@amityco/social-plus-vise 1.4.0 → 1.4.1

package/CHANGELOG.md

@@ -8,6 +8,12 @@ The format is loosely based on [Keep a Changelog](https://keepachangelog.com/en/
 
 No changes yet.
 
+## 1.4.1 — 2026-07-02
+
+### Changed
+- **README evidence refreshed to the current real-E2E scorecard.** The npm README now leads with the retained July 2026 cross-platform scorecard: Vise average **91.9** vs no-Vise average **46.6** (+45.3 points), with Android/iOS/Web runtime-proof rows and the same methodology boundary as `benchmarks/real-e2e/RESULTS.md`.
+- **Outdated benchmark figures removed from the shipped README.** Earlier SDK-compliance, design-conformance, feed-completeness, and Engagement Intelligence dogfood numbers were removed from the public README because they were measured on older builds/protocols and should not be mixed with the current scorecard.
+
 ## 1.4.0 — 2026-07-02
 
 **Entry-state flow + alignment gate.** `vise plan`/`init` now run a persona-free entry-state flow ahead of the build. A broad, multi-surface goal expands into a **blueprint** — the surfaces and experience to build — and for that journey `vise init` requires a human **blueprint sign-off** before it scaffolds anything: the *alignment gate* ("are we building the right thing?"), distinct from and complementary to the deterministic compliance gate ("did we build it the right way?"). The sign-off is a SELECT, not a rubber stamp — when the solution path is undecided it hard-picks `sdk` / `uikit` / `hybrid` first and surfaces candidate experiences, persists the choice, re-arms if the path/journey/design source changes, and cannot be bypassed. (A single-surface init is gated by intake/clarify, not the blueprint sign-off.)

package/README.md

@@ -139,22 +139,29 @@ Vise can ingest your aesthetic from an HTML/CSS prototype (`vise design extract`
 
 ## Evidence
 
-Vise's measured effect is on whether an agent builds the *whole* outcome — pagination, empty and error states, the optional capabilities you asked for — or just the happy path. The durable finding across benchmarks is the **mechanism**, not any single percentage: a checked verification loop beats the same rules pasted into a prompt. What Vise reliably changes is the **stopping condition** — the agent isn't done when the code compiles; it's done when the local contract is green, attested, or blocked on an explicit customer decision.
+Vise's measured effect is on whether an agent builds the *whole* outcome: runtime-visible social surfaces, natural target discovery, SDK setup/login, selected capabilities, design fit, correctness, and the app's own build/runtime proof. The durable finding is the **mechanism**: a checked verification loop changes the stopping condition. The agent is not done when the code compiles; it is done when the local contract is green, attested, runtime-proven, or explicitly blocked on a customer decision.
 
-**Matched comparisons** — Vise's verification loop vs a docs-only baseline, same task and agent:
+**Current real-E2E scorecard** — latest post-improvement Vise runs vs the strongest usable no-Vise controls captured for the same platform families:
 
-| Benchmark | What's measured | Docs-only | With Vise |
-|---|---|---|---|
-| **SDK compliance** | correctness slices passed (greenfield; 9 independent slices, single run per cell, two agents) | 6/9 (67%) | **9/9 (100%)** |
-| **Design conformance** | the brand's *named* design tokens used vs hardcoded values (ambiguous brief, n=3) | ~72% | **~90%** |
+| Platform | Scenario | Vise overall | No-Vise control | Lift | Runtime proof |
+|---|---|---:|---:|---:|---:|
+| **Android** | market-pulse | **94.3** | 31.2 | **+63.1** | 9/9 |
+| **iOS** | luminary-club | **90.1** | 43.0 | **+47.1** | 9/9 |
+| **Web** | pulse-studio | **91.4** | 65.6 | **+25.8** | 8/9 |
+
+Average Vise score: **91.9**. Average no-Vise score: **46.6**. Average lift: **+45.3 points**.
 
-The mechanism is the durable claim: a checked loop beats the same rules in the prompt.
+The scorecard is retained in [`benchmarks/real-e2e/RESULTS.md`](benchmarks/real-e2e/RESULTS.md). It measures capability activation, product outcome, design integration, correctness, efficiency, and runtime evidence. The remaining Web gap was chat runtime proof; community, feed, composer, comment preview, reactions, profile/follow, moderation, and SDK runtime proof were present.
 
-**Feed completeness (capabilities selected):** when the optional capabilities — image, poll, edit — are selected up front, the Vise arm completed **97–100%** of an 11-item feed checklist across three agents (Cursor/Composer, Claude/Sonnet, Codex/GPT). This is *absolute* completeness from answering Vise's capability questions — **not** a lift over a baseline.
+What changed in the winning Vise arms:
 
-**Engagement Intelligence advisory quality:** a 24-case held-out dogfood set supports the advisory positioning: expected available strategies surfaced in **17/17** cases, multiple strategy options were visible in **20/20** strategy-offer cases, rationale/tradeoff evidence was visible in **24/24**, and no-fit plus availability-gated cases preserved their boundaries. The same run retained **4 ranking-calibration concerns**, so this is **not** a claim of top-1 ranking confidence, autonomous strategy selection, business-outcome lift, or a calibrated score.
+- Runtime readiness is treated as a gate before dogfood proof: identity, credentials, target discovery, tenant data, and proof consent must be resolved or explicitly deferred.
+- Agents are steered away from hardcoded `communityId`, `postId`, `channelId`, and `targetId` values; targets must come from host state, SDK queries, SDK create results, or safe fixture state.
+- Runtime errors block proof instead of being treated as cosmetic.
+- Native scoring recognizes real Android/iOS SDK setup and login patterns, not only surface-level type declarations.
+- Generated artifacts are filtered during secret checks so benchmark evidence does not pollute product scoring.
 
-**Boundaries:** these are social.plus's own measurements on greenfield work — self-reported, no third-party audit, measured on earlier builds, and not universal claims. Negative results travel with them: no measured advantage on day-2 bug fixing, and the design-conformance arms were measured at different times (the figure is the robust by-name-token signal, not pixel perfection).
+**Boundaries:** these are social.plus's own measurements, not a third-party audit. The scorecard is a current-loop result, not a same-minute paired A/B rerun: the Vise column uses final post-improvement runs, while the no-Vise column uses the best available controls from earlier benchmark cells. For a formal public benchmark, rerun both arms from clean fixtures under a frozen protocol and score them with the same scorer version. Older README benchmark figures were removed because they came from earlier builds and older protocols.
 
 <sub>Cursor, Claude, Codex, GitHub Copilot, VS Code, and other product names are trademarks of their respective owners; social.plus is not affiliated with or endorsed by them. Benchmark figures are from social.plus's own measurements.</sub>
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@amityco/social-plus-vise",
-  "version": "1.4.0",
+  "version": "1.4.1",
   "description": "Skill-guided deterministic CLI for social.plus SDK integration assistance.",
   "license": "SEE LICENSE IN LICENSE",
   "type": "module",