didi-lot2-backend/backend/services/orchestration-layer/agent-v3/PROMPT_AUDIT.md

DIDI Platform - Audit Complet: Prompt-uri, Separare Lingvistica, Parametri

Data audit: 2026-03-23 Scope: Toate prompt-urile LLM din agent-v3, fluxul de date Framework->Redis->Executor, separare lingvistica RO/EN, parametri injectati, formate raspuns, probleme gasite.

---

CUPRINS

1. Arhitectura generala prompt-uri
2. Inventar complet prompt-uri
3. Techniques - Screening + Deep Analysis
4. AI-Tampered - Disclosure + Screening + Deep + Image
5. Claims - Extraction + Verification
6. Source Assessment - Extraction + Evaluation
7. Verdict - Calculator + Explanation + Virality
8. Vision + Transcription (media pipeline)
9. Fluxul PG -> Redis -> Executor
10. Audit Separare Lingvistica
11. Probleme Gasite (32 issues)
12. Recomandari

---

1. ARHITECTURA GENERALA

Cum ajung prompt-urile la LLM

Admin Dashboard (UI)
    |
    v
didiFramework PUT /api/providers/prompts/:id
    |
    v
PostgreSQL: bos_parammgmt.component_prompt
    |  (component_code, stage_code, system_prompt, user_template)
    |
    v
POST /api/sync-redis (manual trigger)
    |
    v
Redis: didi:config:{component}:{version}:prompts:{stage}
    |  JSON: { "system": "...", "user_template": "..." }
    |
    v
agent-v3 Executor: loadFromRedis() / loadConfig() / loadPrompt()
    |
    v
Template variable replacement: {{text}}, {{dimensions_list}}, etc.
    |
    v
wrapUserContent() - securitate anti-injection
    |
    v
LLM API call (model cascade: primary -> fallback_1 -> fallback_2 -> fallback_3)
    |
    v
JSON parse + Zod schema validation

Trei tipuri de prompt-uri

Tip	Descriere	Editabil din UI?	Exemple
Redis-only	Incarcat din Redis, FARA fallback hardcodat	Da	Techniques screening/deep, AI-Tampered screening/deep, Claims extraction/verification
Redis + Fallback	Redis override cu default hardcodat in cod	Da	Source Assessment extraction/evaluation, Vision prompts, Verdict explanation
Hardcodat	Fix in cod, nu poate fi modificat din UI	Nu	AI Image detection, Disclosure patterns (regex), Virality calculator

---

2. INVENTAR COMPLET PROMPT-URI

15 prompt-uri LLM identificate

#	Componenta	Etapa	Redis Key	Fallback?	Limba	Fisier
1	Techniques	Screening	didi:config:techniques:v3:prompts:screening	NU	EN (Redis)	executor.ts:322
2	Techniques	Deep Analysis	didi:config:techniques:v3:prompts:deep_analysis	NU	EN (Redis)	executor.ts:380
3	AI-Tampered	Screening	didi:config:ai-tampered:v1:prompts:screening	NU	EN (Redis)	executor.ts:519
4	AI-Tampered	Deep Analysis	didi:config:ai-tampered:v1:prompts:deep_analysis	NU	EN (Redis)	executor.ts:579
5	AI-Tampered	Image Detection	N/A (hardcodat)	N/A	EN	ai-tampered-routes.ts:634
6	Claims	Extraction	didi:config:claims:v1:prompts:extraction	NU	EN (Redis)	executor.ts:358
7	Claims	Verification	didi:config:claims:v1:prompts:verification	NU	EN (Redis)	executor.ts:406
8	Source Assess.	Extraction	didi:config:source-assessment:v1:prompts:extraction	DA	EN	executor.ts:124-154
9	Source Assess.	Evaluation	didi:config:source-assessment:v1:prompts:evaluation	DA	EN	executor.ts:156-209
10	Verdict	Explanation	didi:config:pipeline:v1:prompts:verdict_explanation	DA	EN prompt, RO+EN output	verdict-explanation.ts:88-131
11	Vision	Text Extraction	didi:config:vision:v1:prompts:extraction	DA	EN	pipeline-routes.ts:103
12	Vision	Video Frames	didi:config:vision:v1:prompts:video_frames	DA	EN	video-processor.ts
13	Vision	AI Detection (video)	didi:config:vision:v1:prompts:ai_detection	DA	EN	video-processor.ts
14	AI-Tampered	Disclosure Check	N/A (regex)	N/A	EN	executor.ts:190-231
15	Verdict	Virality Calc	N/A (algoritm)	N/A	N/A	virality-calculator.ts

3 prompt-uri NON-LLM (regex/algoritm)

#	Ce face	Tip	Configurabil?
14	Disclosure detection (ChatGPT, Claude, etc.)	Regex patterns	NU (hardcodat)
15	Virality score (emotie, urgenta, reach)	Algoritm numeric	NU
-	Domain analysis (varsta, SSL, blacklist)	API extern + scoring	Partial (Redis)

---

3. TECHNIQUES

3.1 Screening (Etapa 1)

Scop: Detectie rapida a dimensiunilor de manipulare prezente in text.

