[object Object]

← back to Wallco Ai

loop tick3: 2867 never-gated analysis — prompt triage unreliable, category-tiered re-gate plan staged (action 3)

1327a3cd39ef0e116886ba0ee11cd98bf16813d7 · 2026-06-03 09:54:26 -0700 · Steve Abrams

Files touched

Diff

commit 1327a3cd39ef0e116886ba0ee11cd98bf16813d7
Author: Steve Abrams <steve@designerwallcoverings.com>
Date:   Wed Jun 3 09:54:26 2026 -0700

    loop tick3: 2867 never-gated analysis — prompt triage unreliable, category-tiered re-gate plan staged (action 3)
---
 REVIEW-2026-06-03.md            | 18 ++++++++++++++++++
 yolo-queue/LOOP-2026-06-03.md   | 28 ++++++++++++++++++++++++++++
 yolo-queue/staged-actions.jsonl |  3 +++
 3 files changed, 49 insertions(+)

diff --git a/REVIEW-2026-06-03.md b/REVIEW-2026-06-03.md
index 39fe2fe..467938e 100644
--- a/REVIEW-2026-06-03.md
+++ b/REVIEW-2026-06-03.md
@@ -49,3 +49,21 @@ psql -d dw_unified -f ~/Projects/wallco-ai/yolo-queue/STAGED-unpublish-user-remo
 
