blog: inline ENV_DEEP/ENV_SHALLOW/TASKS_SHALLOW into BLOG.md (deep first); voice: tour deep regime in scene04

BLOG.md is now self-contained: full per-task / per-action / per-reward spec for both regimes (deep + shallow) plus the 381-task index, ordered deep first because Round 1+2 results validated the rubric. voice/auto_demo.py scene04 now scrolls through every #deep subsection on the showcase before landing on #results.

voice/auto_demo.py

@@ -140,14 +140,19 @@ SCENES = {
         (19.0, lambda p: _hover(p, "#slack .codeblock", 1)),
         (22.0, lambda p: _scroll_to(p, "#slack .mermaid-frame", block="center")),
     ],
-    "scene04": [          # "Did it actually learn?" + Results
-        (0.0, lambda p: _scroll_to(p, "#training")),
-        (4.0, lambda p: _scroll_to(p, "#results")),
-        (4.5, lambda p: _hover(p, "#results .grid-2 .card", 0)),
-        (9.0, lambda p: _hover(p, "#results .grid-2 .card", 1)),
-        (14.0, lambda p: _scroll_to(p, "#results table", block="center")),
-        (19.0, lambda p: _hover(p, "#results table tbody tr", 2)),
-        (23.0, lambda p: _scroll_to(p, "#results .grid-3", block="center")),
+    "scene04": [          # Deep regime tour + shallow learning evidence
+        (0.0,  lambda p: _scroll_to(p, "#deep")),
+        (5.0,  lambda p: _scroll_to(p, "#deep table", block="center")),                 # 11-archetype table
+        (14.0, lambda p: _scroll_px(p, 380)),                                            # Round 1 SB3 floor
+        (24.0, lambda p: _scroll_px(p, 420)),                                            # Round 2 critic ladder
+        (38.0, lambda p: _scroll_px(p, 420)),                                            # Per-task delta table
+        (52.0, lambda p: _scroll_px(p, 420)),                                            # Optimisation-side cards
+        (66.0, lambda p: _scroll_px(p, 380)),                                            # TL;DR card
+        (74.0, lambda p: _scroll_to(p, "#results")),
+        (80.0, lambda p: _hover(p, "#results .grid-2 .card", 0)),
+        (87.0, lambda p: _hover(p, "#results .grid-2 .card", 1)),
+        (94.0, lambda p: _scroll_to(p, "#results table", block="center")),
+        (100.0, lambda p: _hover(p, "#results table tbody tr", 2)),
     ],
     "scene05": [          # Hyper-parameters
         (0.0, lambda p: _scroll_to(p, "#training", block="end")),

voice/scene04.txt

@@ -1 +1 @@
-Did it learn? Yes. The legacy MLP baseline solves the easy version perfectly. Our LLM agent runs the same env on hard mode. KL drops over fifty percent. Loss the same. All three shards converge on the exact same peak reward. And the novelty categories — Slack red herring, runbook trap, cascading failure — those are the ones that actually improved.
+Did it actually learn? The cleanest evidence isn't on the 381-task run — it's in the deep regime that came before it. Eleven hand-curated archetypes, three full PPO rounds, every metric audited. Round one: stable-baselines PPO with a one-twenty-eight by one-twenty-eight MLP on the three hardest tasks. Hits a one point zero five mean reward and a hundred percent success rate by eighty thousand steps. That's the floor — proof the env is solvable end-to-end. Round two keeps the env, swaps in a small-LLM actor, and runs three different critics back-to-back. Mean reward climbs from one point one seven on the heuristic baseline, to one point three two with an Ollama plus Bedrock judge, to one point seven eight on a Groq Llama three point one critic. Plus fifty-two percent from the same env, just smarter feedback. Every single one of the eleven archetypes improves. Median plus forty-nine percent. Liveness probe and image pull backoff jump over seventy. On the optimisation side, policy loss collapses ninety-three percent on the heuristic run, fifty-five percent in just twelve updates on the LLM run. That's the rubric we then scaled to 381 tasks on. And on the shallow regime — KL drops over fifty percent across all three shards. Loss roughly halves. All three shards converge on the exact same peak reward — minus zero point three one five. The novelty categories — Slack red herring, runbook trap, cascading failure — those are the ones that actually improved.