Redis Key: didi:config:techniques:v3:prompts:screening

Variabile injectate:

• {{dimensions_list}} - Lista celor 8 dimensiuni, format: - D1: Emotional Manipulation - short_description
• {{text}} - Primele 3000 caractere, wrappate in <analyzed_content>...</analyzed_content>

Sursa dimensiuni: didi:framework:dimensions_compact sau didi:config:techniques:v3:dimensions_compact

Format raspuns asteptat (Zod validated):

{
  "detected_dimensions": ["D1", "D3"],
  "confidence_per_dimension": { "D1": 85, "D3": 72 },
  "quick_reasoning": "Text shows emotional appeals and logical fallacies"
}

Early exit: Daca 0 dimensiuni detectate, se opreste (nu ruleaza deep analysis).

3.2 Deep Analysis (Etapa 2)

Scop: Per dimensiune detectata, identifica tehnicile specifice de manipulare.

Redis Key: didi:config:techniques:v3:prompts:deep_analysis

Variabile injectate:

• {{dimension_name}} - ex: "Emotional Manipulation"
• {{dimension_code}} - ex: "D1" (replaced global)
• {{techniques_list}} - Ierarhie completa: subdimensiuni -> tehnici -> indicatori, format markdown
• {{text}} - Primele 12000 caractere, wrappate in <analyzed_content>

Format raspuns asteptat (Zod validated):

{
  "detected_techniques": [
    {
      "technique_id": 12,
      "technique_name": "Appeal to Authority",
      "confidence": 82,
      "intensity": 7,
      "evidence": "Uses unnamed experts to claim credibility"
    }
  ]
}

Paralelizare: Cate un call LLM per dimensiune detectata, toate in paralel.

3.3 Model Cascade

Configurat in Redis didi:config:techniques:v3:stage_assignments:

• Primary + 3 fallbacks per etapa
• Fiecare model are: temperature, max_tokens, timeout_ms

3.4 Scoring (post-LLM)

Din Redis didi:config:techniques:v3:scoring_config:

• count_scaler, intensity_weight, severe_threshold
• manipulation_score = f(tehnici detectate, intensitate, severitate)

---

4. AI-TAMPERED

4.1 Disclosure Check (Etapa 0 - fara LLM)

Tip: Regex pattern matching, hardcodat in executor.ts:190-231

Patterns detectate:

• AI Tool names: ChatGPT, GPT-3/4, Claude, Bard, Llama, Gemini, Copilot, Jasper, Writesonic, Copy.ai, Notion AI, Bing Chat
• Explicit disclosure: [AI-generated], This content was AI generated, etc.
• Partial disclosure: AI tools, AI assistance, used AI

Output: { type: 'explicit' | 'partial' | 'implied' | 'none', text?: string }

PROBLEMA: Patterns hardcodate, nu pot fi actualizate din UI. Nu detecteaza tool-uri noi (Sora, Udio, etc.).

4.2 Screening (Etapa 1)

Redis Key: didi:config:ai-tampered:v1:prompts:screening

Variabile injectate:

• {{categories_list}} - Categorii AI detection: - CODE: NAME - SHORT_DESCRIPTION
• {{text}} - Text wrappat, truncat la SCREENING_TEXT_LIMIT

Format raspuns:

{
  "ai_probability": 75,
  "detected_categories": ["SYNTAX", "STYLE"],
  "confidence_per_category": { "SYNTAX": 80, "STYLE": 65 },
  "quick_indicators": ["uniform sentence length", "lack of typos"],
  "quick_reasoning": "Text exhibits AI-like patterns"
}

Early exit: Daca ai_probability < 20 AND no disclosure AND no categories -> skip deep analysis.

