t3-hardware-scoring（t3 硬件评分）

MantaBase T3 Hardware Audit System — v2.1

Objectives

• Crawl product information → Brand Blind → Triple-Auditor independent scoring → Eagle Eye Validation → Final Judge synthesis
• Output: objective Tool / Toy / Trash classification with Eagle Eye safety veto
• Core principles: evidence-first, parallel audits, zero hallucination, transparent math

Architecture Overview (v2.1)

Main Agent                             Subagents
───────────────────────────────────    ──────────────────────────────
Step 1: Data Collection
Step 2: Organize + Brand Blind
  ↓ write 02-brand-blinded.md
                                       Step 3: ×3 parallel (Tool/Toy/Trash)
                                         · read file paths, print JSON to stdout
                                         · main agent writes 03-*.json files
  ↓ receive JSON, write files
  ↓ auto-build auditor_reports.json
                                       Step 4: ×1 Eagle Eye Validator (conditional)
                                         · AI subagent, not Python script
                                         · outputs adjustments[] diff
  ↓ apply diffs, rebuild JSON
Step 5: synthesize_results.py
Step 6: 99-audit-report.md

Key v2.1 changes:

• Step 4 replaces 3-way Peer Review with single AI-driven Eagle Eye Validator
• Subagents print JSON to stdout; main agent writes all files
• New Eagle Eye pattern: Architectural Implausibility
• Mandatory auto-rebuild of auditor_reports.json from 03-*.json files
• Token guardrails on verbatim_evidence and prompt size

Data Collection Strategy (Step 1-2)

Primary: crawl_product_info.py

python3 scripts/crawl_product_info.py --url https://example.com/product --pretty

Uses Jina AI Reader first, then direct HTTP fallback. Exit code 2 = content insufficient → trigger Exa fallback.

Fallback: Exa MCP (use when crawl fails or is insufficient)

mcporter call 'exa.web_search_exa(query: "{product name} specs review", numResults: 5)'
mcporter call 'exa.company_research_exa(companyName: "{brand}", numResults: 3)'

Exa is the preferred fallback — always available, no API key required. Collect at least 3 independent sources before proceeding.

Additional sources (check in order)

1. Official product page (specs, pricing, marketing claims)
2. Third-party reviews (TechAdvisor, The Verge, Tom's Guide, TrustedReviews)
3. Community discussion (Reddit, forums) for real-user pain points
4. Investigation/news coverage (privacy issues, lawsuits, controversies)

Minimum data threshold

Proceed only when you have:

• At least one official source with tech specs + price
• At least one third-party review with hands-on findings
• Marketing claims verbatim (needed for Eagle Eye Honest check)

Pre-launch data flag

If the product meets ALL of the following, append (Pre-launch) to the case_id and skip Step 4:

• Zero independent hands-on reviews
• ≥3 critical specs undisclosed (battery, weight, connectivity, regulatory)
• Pre-order only / not shipping

Procedure

Step 1: Collect Raw Data

• Run crawl_product_info.py on official product URL
• If exit code 2, run Exa MCP queries
• Save raw extracts to 01-level0-extracts.md with source tags [S1], [S2], etc.

Step 2: Organize + Brand Blind (combined step)

• Read references/organize-guide.md
• Read references/defluff-guide.md
• Merge all sources into one structured document
• Apply Brand Blinding inline (replace brand names with [BRAND], [PRODUCT], [FEATURE])
• Preserve ALL adjectives and marketing claims verbatim — needed by downstream auditors
• Brand-blinded text should be ≤ 1200 words — deduplicate aggressively but keep all unique claims
• Save to 02-brand-blinded.md

Step 3: Triple Auditor Scoring (independent, parallel)

Each Auditor sees only the Brand-Blinded text from Step 2.

Subagent execution rules (mandatory):

• Pass file path to 02-brand-blinded.md — do NOT embed full text in prompt
• Each subagent reads the file, reads their rubric file, and prints JSON to stdout
• Main agent writes files — subagents never call write (prevents timeout file loss)
• verbatim_evidence limits:

- score 0–1: max 1 quote (or [] for score 0) - score 2–3: max 2 quotes

• Recommended timeouts: Tool=150s, Toy=150s, Trash=180s

🟢 Tool Auditor

• Reads references/tool-auditor.md
• Template: references/tool-auditor-template.md
• Score 11 items (0-3 scale), max 33 points
• Output litmus_gate: "Yes"/"No" at top level

🟡 Toy Auditor

• Reads references/toy-auditor.md
• Template: references/toy-auditor-template.md
• Score 11 items (0-3 scale), max 33 points
• Output litmus_gate: "Yes"/"No" at top level

🔴 Trash Auditor

• Reads references/trash-auditor.md
• Reads references/trash-red-flags.md (required for Eagle Eye)
• Score 14 items (0-3 scale), max 42 points
• Eagle Eye triggers → auto-set score to 3 for flagged items
• Output litmus_gate: "Yes"/"No" and critical_issues: [...] at top level

Key principles:

• Each auditor works from Brand-Blinded text only
• No cross-talk between auditors during scoring
• Evidence FIRST, then score — no score without a verbatim quote
• If a feature isn't in the text, it scores 0

