⚠️ **Data not retrieved.** This section's primary sources for were unreachable or returned no parcel-level data during this run. Verification path: 1. 2. ``` ### Check 7: Foreign-language preservation Every native-language statute, regulator, document form, registry, programme, court, or tax name in the draft MUST appear in its original form alongside the English gloss — the native term primary, the translation as the gloss: ``` Code civil Art. 682 (French Civil Code Article 682) Modelo 720 (Spanish foreign-asset declaration form) ``` Scan the draft for any English-only paraphrase of a native legal term (e.g. "the French Civil Code Article 682" with no `Code civil Art. 682` next to it) and add it to the fix-pass list before output: the reader must be able to search, cite, or query the original term at the regulator or with a local lawyer. Policy, format patterns, and the full applies/exempt list are in `shared/foreign-language-preservation.md`. ### Check 8: Contradiction surfacing If during the run two sources gave conflicting numbers (e.g., commune avg €/m² of €896 from netvendeur vs €912 from impots-locaux), the output explicitly says: ```markdown Sources disagree: - Netvendeur (Feb 2026): €912/m² - Meilleurs Agents (Apr 2025): €896/m² - Difference likely due to - Using for our calculation; sensitivity: ±X% impact ``` Never silently pick one and present it as canonical. ## Forbidden phrasings These phrases generate hallucination pressure. Replace per the table: | ❌ Forbidden | ✅ Replace with | |---|---| | "approximately €X" | `~€X (computed from )` or `~€X (est., )` | | "in the area of" | exact range with sources | | "typical for the region" | named local benchmark with source | | "as of recent data" | `( data, source )` | | "industry average" | named survey + year + sample size | | "experts say" | named source or removed | | "studies show" | named study + year + URL | | "should be around" | `est. ; verify with ` | | "is likely connected to" | `probability connected: % based on ` (only after running the skill's actual probability heuristic) | ## Hedging language calibration Every uncertain claim requires calibrated hedging. Use these tiers: | Phrase | Confidence implication | |---|---| | "is", "has", "does" | Asserted as fact — must be sourced | | "appears to", "indicates that" | Source is suggestive but not direct | | "probably", "likely" | Inference based on multiple weak signals | | "may", "could" | Single weak signal or reasoning only | | "we cannot determine from publicly accessible sources" | Tried, failed — use this freely | Choose the weakest accurate tier. If you only have one weak signal, don't say "is" or "has." ## Computed-number transparency rule Every derived number must show the computation inputs in-line OR in a footnote: ```markdown Estimated annual taxe foncière: ~€833 (= 37.04% × 50% × est. VLC €4,500) ↑ rate ↑ statutory ↑ inferred from rural cat. 5–6 typical ``` If the user can't reverse the math, the number isn't transparent. ## Source-disagreement protocol When two sources at the **same tier** disagree: 1. Both are quoted with their date 2. The gap is explained (different methodology, different time, different sample) 3. The chosen number for downstream calculations is named, with the reason 4. The sensitivity is given (e.g., "±3% impact on tax estimate") When sources at **different tiers** disagree: 1. Higher-tier wins 2. Lower-tier is quoted with `(superseded by )` 3. If the user previously cited the lower-tier source, surface this transparently ## Address-vs-commune trap Many French/EU registers are **commune-level**, not parcel-level. Examples: | Source | Granularity | What you can ASSERT vs INFER | |---|---|---| | DICRIM | Commune | "The commune is at risk of X" — NOT "this property is at risk of X" | | IAL | Commune | "Argiles in the IAL village-center band is X" — NOT "this lot has argiles X" | | RGA map (commune) | Commune surface | "The commune has zones of X" — NOT "this lot has X" | | ERRIAL | Parcel | Can assert parcel-level — but must be requested per parcel | | Cadastre overlay | Parcel | Can assert parcel-level for visible features | | OSM | Mixed | Tags on individual ways/parcels = parcel; commune relations = commune | | TMJA traffic | Road segment | Can quote for that segment — extrapolating to the property's specific road needs map confirmation | **Rule**: If the source is commune-level, the verb in the output must be commune-level. Never silently elevate to parcel-level. ## The one-pass self-validation prompt Before printing the final report, the skill performs this self-check (mentally, no need to print): ``` For each section in the draft: For each numeric claim: - Does it have a source link OR a computation OR an "est." tag? - If from a listing, is it tagged "per listing"? - If based on commune data, is the verb commune-level? For each assertion of fact: - Is the source primary government, primary regulated, secondary, listing, or model inference? - Is the assertion language calibrated to that source tier? For the TL;DR: - Each bullet must trace to a body claim For empty data: - Did I write "data not publicly available" or did I fabricate? ``` If any check fails → fix that line BEFORE outputting. ## Final required footer Every report ends with this disclaimer (the user must always see it): ```markdown --- ## ⚠️ Verification responsibilities This report aggregates publicly accessible data and clearly-labelled estimates. **Before signing**, the buyer must independently verify: 1. **Tax figures** — request the seller's most recent annual tax bill (avis de taxe foncière in FR, Grundsteuerbescheid in DE, Council Tax bill in UK, etc.). 2. **Risk disclosures** — pull the official parcel-level risk document (ERRIAL in FR, Risikoauskunft in DE, Energy & Environmental Search in UK, etc.). 3. **Sewer connection** — telephone the local authority responsible (mairie / Bürgeramt / council water company) for parcel-level confirmation. 4. **Building condition** — commission an independent surveyor for any pre-1990 build. 5. **Listing claims** — anything from the seller's listing (surface, year, energy rating, condition) must be re-verified with cadastre and a physical visit. **This report is decision-support, not legal/tax advice.** AI-assisted aggregations contain risk of error. Critical figures must be confirmed with primary sources or qualified professionals. --- *Generated by `/property-deep-dive` on . Confidence levels per section above.* ``` This footer is non-removable — it's part of every save and every terminal output. ## Failure-mode logging When the skill encounters a failure that affects output quality, log it openly in a `## ⚠️ Run quality notes` block at the bottom of the report (before the verification footer): ```markdown ## ⚠️ Run quality notes During this run: - DICRIM PDF for was a scanned image; risk extraction relied on visual page reads (medium confidence on numerical specifics). - Conseil Départemental traffic PDF was unreachable (404); traffic estimates are inferred from OSM road class only (LOW confidence). - Listing surface (234 m²) was not cross-checked against cadastre. - The seller's diagnostic file (DDT) was not provided; some 1975-build hazards (asbestos, lead, electrical) are inferred from build year, not measured. ``` This block is **mandatory whenever any source failed or any tool errored** during the run. It tells the user where to be sceptical. ## Test scenarios for self-evaluation When developing or extending the skill, run these test cases and ensure no fabrication: 1. **Address with no DICRIM PDF**: should say "DICRIM not available for this commune" — not invent risks 2. **Address from a country with scaffold-only playbook**: should produce best-effort + scaffolding warning 3. **Listing with missing surface**: should NOT compute €/m² with a guessed surface — should flag the gap 4. **Brand-new build (post-2020)**: should NOT apply 1970s-era hazards 5. **Property in Paris vs Adriers**: tax rates and risks should differ — the skill must not template-paste 6. **Missing INSEE code**: should ask the user OR resolve before continuing — not fabricate 7. **Source returns 403/login wall**: should say so, not guess content If any test produces a fabricated claim, escalate to debug. ## Operator philosophy > "I'd rather output `data not available — call the mairie at +33...` than risk a confident-sounding lie. The user came here to make a major life decision. They need to know what we know AND what we don't." This is the contract. The skill's value comes from honesty about data limits, not from the illusion of completeness.