4.3 Deep Analysis (Etapa 2)

Redis Key: didi:config:ai-tampered:v1:prompts:deep_analysis

Variabile injectate:

• {{category_name}}, {{category_code}}
• {{indicators_list}} - Indicatori per categorie: - ID: NAME\n DESCRIPTION
• {{text}} - Text complet (fara truncare)

Format raspuns:

{
  "category": "SYNTAX",
  "detected_indicators": [
    { "indicator_id": "SYN_01", "confidence": 85, "evidence": "..." }
  ]
}

4.4 Image AI Detection (Vision)

Tip: HARDCODAT in ai-tampered-routes.ts:634-653

Prompt complet:

Analyze this image to determine if it was AI-generated (by DALL-E, Midjourney,
Stable Diffusion, etc.) or is a real photograph/human-created image.

Look for these AI generation indicators:
1. Anatomical errors: Extra fingers, merged hands, distorted faces
2. Texture anomalies: Overly smooth skin, plastic-like appearance
3. Background artifacts: Blurred or nonsensical backgrounds
4. Lighting inconsistencies: Shadows going different directions
5. Text/writing errors: Garbled text, nonsensical letters
6. Repetitive patterns: Unnatural repetition in textures
7. Watermarks/signatures: AI tool watermarks
8. Style indicators: Characteristic AI art styles
9. Edge artifacts: Unnatural edges, halos
10. Composition issues: Unnatural object placement

Return JSON only:
{ "ai_generated_probability": 75, "indicators": [...], "evidence": "..." }

Vision Cascade: Qwen Local (10.11.10.17:14011) -> Gemini Flash -> GPT-4o Fallback daca toate esueaza: Returns neutral 50% probability PROBLEMA: Nu este configurabil din Redis (key definit in keys.ts dar nefolosit).

4.5 Scoring (post-LLM)

Din Redis didi:config:ai-tampered:v1:scoring_config:

• blend_weights (screening vs deep)
• disclosure_impact multipliers
• thresholds for verdict categories

---

5. CLAIMS

5.1 Extraction (Etapa 1)

Redis Key: didi:config:claims:v1:prompts:extraction

System Prompt (din seed 002):

You are a claim extraction expert. Extract all verifiable factual claims from
the given text. A claim is a statement that can potentially be verified as true
or false. DO NOT include opinions, questions, or subjective statements unless
they are presented as facts.

Variabile injectate:

• {{types_list}} - Tipuri claim din framework: - VF: Verifiable Fact - Can be checked (Web Search)
• {{text}} - Max 10000 caractere, wrappat in <analyzed_content>

Tipuri claim (9): EF, VF, RE, SC, QA, CC, PC, OF, VC

Format raspuns:

{
  "claims": [
    { "text": "exact claim", "type": "VF", "priority": "high", "context": "..." }
  ]
}

IMPORTANT: Max 7 claims verificate (restul silentios filtrate). Claims cu priority='low' sunt sarite.

5.2 Verification (Etapa 2)

Redis Key: didi:config:claims:v1:prompts:verification

Flux: Per claim extras -> M17 Web Search -> LLM verification

M17 Web Search API: POST http://10.11.10.17:51100/v1/gather

{ "claim": "...", "max_search_results": 5, "auto_fallback": true,
  "include_full_text": true, "timeout_seconds": 90, "language": "auto" }

Variabile injectate in prompt verificare:

• {{claim}} - Text claim, wrappat in <extracted_data type="claim">
• {{claim_type}} - ex: "VF - Verifiable Fact"
• {{evidence}} - Rezultate web search formatate: [1] title\nURL: ...\nSource: ...\nContent: ...
• {{statuses}} - Status codes: VT, LT, UV, LF, VF, OP, NV

Format raspuns:

{
  "sources_analysis": [
    { "url": "...", "stance": "SUPPORTS|CONTRADICTS|NEUTRAL", "reliability": "official|news|blog|unknown" }
  ],
  "agreement_score": 75,
  "confidence": 80,
  "status": "VT",
  "reasoning": "explanation"
}

IMPORTANT: Server-ul IGNORA status-ul si agreement_score de la LLM si le recalculeaza din stances! LLM-ul furnizeaza doar analiza surselor, nu decizia finala.

5.3 Scoring (post-LLM)

Din Redis didi:config:claims:v1:scoring_config:

• status_thresholds (VT: min_confidence 85, min_agreement 85, etc.)
• source_reliability_weights (official: 1.2, news: 1.0, blog: 0.7, unknown: 0.5)
• claim_type_weights (EF: 0.95, VF: 0.85, etc.)

---

6. SOURCE ASSESSMENT

6.1 Extraction (Etapa 1)

Redis Key: didi:config:source-assessment:v1:prompts:extraction Fallback hardcodat: DA (DEFAULT_EXTRACTION_PROMPT, executor.ts:124-154)

System: "You extract source attribution metadata from text. Output ONLY valid JSON."

