Separate documentation freshness years from product identifiers

scripts/verify_clean_v3_search_quality.mjs

@@ -86,6 +86,9 @@ async function send(page, prompt) {
     selection_mode: contract?.selection_mode ?? null,
     policy_decisions: metrics?.policy_decisions || [],
     proposed_calls: (metrics?.clean_v3_turn || []).flatMap(message => message.tool_calls || []),
+    runtime_trace: metrics?.clean_v3_turn || [],
+    actual_temperature: metrics?.temperature ?? null,
+    actual_max_output_tokens: metrics?.max_output_tokens ?? null,
     tools: starts, outputs, final,
     evidence: events.filter(event => event.type === 'tool_output').map(event => ({
       tool: canonical(event.tool), arguments: event.command,

services/search/core.py

@@ -260,6 +260,16 @@ def _result_has_query_overlap(query: str, result: dict) -> bool:
         if entity_terms and entity_terms[0] not in result_tokens:
             return False
         model_numbers = {token for token in ordered_query_tokens if token.isdigit()}
+        # Temporal qualifiers are not product identifiers. In particular,
+        # query normalization may append "latest 2026" to a documentation
+        # lookup; an evergreen official page need not put that year in its
+        # title/snippet/URL. Retain actual product numbers (including years
+        # used as model names without an explicit temporal qualifier).
+        temporal_years = set(re.findall(
+            r'\b(?:latest|current|updated|as\s+of)\s+(20\d{2})\b',
+            str(query or ''), re.I,
+        ))
+        model_numbers -= temporal_years
         if model_numbers and not model_numbers.issubset(result_tokens):
             return False
 

src/clean_agent_preview.py

@@ -5662,6 +5662,8 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
         'request_context_tokens': last_request_tokens,
         'tool_schema_count': len(offered),
         'agent_rounds': rounds_used,
+        'temperature': temperature,
+        'max_output_tokens': request_max_tokens,
         'tool_calls': calls,
         'tool_events': executions, 'clean_v3_turn': text_only_clean_trace(history[initial_length:]),
         'policy_decisions': policy_decisions,

tests/test_search_empty_query_relaxation.py

@@ -2,6 +2,20 @@ from services.search import core
 import pytest
 
 
+def test_documentation_freshness_year_is_not_a_product_model_number():
+    result = {'title': 'Enhanced Tracking Protection in Firefox',
+              'url': 'https://support.mozilla.org/en-US/kb/enhanced-tracking-protection-firefox-desktop',
+              'snippet': 'Firefox privacy settings and tracking protection documentation.'}
+    assert core._result_has_query_overlap('Mozilla Firefox privacy documentation latest 2026', result)
+
+
+def test_documentation_filter_keeps_real_product_number_requirements():
+    result = {'title': 'WIKING Miro 3 manual', 'url': 'https://example.org/wiking-miro-3', 'snippet': 'WIKING Miro installation guide'}
+    assert core._result_has_query_overlap('WIKING Miro 3 manual latest 2026', result)
+    assert not core._result_has_query_overlap('WIKING Miro 4 manual latest 2026', result)
+    assert not core._result_has_query_overlap('WIKING Miro 2026 manual', result)
+
+
 @pytest.mark.parametrize('comprehensive', [False, True])
 @pytest.mark.parametrize('transient_error', [False, True])
 def test_empty_results_advance_provider_but_transport_errors_get_one_retry(monkeypatch, tmp_path, comprehensive, transient_error):

tests/test_search_observation_budget.py

@@ -85,6 +85,9 @@ async def test_runtime_does_not_append_unverified_search_result_as_citation(monk
         assert 'old-release' not in final
     assert '[Source:' not in final
     assert not any(event.get('type') == 'error' for event in events)
+    metrics = next(event['data'] for event in events if event.get('type') == 'metrics')
+    assert metrics['temperature'] == 0.0
+    assert metrics['max_output_tokens'] == 768
 
 
 def test_news_intent_survives_query_rewording_without_changing_other_fresh_queries():