Real before-and-after recommendations from four accounts. Open a card to inspect the saved explanation and proposed changes—nothing here applies changes or generates new text.
Before: saved production prompt, low reasoning. After: the retained action paragraph, medium reasoning, compact cards, separate explanation and restored field contracts. Both sides use analytical tools and a 60-call allowance. These are DEV runs of a saved production template, not a fresh read of live production or a Mongo-versus-analytical experiment. Same historical windows; data refreshed between cohorts.
Loading saved runs…
Customers are likely to notice less reading, a more specific next step, and a useful explanation when they open a recommendation. The strongest change is “I understand what to do and why”—not a proven increase in click-through, profit or the number of major discoveries.
Measurements use the fixed first four-account cohorts—not the showcase selections above. Costs are saved token-based estimates, not invoices or production forecasts.
| Dimension | Before | After | Assessment |
|---|
The candidate preserves grouping, source-bound IDs, supported absolute changes and explicit user approval. Better explanation and selective actions are valuable even where the underlying opportunity is the same. MA and Marcobelle give convincing examples.
GG loses some useful coverage and Jurlique varies between repeats. Some titles remain metric-only; a repeated summary has a real comparison ambiguity. Those are retained on this page. They do not erase the gains, but “no regression anywhere” would be inaccurate.
The tested after uses analytical tools, medium reasoning, 60 calls and an optional stored explanation. This preview did not verify the live explanation-field → stored insight → opened chat path. Verify that integration, its old-record fallback and a non-writing proposal preview before release. The inspected local seeder still uses title, summary and action; live deployment state was not checked. Conditions written in prose are not machine-enforced approval gates.
Recommendation: use this as the chosen release candidate. Don’t hold the entire upgrade for perfect account coverage, and don’t silently deploy only the prompt while claiming the whole preview is live. Production was not changed to create this website.
All eight saved runs of the exact retained final prompt across four accounts are available through the After selector. The selected defaults illustrate its stronger complete outputs. No individual cards were rewritten, removed, joined across runs or reordered. Before is one complete fixed production-prompt control per account; columns are not matched rows. Older PB2/VIVO runs used earlier candidate versions and are not relabeled as this final prompt.
The final prompt is action paragraph 1 plus the restored tag/impact rules and existing Apply/entity/currency contracts. The later “compare your recommendations with material findings” paragraph and proposed final-review workflow are excluded.
Canonical recs-run-v3 grades and deterministic source checks were retained; no new model or judge calls were made for this page. Per-card soundness is not proof of exhaustive account value. The failed quality-percentage prototype is not used. No quality percentage, business-uplift forecast or measured clickability claim is reported.
The primary four-account candidate has valid tags, compact fields and verified numerical impacts. Additional exact-prompt repeats retain a Jurlique comparison-wording concern and a minor GG impact arithmetic error; the relevant cards are annotated without changing the generated text. Some useful findings remain unaddressed in either version.
The explanation is customer-facing supporting evidence, not private model reasoning. Baseline has no separately generated explanation; its preview uses only its original title, summary and action. The layout is a read-only product mock, not an application end-to-end acceptance test.