Variabile:

• {{text}} - Primele 2000 caractere
• {{url_context}} - URL daca exista

Output: { publication, author, platform_code, content_type, url_found, queries[] }

6.2 Evaluation (Etapa 2)

Redis Key: didi:config:source-assessment:v1:prompts:evaluation Fallback hardcodat: DA (DEFAULT_EVALUATION_PROMPT, executor.ts:156-209)

Variabile:

• {{publication}}, {{author}}, {{content_type}}, {{platform_code}}
• {{domain_context}} - Rezultat Domain Check API
• {{evidence_summary}} - Rezultate M17 search
• {{source_type_options}}, {{author_options}}, {{platform_options}}
• {{credibility_indicators}}

Output: { source_type_id, author_classification_code, platform_code, credibility_indicators[], publication_confirmed, author_confirmed, reasoning }

6.3 Scoring

Din Redis didi:config:source-assessment:v1:scoring_config:

• axis_weights: publication 0.35, domain 0.25, author 0.25, platform 0.15
• verdict_thresholds: TRUSTED >=70, NEUTRAL >=50, SUSPICIOUS >=30, else UNTRUSTED

6.4 API-uri externe

• M17 Search: POST http://10.11.10.17:51100/v1/search (timeout 20s)
• Domain Check: POST http://<domain-check-host>:11000/api/v1/check/check (timeout 15s)

---

7. VERDICT

7.1 Verdict Calculator (algoritm, fara LLM)

Fisier: verdict-calculator.ts

Formula de baza:

risk_score = SUM(component_score * weight) pentru fiecare componenta activa

Ponderi default (din Redis didi:framework:weights):

• manipulation (techniques): 35%
• claims: 25%
• source: 20%
• ai_tampered: 10%
• context: 10%

Input Profiles (din Redis didi:config:pipeline:v1:input_profiles): 6 profiluri: text_no_url, text_with_url, url, image, audio, video Fiecare profil defineste:

• Ponderi per componenta (suprascriu default-urile)
• Reguli INCONCLUSIVE (min_components, required_any, primary_components)
• Override-uri (false_claims, severe_techniques, undisclosed_ai, untrusted_domain)
• AI disclosure multipliers (explicit, partial, implied, none)

Overrides (bonusuri la risk_score):

• false_claims: +15 per claim fals, max +40
• severe_techniques: +10 daca >= 2 tehnici severe
• undisclosed_ai: +15 daca AI nedezvaltuit
• untrusted_domain: +20/+10/+25 (untrusted/suspicious/blacklisted)
• domain_red_flags: +5 per flag, max +15
• synergy: +5 per componenta peste threshold, max +15

Categorii verdict (din Redis didi:framework:verdicts): RELIABLE (0-15), MOSTLY_RELIABLE (16-30), MIXED (31-55), QUESTIONABLE (56-75), UNRELIABLE (76-90), DISINFORMATION (91-100), INCONCLUSIVE (special)

7.2 Verdict Explanation (LLM review)

Redis Key: didi:config:pipeline:v1:prompts:verdict_explanation Fallback hardcodat: DA (verdict-explanation.ts:88-131)

System Prompt (esenta):

You are the final judge. Use your own knowledge to evaluate.
Treat unverified claims as suspicious if verifiable.
Do NOT rubber-stamp the algorithm. Override when reasoning demands it.

Variabile injectate:

• {{framework_params}} - Categorii, ponderi, flow algoritm, nivele confidence
• {{component_results}} - Rezultate per componenta: [RAN]/[CRASHED]/[SKIPPED] + scoruri
• {{math_verdict}} - Verdict algoritmic: risk_score, confidence, severity, weights, overrides
• {{max_adjustment}} - Cat poate ajusta (default 30 puncte)
• {{baseline_score}} - Scorul matematic de referinta
• {{valid_categories}} - RELIABLE, MOSTLY_RELIABLE, MIXED, QUESTIONABLE, UNRELIABLE, DISINFORMATION, INCONCLUSIVE

FORMAT RASPUNS (singurul cu output bilingv):

{
  "risk_score": 0-100,
  "risk_category": "UNRELIABLE",
  "confidence": 0-100,
  "explanation_ro": "3-5 propozitii in romana",
  "explanation_en": "3-5 sentences in English",
  "adjusted": true,
  "reasoning": "1-2 sentences why adjustment was made"
}

Fallback parsing: Daca JSON fail, regex: RO: ... si EN: ...

