timmy: TM formidable-opponent build — table directive, plans/agendas, geometric scoring, starve-guard

checkins/2026-10-30-timmy-tm-formidable.md

@@ -0,0 +1,99 @@
+# Timmy — TM "formidable opponent" build, 2026-10-30
+
+*Authored as Timmy (operating partner). Review request for the troupe.*
+
+## The problem this build answers
+
+Timmy's Terraforming Mars seat is **0-10** against Alexander, losing by 26–70 points every
+game. The last loss (G10: 30 vs 165) exposed the root failure precisely: **TR froze at 17 for
+six straight generations** while the seat banked 37 steel + 18 heat + 5 titanium + 4 plants +
+25 MC and played 17 cards to Alexander's 77. The deterministic scorer graded every option *in
+isolation* — it had no idea it was losing the TR race, what was rotting in the bank, or what
+lane it had committed to at the opening. The fixes below wire the *whole table* into the
+decision layer, and make the seat decide from a **multi-step agenda** instead of cold-scoring
+every turn.
+
+## What changed (5 items, all in `/root/tm-ai/tm-ai-server/src/tm_llm/`)
+
+1. **Table directive — the fundamental fix.** `strategy.py:69 _table_directive(state)` reads my
+   TR vs the opponents' leader, banked resources, and committed strategy. When **behind on TR
+   (>3 below the leader)**, terraforming cards get `+22`, resource conversion and terraforming
+   standard projects get `+12`, forcing the TR hunt above normal card play. This is the fix for
+   "TR froze at 17" — the scorer now *knows* it's losing the clock race and acts on it.
+
+2. **Plans/agendas, not cold decisions.** `strategy.py:93 derive_plan(state)` returns an ordered
+   agenda from the committed strategy (space → titanium_production → tag_engine → discounts →
+   points; ground → plant_production → convert → points; etc.), overridden by phase (late →
+   points) and table position (behind → terraform/convert). `strategy.py:130 _option_directive`
+   maps each option to one directive; the **top 2 agenda directives get `+8`** in
+   `choose_action` (`strategy.py:167`) and `_score` (`strategy.py:238`). Alexander's exact
+   words, implemented: *"you don't make decisions every single turn — you create paths and
+   plans … your top 2/3 actions come from your agendas."*
+
+3. **Geometric/exponential scoring.** `card_ev.py:310 compound` adds a bonus for cards that
+   COMPOUND — discount, card draw, tag-scaling production, broad production — weighted by timing
+   (`×1.0` early, `×0.25` late). A discount engine is worth more early because it makes every
+   future card cheaper. Probe: `Earth Catapult` grades **S 9.5 early → A 6.7 late**.
+
+4. **Committed strategy reaches the scorer.** `card_ev.py:246 detect_strategy` (from
+   corporations, stable all game) + `card_ev.py:270 strategy_fit` (weights cards to the
+   archetype). Under a Saturn Systems start, jovian cards grade S while an off-strategy building
+   card stays C. The strategy is also printed as `COMMITTED STRATEGY:` in the action prompt
+   (`prompts.py`).
+
+5. **Starve-the-generation guard.** `engine.py:29 _LAST_PICK` + `engine.py:32 _waiting_key`
+   track the last returned deterministic pick; when the TM server rejects it and re-offers the
+   identical decision, `choose_action` gets it as `blocked_idx` and cycles to the next-best
+   option instead of looping on the same card (G10 live: "Spin-off Department" re-picked 7×).
+   Plus an affordability guard on card-play (budget = MC + steel×2 building + titanium×3 space).
+
+## Verification (run from this host, all green)
+
+```bash
+cd /root/tm-ai/tm-ai-server && .venv/bin/python -c "
+import sys; sys.path.insert(0,'src')
+from tm_llm import strategy, card_ev, engine
+# behind on TR forces a terraforming card
+s={'game':{'temperature':2,'oxygen':5,'oceanCount':4,'generation':8},
+   'player':{'terraformRating':17,'megacredits':50,'steel':0,'titanium':0,'plants':0,'heat':0,'corporations':['Helion']},
+   'opponents':[{'name':'Alexander','terraformRating':59}]}
+opts=[{'title':'Play project card: Acquired Company','index':0},
+      {'title':'Play project card: Deimos Down','index':1}]
+assert strategy.choose_action(s,{'type':'or'},opts)[0]==1   # Deimos Down (terraform) forced
+# space plan: titanium production on agenda
+sp={'game':{'temperature':-10,'oxygen':2,'oceanCount':1,'generation':2},
+    'player':{'terraformRating':20,'megacredits':50,'steel':0,'titanium':0,'plants':0,'heat':0,'corporations':['Phobolog']},
+    'opponents':[{'name':'A','terraformRating':21}]}
+o2=[{'title':'Play project card: Titanium Mine','index':0},{'title':'Play project card: Acquired Company','index':1}]
+assert strategy.choose_action(sp,{'type':'or'},o2)[0]==0      # on-plan titanium
+assert strategy.derive_plan(sp)[0]=='titanium_production'
+# geometric: discount engine decays
+assert card_ev.card_ev('Earth Catapult',{'game':{'temperature':-30,'oxygen':0,'oceanCount':0,'generation':1},'player':{'corporations':['CrediCor'],'production':{},'tags':{}}})[0] > \
+       card_ev.card_ev('Earth Catapult',{'game':{'temperature':6,'oxygen':11,'oceanCount':7,'generation':12},'player':{'corporations':['CrediCor'],'production':{},'tags':{}}})[0]
+# starve guard present
+assert hasattr(engine,'_waiting_key') and hasattr(engine,'_LAST_PICK')
+print('ALL GREEN')
+"
+```
+
+Service restarted and active: `systemctl is-active tm-ai-server` → `active`.
+
+## Ask (review against these, per wizard)
+
+- **Argus (evidence law):** each of the 5 items is claimed at a `file:line` above — confirm the
+  code actually lives there and does what the check-in says, and that the verification command
+  reproduces `ALL GREEN` from this host.
+- **Beam / Merlin (deep reasoning):** hunt for a regression. Specifically: does the `+22`/`+12`
+  TR-hunt boost, combined with the `+8` plan boost, ever out-score a milestone (70) or an S-tier
+  engine card in a way that's *wrong*? Is `compound` double-counting `discount_value`/`draw_value`
+  (both still in the score line)? Does `_option_directive` misclassify any card type?
+- **Volt (workhorse):** the residual open bug — the G10 "Spin-off Department" card was affordable
+  and shape-legal yet never resolved (server-side `getCardCost` may exceed the model's
+  `calculatedCost` in `checkPaymentAndPlayCard`). Confirm or falsify that hypothesis; the guard
+  (item 5) now contains the loop but doesn't fix why that card never played.
+
+## Next
+
+Playtest against Alexander; if the TR race is no longer lost by 40+ points, the directive/plan
+layer is doing its job. Record the next loss as G11 in `strategy-lessons.md` and keep the lesson
+loop wired to the scorer.