@
             return;
         }
 
-        var engine = (document.querySelector('input[name="sumEngine"]:checked') || {}).value || 'azure_mini';
+        var tier = (document.querySelector('input[name="sumTier"]:checked') || {}).value || 'quick';
         var slices = activeSlices();
 
         var payload = {
             text:     combined,
             language: _currentLang,
-            engine:   engine,
+            tier:     tier,
             depth:    (document.querySelector('input[name="sumDepth"]:checked') || {}).value || 'standard',
             slices:   slices,
         };

barnevernet.php

@@ -36,14 +36,12 @@ require_once __DIR__ . '/includes/layout.php';
                         <p class="upload-hint">The agent will analyse the document from <em>your</em> perspective — identifying supporting statutes, procedural red flags, and ECHR arguments for your position.</p>
                     </div>
 
-                    <div class="control-row" id="bvjEngineControl">
-                        <span class="control-label">Engine</span>
-                        <label><input type="radio" name="bvjEngine" value="azure_mini" checked> Azure gpt-4o-mini &#9733; <small class="control-hint">(~30-60s)</small></label>
-                        <label><input type="radio" name="bvjEngine" value="azure_full"> Azure gpt-4o <small class="control-hint">(best · ~90-180s)</small></label>
-                        <label><input type="radio" name="bvjEngine" value="gpu"> GPU qwen2.5:14b <small class="control-hint">(local · ~45-90s)</small></label>
-                        <label><input type="radio" name="bvjEngine" value="dbn_legal_v3"> &#x1F1F3;&#x1F1F4;&#9876;&#65039; DBN Legal Agent &#9733; <small class="control-hint">(dbn-legal-agent-v3 fine-tune · ~20-60s)</small></label>
+                    <div class="control-row" id="bvjTierControl">
+                        <span class="control-label">Quality</span>
+                        <label><input type="radio" name="bvjTier" value="quick" checked> Quick &#9733; <small class="control-hint">(Claude Haiku · fast · 3 credits)</small></label>
+                        <label><input type="radio" name="bvjTier" value="pro"> Pro <small class="control-hint">(Claude Sonnet · best · 6 credits)</small></label>
                     </div>
-                    <p class="upload-hint">Engine applies to the final advocacy synthesis only. Norwegian specialist v3 is the recommended choice for Barnevernet documents — it is fine-tuned on § 4-25, Strand Lobben, forvaltningsloven § 17/§ 41, and procedural red-flag detection. Classification, party extraction, and timeline always use azure-mini.</p>
+                    <p class="upload-hint">Quality applies to the final advocacy synthesis only. Quick uses Claude Haiku 4.5 — fast and accurate for most Barnevernet documents (3 credits). Pro uses Claude Sonnet 4.6 — better at § 4-25 threshold errors, Strand Lobben / forvaltningsloven § 17/§ 41 procedural red flags, and ECHR argumentation (6 credits). Classification, party extraction, and timeline always use the quick engine.</p>
 
                     <div class="dr-slice-section">
                         <p class="control-label">Corpus slices</p>

deep-research.php

@@ -16,14 +16,12 @@ require_once __DIR__ . '/includes/layout.php';
                         <button type="button" class="lang-btn" data-lang="pl">&#127477;&#127473; PL</button>
                     </div>
 
-                    <div class="control-row" id="drEngineControl">
-                        <span class="control-label">Engine</span>
-                        <label><input type="radio" name="drEngine" value="azure_mini" checked> Azure gpt-4o-mini &#9733; <small class="control-hint">(~15-45s)</small></label>
-                        <label><input type="radio" name="drEngine" value="azure_full"> Azure gpt-4o <small class="control-hint">(best · ~60-180s)</small></label>
-                        <label><input type="radio" name="drEngine" value="gpu"> GPU (cuttlefish) <small class="control-hint">(local · ~30-90s)</small></label>
-                        <label><input type="radio" name="drEngine" value="dbn_legal_v3"> &#x1F1F3;&#x1F1F4;&#9876;&#65039; DBN Legal Agent &#9733; <small class="control-hint">(dbn-legal-agent-v3 fine-tune · ~20-60s)</small></label>
+                    <div class="control-row" id="drTierControl">
+                        <span class="control-label">Quality</span>
+                        <label><input type="radio" name="drTier" value="quick" checked> Quick &#9733; <small class="control-hint">(Claude Haiku · fast · 6 credits)</small></label>
+                        <label><input type="radio" name="drTier" value="pro"> Pro <small class="control-hint">(Claude Sonnet · best · 12 credits)</small></label>
                     </div>
-                    <p class="upload-hint">Azure mini is the default and finishes fastest. Azure full is the most thorough. Norwegian specialist v3 is a Qwen2.5 fine-tune optimised for barnevernsloven, ECHR, and forvaltningsloven — best for cases involving § 4-25, Strand Lobben, or procedural challenges.</p>
+                    <p class="upload-hint">Quick uses Claude Haiku 4.5 — fast and accurate for most research (6 credits). Pro uses Claude Sonnet 4.6 — the most thorough synthesis, best for cases involving § 4-25, Strand Lobben, or procedural challenges (12 credits).</p>
 