Modele (din Redis didi:config:verdict:v1:available_models):

1. local:qwen3-235b (primary, temp=0.1, 2000 tokens, 45s)
2. openrouter:gemini-flash (fallback, temp=0.1, 2000 tokens, 30s)
3. openrouter:gpt-4o-mini (fallback, temp=0.1, 2000 tokens, 30s)

IMPORTANT: LLM-ul poate AJUSTA scorul cu max +-30 puncte. Ajustarea e clamped la baseline +/- maxAdjustment.

7.3 Virality Calculator (algoritm, fara LLM)

Factori: emotie, urgenta, reach, controversy Output: virality_score 0-100, virality_level, virality_factors

---

8. MEDIA PIPELINE

8.1 Vision - Text Extraction (OCR)

Redis Key: didi:config:vision:v1:prompts:extraction Fallback:

• System: "You are a text extraction specialist. Extract only the meaningful content from images..."
• User: "Extract the main text content from this image. Return ONLY the actual message... If no meaningful text, respond with NO_TEXT_FOUND."

Folosit de: pipeline-routes.ts, claims-routes.ts, routes.ts, component-runner.ts, source-assessment-routes.ts

8.2 Vision - Video Frames (Misinformation)

Redis Key: didi:config:vision:v1:prompts:video_frames Fallback:

• System: "You are a video frame analyst specializing in misinformation detection..."
• User: "Analyze these {{frame_count}} video frames in sequence. Focus on: text overlays, visual manipulation, narrative..."

8.3 Vision - Video AI Detection

Redis Key: didi:config:vision:v1:prompts:ai_detection Fallback:

• System: "You are a video frame analyst specializing in detecting AI-generated content..."
• User: "Analyze these {{frame_count}} frames for: face consistency, lighting coherence, background stability, texture anomalies, motion artifacts... End with AI_CONFIDENCE: <0-100>"

8.4 Vision Model Cascade

Order	Model	Provider	Endpoint	Timeout
1	Qwen3.5-397B-A17B	qwen-local	http://10.11.10.17:14011/v1/chat/completions	60s
2	gemini-2.0-flash-001	openrouter	https://openrouter.ai/api/v1/chat/completions	60s
3	gpt-4o	openrouter	https://openrouter.ai/api/v1/chat/completions	60s

Configurabil din Redis: didi:config:ai-tampered:v1:vision_models

8.5 Transcription (Audio/Video)

Order	Provider	Model	Endpoint	Timeout
1	M17-Whisper	whisper-large-v3	http://10.11.10.17:11000/audio/v1/transcriptions	180s
2	Groq-Whisper	whisper-large-v3-turbo	https://api.groq.com/openai/v1/audio/transcriptions	120s
3	OpenAI-Whisper	whisper-1	https://api.openai.com/v1/audio/transcriptions	120s

• Limba: auto-detect (Whisper)
• Limba detectata returnata dar NU folosita downstream
• Min transcript: 10 caractere (sub = fallback la urmatorul provider)

8.6 Video Processing Pipeline

1. Download video (yt-dlp / direct fetch)
2. Check durata (max 180s)
3. Extract frames ffmpeg (interval 5s, max 10 frames, min 3)
4. Extract audio ffmpeg -> transcribe
5. Analyze frames via vision cascade
6. Merge: [AUDIO TRANSCRIPT]\n...\n\n[VISUAL ANALYSIS]\n...

---

9. FLUX PG -> REDIS -> EXECUTOR

Tabelul component_prompt (PostgreSQL)

bos_parammgmt.component_prompt (
  prompt_id SERIAL PK,
  component_code VARCHAR(50),  -- 'techniques', 'ai-tampered', 'claims', etc.
  stage_code VARCHAR(50),      -- 'techniques_screening', 'claims_extraction', etc.
  system_prompt TEXT,
  user_template TEXT,
  description TEXT,
  UNIQUE (component_code, stage_code)
)

Sync Redis (sync-redis.ts)

fetchPrompts() -> SELECT * FROM component_prompt
  |
  v
Group by component_code -> { "techniques": { "screening": {system, user_template} } }
  |
  v
For each component+stage:
  shortStage = stage_code.replace(component_prefix, '')
  redis.SET("didi:config:{comp}:{version}:prompts:{shortStage}", JSON)

Version Map

techniques -> v3
ai-tampered -> v1
claims -> v1
source-assessment -> v1
pipeline -> v1
vision -> v1

Executor Loading Patterns

Componenta	Pattern	Fallback?
Techniques	loadFromRedis('prompts:screening') -> throw if missing	NU
AI-Tampered	loadFromRedis('prompts:screening') -> throw if missing	NU
Claims	loadConfig('prompts:extraction') -> throw if missing	NU
Source Assessment	loadPrompt('extraction', DEFAULT) -> return default	DA
Verdict Explanation	redis.get(key) -> use hardcoded	DA
Vision prompts	redis.get(key) -> use hardcoded	DA

PROBLEMA CRITICA: Techniques, AI-Tampered si Claims CRAPA daca Redis nu are prompt-urile. Source Assessment si Verdict au fallback.

