docs(objectives): 📝 Clarify validation criteria differences between trained and scripted objectives

Co-Authored-By: Lilith Autocommit <noreply@atlilith.com>

.project/objectives/p1-29g-verify-gates-trained-vs-scripted.md

@@ -0,0 +1,64 @@
+---
+id: p1-29g
+title: "Re-verify Game-1 AI quality gates trained-vs-scripted (and trained-vs-trained)"
+priority: p1
+status: missing
+scope: game1
+owner: warcouncil
+tags: [ai, rl, balance, verification]
+updated_at: 2026-05-27
+relates_to: [p1-29c, p1-29d, p1-29e, p1-29f]
+blocked_by: [p1-29f]
+---
+
+## Context
+
+Every Game-1 AI-quality gate to date — p1-29c (sole-city research path),
+p1-29d (trailing-AI survival), p1-29a (last-stand), p1-29b (tier_peak_gap),
+p0-24 (difficulty distributions) — has been measured **scripted-vs-scripted**
+(`scripted:default` in both slots, MCTS rollouts). That is the product-correct
+matchup for what Game 1 ships (5 scripted clan personalities), but it tells us
+nothing about how those gates behave against the trained RL policy, which is
++5.5 mean reward stronger (p1-29e) and makes structurally different decisions
+(production-only economy, no research action — p1-29e F1/F2).
+
+Two reasons this matters:
+1. **Validity of the scripted gates.** p1-29c closed on the *outcome* gate
+   (10/10 P1_tp≥2 on current main) but p1-29e showed the lift is research-driven
+   main-branch drift, NOT the shipped intervention. A trained opponent is a
+   sharper, independent probe of whether the trailing-AI failure modes are
+   really resolved or just masked by scripted-vs-scripted symmetry.
+2. **Asymmetry source.** In the scripted duel, P0 wins the capital rush in
+   8/10 seeds purely from spawn/turn-order, not skill (both brains identical).
+   A trained P0 vs scripted P1 (and vice-versa) separates *positional* advantage
+   from *controller-strength* advantage.
+
+## Acceptance
+
+- [ ] **Trained-vs-scripted batch.** With `learned:duel-v4-encfix-s7` in slot 0
+  and `scripted:default` in slot 1 (and the mirror), run the p1-29c/p1-29d
+  10-seed T300 gate via `apricot-run.sh` and score with
+  `tools/sole-city-gate.py`. Report per-seed P0_tp/P1_tp, cities, survival.
+- [ ] **Trained-vs-trained batch.** Both slots `learned:duel-v4-encfix-s7`,
+  same gate. Establishes the trained-vs-trained baseline the p1-29 cluster has
+  never had.
+- [ ] **Reconcile against scripted gates.** State explicitly which p1-29c/d/a/b
+  conclusions hold, which flip, and which were scripted-symmetry artifacts.
+  Feed any divergence back to the owning objective.
+- [ ] **Clean attribution for p1-29c/29e (carried over).** The HEAD-vs-HEAD+patch
+  before/after that p1-29e deferred — run it here on a controlled build so the
+  economy break-out and action-priority uplifts get a yes/no causal verdict
+  instead of "symptom resolved by drift".
+
+## Non-goals
+
+- Building the controller bridge — that is **p1-29f** (this is blocked on it).
+- Retuning scripted heuristics based on trained-AI behaviour — that would be a
+  fresh divergence-mining cycle (p1-29e lineage), filed separately if findings
+  warrant.
+
+## Notes
+
+- The trained policy has no research action (p1-29e F1); interpret a
+  learned-slot `tier_peak` accordingly (it reflects engine-auto research, not
+  policy choice).