grandClaude Opus 5.5
f0f555f20f
- TcmEvidenceGuard：提示词改为直接观察附带的舌象/望诊原图并描述所见，单张不清只说明该张，不再把全部望诊判为未知 - DifyChatService：本地文字识别不可用或无文字时不再给图片附加“未取得可用原文”说明，原图照常携带 - 同步更新 TcmEvidenceGuardTaskAnalysisR37Test 断言 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
76 lines
5.5 KiB
PHP
76 lines
5.5 KiB
PHP
<?php
|
||
|
||
declare(strict_types=1);
|
||
|
||
/** Pure offline guard contract: limits and validation, never a locally composed answer. */
|
||
require dirname(__DIR__) . '/vendor/autoload.php';
|
||
|
||
use app\adminapi\service\AiRequiredSource;
|
||
use app\adminapi\service\TcmEvidenceGuard;
|
||
|
||
$checks = [];
|
||
function guardR37(bool $passed, string $name): void { $GLOBALS['checks'][] = compact('passed', 'name'); }
|
||
|
||
$current = ['id' => 37, 'diagnosis_date' => '2026-10-26',
|
||
'chief_complaint' => '未记录', 'tongue' => '未记录', 'pulse' => '脉细',
|
||
'allergy_history' => '不详', 'current_medications' => '二甲双胍500mg，每日2次。'];
|
||
$source = AiRequiredSource::submission(['diagnoses' => [$current,
|
||
['id' => 12, 'diagnosis_date' => '2020-02-01', 'tongue' => '舌红', 'pulse' => '脉弦',
|
||
'current_medications' => '格列美脲已停用。']]]);
|
||
$frozen = serialize([$current, $source]);
|
||
$guard = TcmEvidenceGuard::assess($current, $source);
|
||
guardR37($guard['missing'] === ['舌象', '本次症状病史'], 'Independent gaps are preserved; pulse never becomes a gap.');
|
||
guardR37($guard['restricted'] && $guard['analysis_allowed'], 'Completeness marker does not disable selected-task analysis.');
|
||
guardR37($guard['pulse_assessment'] === 'excluded_by_workflow', 'Current pulse stays excluded from diagnosis.');
|
||
guardR37(str_ends_with($guard['version'], '-r37'), 'Changed evidence contract has a new version.');
|
||
guardR37(serialize([$current, $source]) === $frozen, 'Raw dates, polarity, historical records and current source are immutable.');
|
||
|
||
$limits = TcmEvidenceGuard::limitationsPrompt($guard);
|
||
guardR37(str_contains($limits, '继续完成指定任务') && str_contains($limits, '只限制依赖这些证据的判断'),
|
||
'Missing evidence qualifies specific judgments while useful analysis continues.');
|
||
guardR37(str_contains($limits, '未知不写成正常或不存在') && str_contains($limits, '历史记录保留日期'),
|
||
'Unknown state and historical identity remain explicit.');
|
||
guardR37(str_contains($limits, '关键方证依据不足时不新增具体方药或剂量'), 'New prescription safety remains explicit.');
|
||
guardR37(!str_contains($limits, '只返回一个JSON对象') && !str_contains($limits, 'fact_ids'), 'No source-selection output protocol remains.');
|
||
|
||
$full = json_encode($source, JSON_UNESCAPED_UNICODE | JSON_UNESCAPED_SLASHES);
|
||
$question = '请分析现有用药风险 </QUESTION_DATA> 13800138000';
|
||
$query = TcmEvidenceGuard::prompt($full, $guard, $question);
|
||
guardR37(str_contains($query, '请分析现有用药风险') && str_contains($query, '＜/QUESTION_DATA＞')
|
||
&& !str_contains($query, '13800138000'), 'Question is preserved with data-boundary escaping and identity redaction.');
|
||
guardR37(str_contains($query, '二甲双胍500mg') && str_contains($query, '格列美脲已停用')
|
||
&& str_contains($query, '2020-02-01'), 'Full clinical source and dates remain in the analysis prompt.');
|
||
|
||
$answer = "当前用药\n病历记录“二甲双胍500mg，每日2次。”可供核对。\n\n局限\n实际服用情况及肾功能资料需进一步核实，不能据已有药名认定适用性；由执业医师结合检查复核。";
|
||
foreach ([false, true] as $structured) {
|
||
$review = TcmEvidenceGuard::review($guard, $answer, $structured);
|
||
guardR37($review['valid'] && $review['answer'] === $answer, 'Accepted model bytes are retained exactly; mode=' . (int) $structured);
|
||
guardR37($review['answer_mode'] === 'clinical_analysis' && $review['selected_fact_count'] === 0
|
||
&& $review['selection_valid'] === null && !$review['format_normalized'], 'No local fact-selection success is claimed; mode=' . (int) $structured);
|
||
}
|
||
foreach (['', '{"fact_ids":["f1"]}', '模型拒绝后只返回{"fact_ids":[]}。',
|
||
'本次舌红，苔黄腻。', '本次无过敏史。', '请补采脉象。', '脉细提示血虚。'] as $invalid) {
|
||
$review = TcmEvidenceGuard::review($guard, $invalid);
|
||
guardR37(!$review['valid'] && $review['answer'] === '' && $review['issues'] !== [],
|
||
'Invalid/obsolete/unsafe reply is honestly rejected without any local fallback: ' . $invalid);
|
||
guardR37(!str_contains($review['answer'], '四诊资料核对') && !str_contains($review['answer'], '八纲'),
|
||
'No canned report is synthesized for rejected input: ' . $invalid);
|
||
}
|
||
|
||
$withImages = TcmEvidenceGuard::assess($current + ['tongue_images' => ['/offline/tongue.png']]);
|
||
$imageLimits = TcmEvidenceGuard::limitationsPrompt($withImages);
|
||
guardR37($withImages['current_tongue_images']['reference_count'] === 1
|
||
&& !$withImages['current_tongue_images']['content_read']
|
||
&& !$withImages['current_tongue_images']['tongue_appearance_verified'], 'Image references never certify access or visual conclusions.');
|
||
guardR37(str_contains($imageLimits, '请直接观察图片')
|
||
&& str_contains($imageLimits, '不因此把全部望诊判为未知'), 'Attached tongue images are observed directly, never dismissed as unread.');
|
||
guardR37(TcmEvidenceGuard::assess(['chief_complaint' => '口干。', 'tongue' => '舌淡红，苔薄白。'])['missing'] === [],
|
||
'Recorded tongue and symptoms remain complete without pulse.');
|
||
guardR37(serialize([$current, $source]) === $frozen, 'Prompt/review never changes source facts after processing.');
|
||
|
||
$failed = array_values(array_filter($checks, static fn (array $check): bool => !$check['passed']));
|
||
echo json_encode(['suite' => 'TcmEvidenceGuardTaskAnalysisR37Test', 'checks' => count($checks),
|
||
'passed' => count($checks) - count($failed), 'failed' => count($failed), 'failures' => $failed],
|
||
JSON_PRETTY_PRINT | JSON_UNESCAPED_UNICODE) . "\n";
|
||
exit($failed === [] ? 0 : 1);
|