---

10. AUDIT SEPARARE LINGVISTICA

Matrice limba per componenta

Componenta	Limba System Prompt	Limba User Template	Limba LLM Output	Limba Finala User
Techniques Screening	EN	EN + data any lang	EN (JSON codes)	Coduri dimensiuni (D1-D8)
Techniques Deep	EN	EN + data any lang	EN (JSON codes)	Tehnici: id + name EN
AI-Tampered Screening	EN	EN + data any lang	EN (JSON codes)	Categorii: codes
AI-Tampered Deep	EN	EN + data any lang	EN (JSON codes)	Indicatori: codes
AI-Tampered Image	EN	EN	EN (JSON)	ai_probability + evidence EN
Claims Extraction	EN	EN + data any lang	EN (JSON)	Claims in limba originala textului
Claims Verification	EN	EN + evidence any lang	EN (JSON)	Status codes + reasoning EN
Source Extraction	EN	EN + data any lang	EN (JSON)	Publication/author names
Source Evaluation	EN	EN + evidence any lang	EN (JSON)	Codes + reasoning EN
Verdict Explanation	EN	EN	RO + EN	explanation_ro + explanation_en
Vision Text Extraction	EN	EN	Any (text content)	Text extras in limba imaginii
Vision Video Frames	EN	EN	EN	Analiza vizuala EN
Vision AI Detection	EN	EN	EN	Evidence EN

Constatari cheie

1. SINGURUL output bilingv este Verdict Explanation - produce explanation_ro + explanation_en
2. Toate prompt-urile sunt in engleza - indiferent de limba textului analizat
3. Textul utilizatorului poate fi in orice limba - LLM-ul primeste text RO/EN/etc wrappat in <analyzed_content>
4. Claims sunt extrase in limba originala - daca textul e in romana, claims sunt in romana
5. Evidence de la web search poate fi in orice limba - M17 are language: 'auto'
6. Codurile sunt language-neutral - D1, VT, SYNTAX etc. nu depind de limba
7. Descrierile tehnicilor din framework prefera .en hardcodat (executor.ts:475): tech.description?.en
8. Categoriile verdict din framework au descrieri in ROMANA (din seed)
9. Nu exista parametru de limba pasat la niciun executor

Probleme lingvistice identificate

#	Problema	Severitate	Locatie
L1	Claims extrase in RO dar verificate cu prompt EN - LLM poate confunda	MEDIE	claims/executor.ts
L2	Technique descriptions hardcodat .en - nu exista fallback .ro	MICA	techniques/executor.ts:475
L3	Evidence M17 poate fi in RO dar prompt-ul de verificare e EN	MICA	claims/executor.ts:474
L4	Verdict explanation cere RO+EN dar nu specifica "Romanian" explicit in prompt	MEDIE	verdict-explanation.ts
L5	Disclosure patterns doar EN (nu detecteaza "generat de AI" in romana)	MEDIE	ai-tampered/executor.ts:190
L6	Vision prompts doar EN - nu specifica limba textului din imagine	MICA	vision.ts
L7	Transcription detecteaza limba dar nu o paseaza downstream	MICA	transcription.ts
L8	Framework verdict categories au descrieri RO in DB dar coduri EN - mixing	INFO	didiFramework seed

---

11. PROBLEME GASITE (32 issues)

CRITICE (3)

#	Problema	Impact	Locatie
C1	Techniques/AI-Tampered/Claims CRAPA daca prompt-urile lipsesc din Redis (no fallback)	Serviciul devine inoperabil dupa un flush Redis	techniques/executor.ts, ai-tampered/executor.ts, claims/executor.ts
C2	Float gap in verdict categories: 90.x nu se potriveste UNRELIABLE(76-90) nici DISINFORMATION(91-100) -> RELIABLE	Verdic complet gresit (deja documentat in VERDICT_BUGS_AUDIT.md Bug #1)	verdict-calculator.ts
C3	Sync Redis este MANUAL - nu exista auto-sync la update prompt	Dupa editare prompt din UI, trebuie trigger manual POST /api/sync-redis	sync-redis.ts