Step 3.5: Auto-rebuild auditor_reports.json (mandatory)

After all three auditors return, the main agent MUST:

1. Read each 03-tool-auditor.json, 03-toy-auditor.json, 03-trash-auditor.json
2. Rebuild auditor_reports.json programmatically: merged = {"tool": {"total_score": tool["total_score"], "litmus_gate": tool["litmus_gate"], "extract_for_report": tool["extract_for_report"]}, "toy": {"total_score": toy["total_score"], "litmus_gate": toy["litmus_gate"], "extract_for_report": toy["extract_for_report"]}, "trash": {"total_score": trash["total_score"], "litmus_gate": trash["litmus_gate"], "critical_issues": trash["critical_issues"], "extract_for_report": trash["extract_for_report"]}}
3. Sanity check before proceeding:

- merged.trash.total_score == sum of all 14 items in 03-trash-auditor.json - len(merged.trash.critical_issues) == count of Eagle Eye triggers in 03-trash-auditor.json - If mismatch: re-read the source file (never hand-edit scores)

**🚨 Never manually write auditor_reports.json from memory or summaries — always rebuild from 03-*.json files.**

Step 4: Eagle Eye Validator (conditional, AI-driven)

Read references/eagle-eye-validator.md

Trigger conditions — run Step 4 if ANY is true:

• Trash report critical_issues is non-empty
• Any auditor has any item scored 3
• Composite score is in Gray Zone (-10 to +10)

Otherwise: skip directly to Step 5.

Execution:

• Spawn 1 subagent (timeout: 120s)
• Subagent reads 02-brand-blinded.md + only the flagged items from 03-*.json
• Performs three checks:

1. Eagle Eye trigger validity — do both conflicting quotes exist verbatim? 2. Architectural Plausibility — are on-device/local claims technically feasible for the form factor? (NEW in v2.1) 3. Score=3 evidence check — does evidence actually support the 3-level rubric?

• Outputs adjustments[] diff (not a full re-score)
• Main agent applies diffs to auditor_reports.json and re-saves

Step 5: Final Judge Synthesis

python3 scripts/synthesize_results.py --input auditor_reports.json --text

Or manually per references/t3-classification.md:

1. Normalize scores: Tool=(raw/33)×100, Toy=(raw/33)×100, Trash=(raw/42)×100
2. Composite = max(NormTool, NormToy) - NormTrash
3. Check primary conditions (need 2+ of 4 conditions per category)
4. Check secondary conditions
5. Apply Eagle Eye Veto (if Trash critical_issues non-empty → force Trash into label)

Step 6: Generate Audit Report

Output file: 99-audit-report.md Directory: tmp/reports/t3-{YYYY-MM-DD}-{slug}/

Follow YAML schema in references/report-schema.md.

Output Format for Messaging Surfaces (WhatsApp / Telegram)

When delivering results in a chat interface, do not paste the full file. Use this condensed format instead:

🏷️ T3 Audit — {Product Name}
Price: ${price} | Date: {YYYY-MM-DD}

📊 Scores
🟢 Tool:  {raw}/{33} → {norm}/100
🟡 Toy:   {raw}/{33} → {norm}/100
🔴 Trash: {raw}/{42} → {norm}/100
Composite: {composite:+.1f}

🏆 Classification: {final_label}
Confidence: {confidence}

{If Eagle Eye:}
🚨 Eagle Eye: {trigger name}
"{verbatim quote A}"
vs.
"{verbatim quote B}"

🟢 Tool highlights (3 bullets)
🟡 Toy highlights (3 bullets)
🔴 Trash issues (3 bullets)

💬 One-line verdict

Rules for messaging output:

• No markdown tables (WhatsApp doesn't render them)
• Use bold for emphasis, bullet lists for details
• Keep total message under 600 words
• If Eagle Eye triggered, always show the conflicting quotes

Important Constraints

🚨 Information Isolation (critical)

• Brand-Blinded text is the ONLY input for auditors
• Auditors must never see original brand names, marketing copy, or each other's reports during scoring
• Final Judge reads only the three auditor JSON outputs, not original product text

📊 Objective data completeness

• Collect specs, performance, reliability, market, and cost data
• Brand Blinding must retain all specs and numbers
• Every score must link to extracted verbatim evidence

Eagle Eye enforcement

• Trash Auditor MUST check every item against references/trash-red-flags.md
• Eagle Eye trigger → automatic score 3 → populate critical_issues[]
• Eagle Eye Validator (Step 4) re-checks triggers including Architectural Plausibility
• Final Judge must respect Eagle Eye Veto regardless of composite math

synthesize_results.py input format

The script expects a JSON file like:

{
  "tool":  { "total_score": <int 0-33>, "litmus_gate": "Yes|No", ... },
  "toy":   { "total_score": <int 0-33>, "litmus_gate": "Yes|No", ... },
  "trash": { "total_score": <int 0-42>, "litmus_gate": "Yes|No",
             "critical_issues": ["<eagle eye trigger text>", ...], ... }
}

Resource Index

Scripts

• scripts/crawl_product_info.py — Multi-strategy product crawler (Jina → direct HTTP fallback)
• scripts/synthesize_results.py — Final Judge math (normalization + Eagle Eye + Litmus Gate)

References (v2.1)