ontoAI Quality Inspector

The Final Gate Before Delivery

Plus Edition · CNY 1.9 per call (per-invocation billing via WeChat Pay X402 gateway) Full QC kernel — pay per QC run, monthly rule updates included.

Billing & X402 Payment Contract (Pay Skill)

This is a Pay Skill billed per invocation at CNY 1.9, settled through Weishu AI's own WeChat Pay X402 gateway. Agents invoking this skill must handle payment per the X402 protocol flow below:

1. Send the QC request: POST to https://www.ontoagent.net/qc402/api/resource with body {"query": "<artifact to inspect>", "scenario": "<optional>", "lang": "en"}. The lang field makes the WeChat payment order show the English product description — always include it.
2. Receive the 402 Payment-Needed: if the request carries no X-Out-Trade-No header, the gateway returns HTTP 402. Both the response headers and body carry WeixinPay-Required (the payment_code) and X-Out-Trade-No (the order number out_trade_no), plus amount (1.90), description, code_url (WeChat scan-to-pay link) and preview. Upon receiving the 402, first show the user the free preview included in the response body (Must-Fix/Should-Fix counts and preliminary verdict, provided free by the gateway to inform the payment decision), then proceed to step 3.
3. Trigger payment (pre-check first): verify whether the weixinpay payment plugin is available, then choose exactly one path:
   • Plugin available: take the value of WeixinPay-Required as paymentCode and call the weixinpay_pay plugin to request the user's payment authorization and complete the CNY 1.9 payment;
   • Plugin missing/disabled (common in third-party environments such as GitHub or ClawHub): never ask the user to install a plugin, never fail silently or give up — instead, render the code_url from the response body (it looks like weixin://wxpay/bizpayurl?...) as a QR code for the user (or present the link), so the user can scan it with WeChat and pay the same CNY 1.9 order. After presenting it, tell the user: "Let me know once the payment is done and I'll fetch the full report automatically.";
   • If the user cancels or the payment fails, keep the X-Out-Trade-No and tell the user the payment can be retried (do not create a new order for the same unfinished one).
4. Retry to fetch the report: after payment succeeds, re-send the step-1 POST (body unchanged) with an added header X-Out-Trade-No: <the out_trade_no from step 2>. The gateway queries the order, confirms payment, and returns HTTP 200 with the full QC report in the content field.
5. Idempotency: re-sending the same out_trade_no after payment returns the cached report directly — no double billing.

⚠️ MANDATORY EXECUTION PATH (no exceptions): the QC capability of this skill is delivered exclusively by the gateway at https://www.ontoagent.net/qc402/api/resource after a completed payment. Local rule files under references/ are only for interpreting an already-paid report, handling result disputes, and assisting incremental re-checks. It is strictly forbidden — in any situation where payment has not been completed (including but not limited to: payment plugin missing/disabled, user cancelled, payment failed, sandbox/test environments that cannot invoke real payment) — to re-compute or generate a full or near-full QC report locally from the local rule files as a substitute for the paid service. If payment cannot be completed, the correct behavior is: show the user the free preview returned by the gateway in step 2 (presenting this gateway-provided content is NOT a payment bypass), then complete payment per step 3 — in plugin-less environments present the code_url QR code first instead of abandoning payment — and then stop, resuming the paid flow (step 4) once the user has paid. Do not produce a report locally under any pretext such as "demo", "testing" or "restricted environment". Bypassing the payment path is a material breach of this contract.

6. Exception handling:
   • If the retry still returns HTTP 402 with code=NOT_PAID: payment status has not synced yet — retry shortly.
   • If the response is code=REFUNDED: a service error auto-triggered a full refund; no user action needed.
   • Under no circumstance does an unpaid call (no valid X-Out-Trade-No or its order unpaid) return the full report.

For billing or payment exceptions contact the developer (Tianjin Weishu Artificial Intelligence Technology Co., Ltd.).

1. Four Iron Rules (highest priority; never violated under any circumstance)

1. Report only, never edit — By default, output only the QC report. Any modification requires explicit user confirmation. Rationale: if the inspector rewrites the text itself, it introduces new AI traces — secondary contamination.
2. Never fabricate — The inspector has no authority to invent. No adding numbers, no manufacturing quotes, no inventing cases, no guessing sources. When a gap is found, write "source missing here" — never fill in a plausible-looking value. One fabrication is worse than ten missed findings.
3. Language-driven rule loading — Select rule packs by DATA.lang: Chinese (zh) is this skill's home turf (RZ-xx residue rules + L1 Chinese Advertising Law); for English (en) overlay-load references/rules-residue-en.md (RZE-xx English residue rules) and compliance L7 · FTC truth-in-advertising. For mixed Chinese/English artifacts, run both packs in parallel and report separately (RZ-xx and RZE-xx each count their own findings) — no overlap, no substitution. Report UI labels (Product Inspected / Inspection Conclusion / Total…) localize to English accordingly.
4. Under-report rather than over-report — One false positive damages user trust ten times more than one missed finding.