MARI (10)

#	Problema	Impact	Locatie
M1	Image AI detection prompt HARDCODAT, nu foloseste Redis key (definit dar nefolosit)	Nu poate fi actualizat fara deploy	ai-tampered-routes.ts:634
M2	Disclosure patterns HARDCODATE (nu detecteaza tool-uri noi: Sora, Udio, Flux)	False negatives pe AI tool-uri noi	ai-tampered/executor.ts:190-231
M3	Max 7 claims verificate, restul silentios filtrate (utilizatorul nu stie)	Texte lungi pierd claims importante	claims/executor.ts:415
M4	Claims cu priority='low' sarite silentios fara indicator in raspuns	Utilizatorul nu stie ce a fost exclus	claims/executor.ts:408
M5	Server ignora status-ul LLM si agreement_score dar nu logeaza discrepanta	Debug dificil cand server si LLM dau rezultate diferite	claims/executor.ts:507
M6	Techniques descriptions hardcodat .en prefer English, fara fallback .ro	Daca prompt-ul e RO dar descrierile sunt EN = mixing	techniques/executor.ts:475
M7	Prompt keys Techniques nu sunt in keys.ts (missing from registry)	Greu de auditat, inconsistenta cod	shared/redis/keys.ts
M8	Text truncat silentios (3000/10000/12000 chars) fara log sau avertizare	Context pierdut pentru texte lungi, utilizator neinformat	techniques/executor.ts, claims/executor.ts
M9	Disclosure patterns doar EN - nu detecteaza "generat de AI" sau "creat cu inteligenta artificiala"	Miss pe continut romanesc	ai-tampered/executor.ts:190
M10	Transcription detecteaza limba dar NU o paseaza la executor	Executorul nu stie ca textul e RO/EN/FR	transcription.ts, video-processor.ts

MEDII (12)

#	Problema	Impact	Locatie
m1	Zod schemas folosesc .passthrough() - extra fields trec nevalidate	LLM hallucinations trec netectate	toate executoarele
m2	Fallback defaults prea generice: ai_probability=0 vs "analysis failed"	Nu se poate distinge 0% real de eroare	ai-tampered/executor.ts
m3	Evidence formatata ca text plain, nu JSON structurat	LLM poate interpreta gresit	claims/executor.ts:462
m4	Indicators din framework incarcati dar NU pasati in prompt (Techniques)	LLM nu stie ce indicatori sa caute	techniques/executor.ts:473
m5	Toate modelele fallback primesc acelasi prompt (no model-specific tuning)	Modele slabe pot esua pe prompt complex	toate executoarele
m6	Verdict explanation parsing are fallback regex dar e fragil	Daca LLM nu respecta formatul, pierde explicatia	verdict-explanation.ts:523
m7	Verdict explanation cere RO+EN dar "Romanian" nu apare explicit in prompt	Depinde de LLM sa ghiceasca limba din "explanation_ro" field name	verdict-explanation.ts:101
m8	Video AI detection prompt cere "AI_CONFIDENCE: X" la final - fragil	Orice variatie in format pierde scorul	video-processor.ts
m9	Temperature fixe per etapa, nu per complexitate text	Text scurt vs lung poate necesita temperature diferite	toate executoarele
m10	Status code mismatch: prompt Claims listeaza 7 coduri (incl NV) dar framework poate avea 6	LLM poate returna cod inexistent	claims seed vs framework
m11	Vision model Qwen converteste URL-uri public->internal dar Gemini/GPT nu pot accesa MinIO intern	Fallback la Gemini/GPT poate esua pe imagini MinIO	vision.ts:189-196
m12	Source Assessment models au timeout 15s - prea scurt pentru modele lente	Timeout prematur pe Qwen local daca e incarcat	source-assessment/executor.ts

MICI / INFORMATIONALE (7)

#	Problema	Impact	Locatie
i1	Nu exista versionare prompt-uri (nu se stie ce versiune ruleaza)	Rollback imposibil	component_prompt table
i2	Console logging in engleza cu emoji-uri, nu structured logging	Parse dificil in monitoring	toate executoarele
i3	Framework categories verdict au descrieri RO in seed dar coduri EN	Inconsistenta cosmetica	didiFramework seed
i4	Claim types au 9 coduri dar unele rareori folosite (PC, VC)	Polueaza prompt-ul de extragere	claims framework
i5	wrapUserContent() elimina </analyzed_content> dar nu alte tag-uri XML	Injection partial posibil cu alte tag-uri	prompt-safety.ts
i6	Video frames: min 3, max 10, interval 5s - hardcodat	Nu se poate ajusta per analiza	video-processor.ts
i7	Redis keys permanente (no TTL) - daca sync esueaza, date vechi raman la infinit	Date potentiale stale	sync-redis.ts