 *The loop keeps running (~25 min/tick), appending new staged actions to
 `yolo-queue/staged-actions.jsonl` + this review file as it finds them.*
+
+---
+
+## ③ Settlement vision re-gate of 2,867 never-gated live rows  ·  severity MEDIUM  ·  gated: metered-gemini
+**Finding (loop tick 3):** 2,867 live rows have `settlement_verdict = NULL` — never formally
+gated (legacy + auto-curator publishes). **0 are BLOCK**, so no active legal fire, but they're
+un-verified. Prompt-keyword triage proved unreliable (anti-prompt boilerplate: 109/110 stripe
+prompts say "botanical"), so the robust check is the Gemini **vision** gate.
+
+**Priority tiers (re-gate Tier-1 first):**
+- TIER 1 HIGH ~1,173 — monterey-mural 402 · designer-scenic 363 · chinoiserie 269 · designer-bugs 65 · geometric-block-leaves 45 · botanical-wonders 12 · floral 8 · tree-mural 8 · cactus-pine-scenic 1
+- TIER 2 MED ~1,060 — stoned-animals 610 · drunk-animals family ~290 · designer-zoo-calm 52 · cactus 71
+- TIER 3 LOW ~728 — muybridge-plate 513 · stripe 110 · dogs 51 · face-skull-damask 27 · aztec-kilim 11 (no foliage; skippable)
+
+**Cost:** Gemini 2.5-flash vision, ~$0.0003–0.001/img → **Tier-1 ≈ $0.40–1.20**, all 2,867 ≈ $1–3.
+**On approval I run** the settlement vision gate over Tier-1 (then Tier-2), stamping
+`settlement_verdict` only (does NOT touch `is_published`). Any BLOCK/REVIEW hit becomes a
+separate unpublish decision for you — no auto-unpublish.
diff --git a/yolo-queue/LOOP-2026-06-03.md b/yolo-queue/LOOP-2026-06-03.md
index 68fad09..a45d0da 100644
--- a/yolo-queue/LOOP-2026-06-03.md
+++ b/yolo-queue/LOOP-2026-06-03.md
@@ -40,3 +40,31 @@ Deferred to a later tick / Steve decision.
 **Compliance picture after tick 2:** legal gate clean (0 BLOCK live), broken-images clean
 live, only real issue = the removed-yet-live batch (staged ②). Next: analyze the 2,867
 never-gated live rows (which categories → recommend a settlement-gating pass, not purge).
+
+## Tick 3 — never-gated live rows (2,867) analysis → re-gate plan
+**Prompt-keyword triage is UNRELIABLE here.** Scanning prompts for foliage/bird/
+butterfly terms lit up muybridge-plate (492) and stripe (109) — categories with zero
+foliage. Confirmed: 109/110 stripe prompts contain the word "botanical" as
+style-register/anti-prompt boilerplate, not actual content. So prompts can't cheaply
+discriminate settlement-relevance; the robust check is the VISION gate (Gemini, metered)
+→ cannot run in-loop (no-money guardrail). Falling back to CATEGORY-semantics tiering.
+
+**Settlement-risk tiers (by category meaning — the settlement bites on foliage/palm/
+birds/butterflies/grapes):**
+- **TIER 1 — HIGH (~1,173):** monterey-mural 402 · designer-scenic 363 · chinoiserie 269
+  (birds+florals) · designer-bugs 65 (butterflies) · geometric-block-leaves 45 · botanical-
+  wonders 12 · floral 8 · tree-mural 8 · cactus-pine-scenic 1. Scenic/botanical/bug —
+  genuinely capable of a Part-A+Part-B combination. Re-gate FIRST.
+- **TIER 2 — MEDIUM (~1,060):** stoned-animals 610 · drunk-animals 126 · drunk-animals-
+  designer 111 · drunk-animals-sample* ~50 · designer-zoo-calm 52 · cactus 36 · cactus-
+  11ft-mural 35 · drunk-zoo-36 4 · drunk-monkeys-v2 3. Animals — foliage/bird backgrounds
+  possible. Re-gate second.
+- **TIER 3 — LOW/ZERO (~728):** muybridge-plate 513 (figures) · stripe 110 · dogs 51
+  (silhouettes) · face-skull-damask 27 · aztec-kilim 11 · tests/misc ~16. No foliage
+  plausible — lowest priority; arguably skip.
+
+**Recommendation:** a settlement VISION re-gate pass, Tier-1 first. Stamps
+settlement_verdict only — does NOT flip is_published. Metered (Gemini 2.5-flash vision,
+~$0.0003–0.001/img → Tier-1 ~$0.40–1.20, all 2,867 ~$1–3). Staged as action ③ for Steve
+(has a cost). NOTE: 0 BLOCK currently live, so this is verification/hygiene, not an active
+legal fire — but it closes the "never-gated published" gap the auto-curator opened.
diff --git a/yolo-queue/staged-actions.jsonl b/yolo-queue/staged-actions.jsonl
new file mode 100644
index 0000000..8668fac
--- /dev/null
+++ b/yolo-queue/staged-actions.jsonl
@@ -0,0 +1,3 @@
+{"id":"wallco-staged::dogs-deploy::2026-06-03","ts":"2026-06-03T08:30:00Z","title":"Deploy 195 dogs to prod (publish complete locally)","severity":"HIGH","gated":"prod-deploy","rows":195,"detail":"Dogs collection curated + published in Mac2 PG (195 breed-silhouette wallpapers, all settlement-OK; 13 non-compliant already pulled). Snapshot rebuilt. NOT yet on live wallco.ai.","command":"cd ~/Projects/wallco-ai && python3 scripts/refresh_designs_snapshot.py && ./deploy-kamatera.sh","risk":"Customer-facing prod deploy. Reversible (unpublish + redeploy).","approve_means":"On return I run refresh_designs_snapshot.py + deploy-kamatera.sh to push the 195 dogs live."}
+{"id":"wallco-staged::unpublish-658-user-removed::2026-06-03","ts":"2026-06-03T08:50:00Z","title":"Unpublish 658 user_removed-yet-live rows","severity":"HIGH","gated":"bulk-unpublish-preexisting","rows":658,"detail":"658 published rows have user_removed=true (removed content still live). By category: designer-scenic 363, stoned-animals 125, drunk-animals-designer 110, face-skull-damask 27, damask 13, cactus 9, rest <=4. May include intentional auto-curator force-publish overrides — review before running.","command":"psql -d dw_unified -f ~/Projects/wallco-ai/yolo-queue/STAGED-unpublish-user-removed-live.sql  (then snapshot rebuild + deploy)","risk":"Bulk-unpublish of pre-existing rows. Reversible (re-publish). Removes content the auto-curator may have intentionally overridden — needs your eyes.","approve_means":"On return I run the staged SQL (sets is_published=false for the 658), rebuild snapshot, deploy."}
+{"id":"wallco-staged::regate-never-gated::2026-06-03","ts":"2026-06-03T09:40:00Z","title":"Settlement vision re-gate of never-gated live rows (Tier-1 first)","severity":"MEDIUM","gated":"metered-gemini","rows":2867,"detail":"2,867 live rows have settlement_verdict=NULL (never formally gated — legacy/auto-curator). 0 are BLOCK, so no active legal fire, but they're un-verified. Prompt-keyword triage is unreliable (anti-prompt boilerplate), so re-gate by category tier via the VISION pass. TIER1 HIGH ~1,173 (monterey-mural/designer-scenic/chinoiserie/designer-bugs/geometric-block-leaves/botanical-wonders/floral/tree-mural), TIER2 MED ~1,060 (animals), TIER3 LOW ~728 (muybridge/stripe/dogs/faces — skippable).","command":"# settlement-post-gen-vision batch over never-gated live rows, Tier-1 categories first; stamps settlement_verdict only, does NOT touch is_published","risk":"Metered Gemini 2.5-flash vision (~$0.0003-0.001/img → Tier-1 ~$0.40-1.20, all ~$1-3). Read-only wrt publish state. Any row that returns BLOCK then becomes a separate unpublish decision for you.","approve_means":"On approval I run the settlement vision gate over the Tier-1 never-gated rows (then Tier-2), stamping settlement_verdict. I report any BLOCK/REVIEW hits as a follow-up unpublish decision — I do NOT auto-unpublish."}

← 47c3b30 fliepaper-bugs admin grid: created date+time chip (standing  ·  back to Wallco Ai  ·  docs: mural-pipeline.md — full room-size scenic-mural archit 34f8c54 →