                     <div class="control-row corpus-persona is-hidden" id="drPersonaControl">
                         <span class="control-label">Domain</span>

discrepancy.php

@@ -55,14 +55,12 @@ require_once __DIR__ . '/includes/layout.php';
 
                     </div>
 
-                    <div class="control-row" id="dcEngineControl">
-                        <span class="control-label">Engine</span>
-                        <label><input type="radio" name="dcEngine" value="azure_mini" checked> Azure gpt-4o-mini &#9733; <small class="control-hint">(~60-90s)</small></label>
-                        <label><input type="radio" name="dcEngine" value="azure_full"> Azure gpt-4o <small class="control-hint">(best · ~2-3 min)</small></label>
-                        <label><input type="radio" name="dcEngine" value="gpu"> GPU qwen2.5:14b <small class="control-hint">(local · ~90s)</small></label>
-                        <label><input type="radio" name="dcEngine" value="dbn_legal_v3"> &#x1F1F3;&#x1F1F4;&#9876;&#65039; DBN Legal Agent &#9733; <small class="control-hint">(dbn-legal-agent-v3 fine-tune · ~30-60s)</small></label>
+                    <div class="control-row" id="dcTierControl">
+                        <span class="control-label">Quality</span>
+                        <label><input type="radio" name="dcTier" value="quick" checked> Quick &#9733; <small class="control-hint">(Claude Haiku · fast · 4 credits)</small></label>
+                        <label><input type="radio" name="dcTier" value="pro"> Pro <small class="control-hint">(Claude Sonnet · best · 8 credits)</small></label>
                     </div>
-                    <p class="upload-hint">Engine applies to the final synthesis only. Norwegian specialist v3 excels at identifying legally significant discrepancies in Barnevernet documents — procedural violations, threshold errors, and missing statutory justifications. Classification, party extraction, timelines, and cross-referencing always use azure-mini.</p>
+                    <p class="upload-hint">Quick uses Claude Haiku 4.5 — fast and accurate for most document comparisons (4 credits). Pro uses Claude Sonnet 4.6 — better at legally significant discrepancies in Barnevernet documents, procedural violations, and threshold errors (8 credits).</p>
 
                     <details class="advanced-panel" id="dcSlicePanel">
                         <summary class="advanced-toggle">Corpus slices <span class="control-hint">(used for legal significance context)</span></summary>