---

12. RECOMANDARI

Prioritate 1 - Fix imediat (fara impact frontend)

1. Adauga fallback prompts la Techniques, AI-Tampered, Claims (ca Source Assessment)

   • Hardcodeaza default-uri in executor care se folosesc daca Redis e gol
   • Estimare: 1-2 ore per componenta

2. Muta Image AI detection prompt in Redis (foloseste ConfigKeys.visionPromptAiDetection deja definit)

   • Estimare: 30 min

3. Adauga disclosure patterns in romana

   • "generat de AI", "creat cu inteligenta artificiala", "produs de ChatGPT" etc.
   • Estimare: 1 ora

4. Fix keys.ts - adauga prompt keys lipsa pentru Techniques

   • Estimare: 15 min

Prioritate 2 - Imbunatatiri lingvistice

5. Adauga parametru language la toate executoarele

   • Detectat automat din transcription sau configurat per analiza
   • Pasează limba in prompt: "The text is in {{language}}. Analyze accordingly."
   • Estimare: 2-3 ore

6. Specifica explicit "Romanian" in verdict explanation prompt

   • "Write explanation_ro in Romanian language" nu doar field name
   • Estimare: 30 min (in Redis via UI)

7. Paseaza limba detectata de transcription la executor

   • transcription.ts deja returneaza language - trebuie propagat
   • Estimare: 1 ora

Prioritate 3 - Imbunatatiri calitate

8. Logeaza text truncation - cand textul depaseste limita, log + metadata in raspuns
9. Logeaza discrepanta LLM vs server la Claims verification (cand server overrides LLM)
10. Indica in raspuns claims sarite - adauga skipped_claims_count in output
11. Auto-sync Redis la update prompt (webhook sau trigger in providers.ts)
12. Adauga prompt versioning - version column in component_prompt + Redis key cu versiune

Prioritate 4 - Arhitecturala

13. Standardizeaza pattern-ul de incarcare prompts - toate executoarele sa foloseasca acelasi mecanism (cu fallback)
14. Adauga health check prompt-uri - endpoint care verifica ca toate prompt-urile exista in Redis
15. Structured logging - JSON logs cu session_id, component, stage, model, duration
16. Model-specific prompt variations - prompt-uri optimizate per tier de model

---

ANEXA: Redis Keys Complete Map

# Framework data (permanent, synced manual)
didi:framework:techniques          # Ierarhie completa tehnici
didi:framework:dimensions_compact  # Lista compacta dimensiuni
didi:framework:claims              # Tipuri, statusuri, confidence claims
didi:framework:verdicts            # Categorii verdict, risk mappings
didi:framework:weights             # Ponderi componente + scenarii
didi:framework:sources             # Evaluare surse
didi:framework:providers           # Config LLM providers + API keys

# Prompt-uri per componenta (permanent, synced manual)
didi:config:techniques:v3:prompts:screening
didi:config:techniques:v3:prompts:deep_analysis
didi:config:ai-tampered:v1:prompts:screening
didi:config:ai-tampered:v1:prompts:deep_analysis
didi:config:claims:v1:prompts:extraction
didi:config:claims:v1:prompts:verification
didi:config:source-assessment:v1:prompts:extraction
didi:config:source-assessment:v1:prompts:evaluation
didi:config:pipeline:v1:prompts:verdict_explanation
didi:config:vision:v1:prompts:extraction
didi:config:vision:v1:prompts:video_frames
didi:config:vision:v1:prompts:ai_detection

# Stage assignments (modele LLM per etapa)
didi:config:techniques:v3:stage_assignments
didi:config:ai-tampered:v1:stage_assignments
didi:config:claims:v1:stage_assignments
didi:config:source-assessment:v1:stage_assignments

# Scoring configs
didi:config:techniques:v3:scoring_config
didi:config:ai-tampered:v1:scoring_config
didi:config:claims:v1:scoring_config
didi:config:source-assessment:v1:scoring_config

# Pipeline config
didi:config:pipeline:v1:component_config
didi:config:pipeline:v1:input_profiles
didi:config:pipeline:v1:verdict_config
didi:config:pipeline:v1:session_config

# Vision models
didi:config:ai-tampered:v1:vision_models

# Verdict models
didi:config:verdict:v1:available_models