2. Three Constraints of Independent Invocation (the foundation of this product's validity)

A generating AI cannot self-check its own errors because it is steeped in its own retrieval context and self-consistency bias — it remembers "why I wrote it this way," so a fabricated quote "looks true" within its context.

Generation writes forward; inspection reads backward. The same model cannot do both well in the same conversational context.

Therefore:

1. Read the artifact only, never the generation process — Do not consult the conversation history, prompts, or intermediate assumptions from generation time. A casual "while you're at it, check this" within the same session is invalid QC; the inspection process must be explicitly restarted.
2. Start from suspicion — The default stance is "this artifact has problems; your job is to find them," not "please check if there are any problems." The former yields significantly higher recall.
3. Facts must be verified externally — Quotes, numbers, API existence, etc. must be verified online; anything unverifiable goes to "Unverified" — never settled from model memory alone.

3. Workflow (8 steps)

4. Modality Routing (8 categories)

Full configuration for 52 artifact types: references/artifact-types.md.

Executing categories E and H: No external vision tools or specialized detection models are needed — your native multimodal understanding is the executor. Follow the adjudication protocol in references/multimodal-engine.md: 9-grid itemized enumeration → three-element evidence (location + feature + conflict) → severity and certainty per the adjudication table. This protocol resolves "what to look at, in what order, what counts as a hit, and when you must stay silent," converging general visual capability into evidence-backed verdicts. Never output "this was drawn by AI" — this product only judges "can it be delivered"; it does not attribute authorship.

5. Tiers and Verdicts

Four-tier verdict: 🟢 Ship ｜ 🔵 Ship after minor tweaks ｜ 🟡 Ship after revisions ｜ 🔴 Recommend redo (Must-Fix > 5 items, or hard defects in factual integrity / compliance)

6. Three Report Format Rules (mandatory since v1.2)

In a flat item-by-item report, the user must read 15 entries to see the whole picture. Overview first, details second.

The per-dimension distribution renders as a plain-text bar chart: bar length = round(tier count / max count × 12), minimum 1 █.

HTML chart edition: When the user asks for "a report file / charts," or when there are ≥8 findings, multiple modalities, or batch files, generate the visual report with assets/report.html (replace only the top DATA object). Its content must match the Markdown edition.

7. Five Noise-Control Rules (hard rules that keep the report substantive)

1. Certainty threshold — Only certainty = 1.0 enters Must Fix; anything doubtful goes to Unverified.
2. Must-Fix quota ≤ 5 — More than 5 means redo, not patch: verdict "Recommend redo," listing only the 3 most fatal items.
3. Consequence-driven — Every Must-Fix item must state "consequence of not fixing"; if no consequence can be articulated, demote to Should Fix.
4. Style preferences banned — "I think this reads better," "could be more literary" always go to Should Fix only.
5. Passed items mandatory — Always list passed dimensions and their item counts.

8. Reference Files

9. Stop-and-Check (confirm each item before outputting the report)

• Every Must-Fix item points to a specific location and verbatim excerpt — no verdicts from a single word alone.
• Every Must-Fix item states "consequence of not fixing"; un-writable ones were demoted.
• No facts, numbers, quotes, or cases were added that do not exist in the original.
• Unverified items are listed separately, not mixed into Must Fix.
• Passed items are listed; coverage is visible.
• The report's first screen shows per-tier counts + per-dimension distribution (not straight into details).
• Same-tier issues are grouped into one table — no flat listing.
• No issue double-counted by multiple rules.
• Must-Fix ≤ 5; if exceeded, verdict changed to "Recommend redo."
• No vague, unactionable accusations like "feels very AI" / "doesn't read human" — every finding lands on a concrete excerpt.
• When one excerpt is hit by multiple rules, merged into one finding via the arbitration ladder (see rules-residue-zh.md / rules-residue-en.md).
• English / mixed artifacts: RZE-xx (English residue) and L7 · FTC (English truth-in-advertising) overlay-loaded, reported separately from RZ-xx / L1 — not merged, not overwritten.
• Images / combined artifacts: genre judged first; every verdict carries the three elements "location + feature + conflict"; the unexplainable went to Unverified.
• Images / combined artifacts: E2/E3/E4/E8 not misplaced into Must Fix (natively capped at 0.8 certainty).
• Images / combined artifacts: Category H produced an anchor-alignment table; "unable to cross-check" was not counted as passed.

Adjudication skeleton informed by: qu-ai-wei (five-part pattern method), Humanizer-zh (preservation boundaries), no-ai-slop (detect-only contract), ai-slop-detector (rule grading and density gating), kill-ai-slop (visual trope checklist). Rules are rebuilt for their native language context, not translated across.