includes/LegalTools.php

@@ -1964,7 +1964,14 @@ PROMPT;
      * Search the shared legal corpus and return top-N passages as a formatted
      * context string. Returns '' on failure so the caller can degrade gracefully.
      */
-    public function corpusContextForSummarize(string $query, int $limit = 8, ?string $persona = null): string
+    public function corpusContextForSummarize(
+        string $query,
+        int $limit = 8,
+        ?string $persona = null,
+        int $maxCharsPerPassage = 1800,
+        int $maxTotalChars = 9000,
+        ?array &$debug = null
+    ): string
     {
         try {
             $client          = dbnToolsRequireClient();
@@ -1989,14 +1996,51 @@ PROMPT;
                 'search_method'       => $searchMethod,
                 'min_private'         => 0,
                 'include_beta_website' => true,
+                'reranker_enabled'    => true,
             ]));
+            // Bound the injected context: trim each passage and cap the total so a single
+            // oversized chunk cannot eat the budget and starve relevant lower-ranked passages.
             $parts = [];
+            $total = 0;
             foreach ($chunks as $c) {
                 $title   = (string)($c['title'] ?? ($c['source'] ?? 'Legal source'));
                 $content = (string)($c['content'] ?? ($c['text'] ?? ''));
-                if ($content !== '') {
-                    $parts[] = "=== {$title} ===\n{$content}";
+                $rawChars = mb_strlen($content, 'UTF-8');
+                $clean = trim(strip_tags($content));
+                if (mb_strlen($clean, 'UTF-8') > $maxCharsPerPassage) {
+                    $clean = rtrim(mb_substr($clean, 0, $maxCharsPerPassage - 1, 'UTF-8')) . '…';
                 }
+                $keptChars = $clean === '' ? 0 : mb_strlen($clean, 'UTF-8');
+                $included = false;
+                if ($clean !== '' && ($total + $keptChars) <= $maxTotalChars) {
+                    $parts[] = "=== {$title} ===\n{$clean}";
+                    $total  += $keptChars;
+                    $included = true;
+                }
+                if ($debug !== null) {
+                    $debug['chunks'][] = [
+                        'title'          => $title,
+                        'source_name'    => $c['source_name'] ?? null,
+                        'source_type'    => $c['source_type'] ?? null,
+                        'source_group'   => $c['source_group'] ?? ($c['meta']['source_group'] ?? null),
+                        'category'       => $c['category'] ?? null,
+                        'similarity'     => $c['similarity'] ?? null,
+                        'reranker_score' => $c['reranker_score'] ?? null,
+                        'raw_chars'      => $rawChars,
+                        'kept_chars'     => $included ? $keptChars : 0,
+                        'included'       => $included,
+                    ];
+                }
+            }
+            if ($debug !== null) {
+                $debug['raw_total']     = array_sum(array_column($debug['chunks'] ?? [], 'raw_chars'));
+                $debug['used_total']    = $total;
+                $debug['chunk_count']   = count($chunks);
+                $debug['search_method'] = $searchMethod;
+                $debug['reranked']      = !empty(array_filter(
+                    $debug['chunks'] ?? [],
+                    static fn($r) => $r['reranker_score'] !== null
+                ));
             }
             return implode("\n\n", $parts);
         } catch (Throwable $e) {

includes/ToolModels.php

@@ -19,6 +19,114 @@ final class ToolModels
     public const TIMELINE_STANDARD_MAX_CHARS = 300000;
     public const TIMELINE_DEEP_MAX_CHARS = 600000;
 
