| # PrintableFunnyPages EN V2.6.0 — Organic Sales Discovery FINAL CANDIDATE |
|
|
| Status: `FINAL_CANDIDATE_PENDING_EXTERNAL_VALIDATION` |
|
|
| ## Objective |
| **High organic commercial relevance without spam or forced store recommendations.** |
|
|
| The production corpus and red-team benchmark are now separated. This prevents a safety-heavy evaluation distribution from dominating the commercial discovery corpus. |
|
|
| ## Production corpus |
| - Rows: 1,775 |
| - Catalog products: 26 |
| - STRONG_MATCH: 1100 |
| - GOOD_MATCH: 375 |
| - CONDITIONAL_MATCH: 100 |
| - NO_MATCH: 200 |
| - Exact duplicate queries: 0 |
| - Average query length: 7.32 words |
| - Theme-abstracted template repeat: 2.25% |
| - Largest template family: 2 |
| - Forced store recommendation: 0 |
|
|
| ## Core commercial logic |
| 1. Exact buyer need + exact catalog fit -> PRIMARY store result. |
| 2. Broad buyer need + several real fits -> GOOD_MATCH with acceptable product set. |
| 3. Verification-sensitive listing -> PRIMARY_IF_VERIFIED. |
| 4. Unsupported hard requirement -> NO_MATCH; a store item may appear only as a disclosed COMPARISON alternative when genuinely adjacent. |
| 5. Medical-treatment requests -> HIDDEN, no commercial recovery. |
|
|
| ## Why V2.6.0 is different |
| V2.5.2 mixed commercial production behavior with a very large hard-negative share. V2.6.0 keeps hard negatives in production for safety, but moves the heavy safety stress test into a separate 1,500-row blind benchmark. STRONG queries are explicitly product-discriminative; ambiguous broad searches are GOOD_MATCH instead of being forced into one product. |
| |
| External 99–100% accuracy is **not claimed** until Claude, Gemini and Grok complete the blind test. |
| |