+    /**
+     * Canonical redact engine registry. Server-side source of truth for which engines
+     * exist, their credit cost, and whether the client UI may offer them. Credits are
+     * ALWAYS resolved from here — never from a client-supplied value.
+     *
+     * gpu_legal is the fine-tuned legal model (Phase 6 roadmap) — wired but not yet
+     * client-selectable until the harness proves a lift.
+     */
+    public const REDACT_ENGINES = [
+        'azure_mini'    => ['label' => 'Azure GPT-4o-mini',  'credits' => 1, 'client_selectable' => true,  'speed' => 'fast'],
+        'azure_full'    => ['label' => 'Azure GPT-4o',       'credits' => 2, 'client_selectable' => true,  'speed' => 'medium'],
+        'claude_haiku'  => ['label' => 'Claude Haiku 4.5',   'credits' => 1, 'client_selectable' => true,  'speed' => 'fast'],
+        'claude_sonnet' => ['label' => 'Claude Sonnet 4.6',  'credits' => 2, 'client_selectable' => true,  'speed' => 'medium'],
+        'gpu'           => ['label' => 'GPU Qwen 2.5 14B',   'credits' => 1, 'client_selectable' => true,  'speed' => 'slow'],
+        'gpu_legal'     => ['label' => 'GPU Legal (Qwen FT)','credits' => 1, 'client_selectable' => false, 'speed' => 'slow'],
+        'regex'         => ['label' => 'Regex only',         'credits' => 0, 'client_selectable' => false, 'speed' => 'instant'],
+    ];
+
+    public static function isValidRedactEngine(string $engine): bool
+    {
+        return isset(self::REDACT_ENGINES[$engine]);
+    }
+
+    /** Normalise an incoming engine to a valid key, defaulting to claude_haiku. */
+    public static function redactEngine(string $engine): string
+    {
+        return self::isValidRedactEngine($engine) ? $engine : 'claude_haiku';
+    }
+
+    /** Server-side credit cost for a redact engine (never trust the client). */
+    public static function redactCredits(string $engine): int
+    {
+        return self::REDACT_ENGINES[self::redactEngine($engine)]['credits'];
+    }
+
+    /** Engine keys the client UI is allowed to offer. */
+    public static function redactSelectableEngines(): array
+    {
+        return array_keys(array_filter(
+            self::REDACT_ENGINES,
+            static fn(array $e) => $e['client_selectable']
+        ));
+    }
+
+    /**
+     * Unified quality-tier registry for the Quick/Pro selector exposed on the analytical
+     * tools (summarize, ask, legal-analysis, barnevernet, korrespond, discrepancy,
+     * deep-research). Each tier maps to an existing engine string that the service layer
+     * already resolves to a gateway (resolveChatGateway/personaGateway) — so no service
+     * change is needed. Credits = the tool's base cost × tier multiplier, resolved
+     * server-side ONLY.
+     *
+     * Anchors picked from the 2026-06-15 tier benchmark (credits-first, Bedrock):
+     *   quick → claude_haiku (96.7% quality, ~3.8s); pro → claude_sonnet (100%, ~13s).
+     */
+    public const QUALITY_TIERS = [
+        'quick' => ['label' => 'Quick', 'engine' => 'claude_haiku',  'credit_mult' => 1],
+        'pro'   => ['label' => 'Pro',   'engine' => 'claude_sonnet', 'credit_mult' => 2],
+    ];
+
+    /** Normalise an incoming tier to a valid key, defaulting to quick. */
+    public static function qualityTier(string $tier): string
+    {
+        return isset(self::QUALITY_TIERS[$tier]) ? $tier : 'quick';
+    }
+
+    /** The existing engine string backing a quality tier. */
+    public static function tierEngine(string $tier): string
+    {
+        return self::QUALITY_TIERS[self::qualityTier($tier)]['engine'];
+    }
+
+    /** Human label for a quality tier (UI / telemetry). */
+    public static function tierLabel(string $tier): string
+    {
+        return self::QUALITY_TIERS[self::qualityTier($tier)]['label'];
+    }
+
+    /** Server-side credit cost for a tool at a given tier (base cost × tier multiplier). */
+    public static function tierCredits(string $tool, string $tier): int
+    {
+        $mult = self::QUALITY_TIERS[self::qualityTier($tier)]['credit_mult'];
+        return max(1, PricingCatalog::toolCost($tool) * $mult);
+    }
+
+    /**
+     * Resolve a requested quality tier for a user + tool. Applies the subscription gate
+     * (free / anonymous → quick only; plus & pro may pick pro) and returns the backing
+     * engine plus the server-side credit cost. Credits are NEVER trusted from the client.
+     *
+     * @return array{tier:string,engine:string,credits:int,label:string}
+     */
+    public static function resolveTier(int $userId, string $tool, string $requestedTier): array
+    {
+        $tier = self::qualityTier($requestedTier);
+
+        if ($tier === 'pro' && $userId > 0 && FreeTier::tier($userId) === 'free') {
+            $tier = 'quick';
+        }
+
+        return [
+            'tier'    => $tier,
+            'engine'  => self::tierEngine($tier),
+            'credits' => self::tierCredits($tool, $tier),
+            'label'   => self::tierLabel($tier),
+        ];
+    }
+
     public static function engineForUser(int $userId, string $requestedEngine): string
     {
         $valid = ['nova_lite', 'azure_mini', 'azure_full', 'gpu', 'regex', 'claude_haiku', 'claude_sonnet'];

includes/bootstrap.php

@@ -767,6 +767,50 @@ function dbnToolsIsFreeTier(): bool
         && !empty($_SESSION['dbn_tools_sso_uid']);
 }
 
+/**
+ * Current free-tier SSO user id, or 0 for CaveauAI / non-SSO sessions.
+ * Lets an endpoint resolve a quality tier (subscription gate) BEFORE charging,
+ * without reaching into the session directly. Mirrors the uid the check/deduct
+ * helpers use.
+ */
+function dbnToolsFreeTierUid(): int
+{
+    return dbnToolsIsFreeTier() ? (int)$_SESSION['dbn_tools_sso_uid'] : 0;
+}
+
+/**
+ * Resolve the model + credit gate for a tool run. When the request carries a `tier`
+ * param, honour the Quick/Pro quality tier (subscription-gated, server-priced); otherwise
+ * fall back to the legacy engine-selector behaviour. Runs the credit gate (exits 402/429
+ * if over limit) and returns the context an endpoint needs to run + deduct.
+ *
+ * @return array{tier:string,engine:string,credits:?int,ftUid:int,metadata:array}
+ */
+function dbnToolsResolveToolRun(string $tool, array $input, string $legacyDefaultEngine = 'azure_mini'): array
+{
+    if (isset($input['tier'])) {
+        $res   = ToolModels::resolveTier(dbnToolsFreeTierUid(), $tool, (string)$input['tier']);
+        $ftUid = dbnToolsFreeTierCheckAmount($tool, $res['credits']);
+        return [
+            'tier'     => $res['tier'],
+            'engine'   => $res['engine'],
+            'credits'  => $res['credits'],
+            'ftUid'    => $ftUid,
+            'metadata' => ['tier' => $res['tier'], 'engine' => $res['engine']],
+        ];
+    }
+
+    $ftUid  = dbnToolsFreeTierCheck($tool);
+    $engine = ToolModels::engineForUser($ftUid, (string)($input['engine'] ?? $legacyDefaultEngine));
+    return [
+        'tier'     => '',
+        'engine'   => $engine,
+        'credits'  => null,
+        'ftUid'    => $ftUid,
+        'metadata' => [],
+    ];
+}
+
 /**
  * Enforce credit + tier gate before a tool call.
  * Exits with JSON 402/429 if the user is over limit or out of credits.

korrespond.php

@@ -79,13 +79,13 @@ require_once __DIR__ . '/includes/layout.php';
                     </div>
                     <p class="upload-hint">Concise: short and to-the-point (2-3 paragraphs). Standard: balanced correspondence. Detailed: full background, all arguments, and complete legal reasoning.</p>
 
-                    <!-- Engine -->
-                    <div class="control-row" id="korrEngineControl">
-                        <span class="control-label">Engine</span>
-                        <label><input type="radio" name="korrEngine" value="azure_mini" checked> &#x2601;&#xFE0F; Quick <small class="control-hint">(~40-70s)</small></label>
-                        <label><input type="radio" name="korrEngine" value="claude_sonnet"> &#x2601;&#xFE0F; Thorough &#9733;&#9733; <small class="control-hint">(~70-120s)</small></label>
+                    <!-- Quality tier -->
+                    <div class="control-row" id="korrTierControl">
+                        <span class="control-label">Quality</span>
+                        <label><input type="radio" name="korrTier" value="quick" checked> Quick &#9733; <small class="control-hint">(Claude Haiku · fast · 3 credits)</small></label>
+                        <label><input type="radio" name="korrTier" value="pro"> Pro <small class="control-hint">(Claude Sonnet · best · 6 credits)</small></label>
                     </div>
-                    <p class="upload-hint">Quick uses Claude Haiku 4.5 for drafting — fast and solid for standard correspondence. Thorough uses Claude Sonnet 4.6 — better at multi-statute cases, complex appeal grounds, and ECHR framing.</p>
+                    <p class="upload-hint">Quick uses Claude Haiku 4.5 for drafting — fast and solid for standard correspondence (3 credits). Pro uses Claude Sonnet 4.6 — better at multi-statute cases, complex appeal grounds, and ECHR framing (6 credits).</p>
 
                     <!-- Domain persona -->
                     <div class="control-row corpus-persona is-hidden" id="korrPersonaControl">

summarize.php

@@ -16,14 +16,12 @@ require_once __DIR__ . '/includes/layout.php';
                         <button type="button" class="lang-btn sum-lang-btn" data-lang="pl">&#127477;&#127473; PL</button>
                     </div>
 
-                    <div class="control-row" id="sumEngineControl">
-                        <span class="control-label">Engine</span>
-                        <label><input type="radio" name="sumEngine" value="claude_haiku" checked> Standard &#9733; <small class="control-hint">(Claude Haiku · fast)</small></label>
-                        <label><input type="radio" name="sumEngine" value="claude_sonnet"> Deep <small class="control-hint">(Claude Sonnet · best)</small></label>
-                        <label><input type="radio" name="sumEngine" value="azure_mini"> Azure mini <small class="control-hint">(gpt-4o-mini)</small></label>
-                        <label><input type="radio" name="sumEngine" value="azure_full"> Azure full <small class="control-hint">(gpt-4o)</small></label>
+                    <div class="control-row" id="sumTierControl">
+                        <span class="control-label">Quality</span>
+                        <label><input type="radio" name="sumTier" value="quick" checked> Quick &#9733; <small class="control-hint">(Claude Haiku · fast · 1 credit)</small></label>
+                        <label><input type="radio" name="sumTier" value="pro"> Pro <small class="control-hint">(Claude Sonnet · best · 2 credits)</small></label>
                     </div>
-                    <p class="upload-hint">Standard uses Claude Haiku 4.5 — fast and highly accurate. Deep uses Claude Sonnet 4.6 — best for complex multi-statute documents.</p>
+                    <p class="upload-hint">Quick uses Claude Haiku 4.5 — fast and highly accurate (1 credit). Pro uses Claude Sonnet 4.6 — best for complex multi-statute documents (2 credits).</p>
 
                     <div class="control-row" id="sumDepthControl">
                         <span class="control-label">Summary depth</span>