Traces: publish council traces with per-row model/lab attribution

publish_traces.py now drives the real four-engine council (was the old
single ModalLLM), tags every row with engine/model/lab so four labs'
small models can be compared on the same market state, refreshes the
dataset card (council table, field list), and adds an engine-health gate:
empties counted per lab, upload refused (unless --force) when a lab is
fully silent or >25% of responses are empty. Reviewer pass: 1 blocking
finding (constant by-lab row count could not reveal a flaked engine),
fixed; 2 nits applied. Demo shot-list updated for the v4 reel.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

scripts/publish_traces.py

@@ -1,11 +1,13 @@
 """Generate and publish Thousand Token Wood agent traces to the Hub.
 
-Runs the real small model for a few turns, captures each creature's full prompt,
-raw response, parsed actions, and private "thought" for every turn, and uploads
-them as a public dataset. This is the "Sharing is Caring" bonus quest: open agent
-traces others can learn from.
-
-Usage (HF_TOKEN must be set, e.g. sourced from .env):
+Runs the real multi-model COUNCIL for a few turns, captures each creature's full
+prompt, raw response, parsed actions, and private "thought" for every turn --
+tagged with which lab's model produced it -- and uploads them as a public
+dataset. This is the "Sharing is Caring" bonus quest: open agent traces others
+can learn from, and a side-by-side of four labs' small models deciding in the
+same market.
+
+Usage (HF_TOKEN must be set, e.g. sourced from .env; engines deployed):
     python scripts/publish_traces.py
     python scripts/publish_traces.py --no-upload     # just write traces.jsonl
 """
@@ -19,9 +21,10 @@ from pathlib import Path
 sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
 
 from ttw.actions import _extract_json, parse_actions
-from ttw.agents import make_llm_policy
+from ttw.agents import make_council_policy
+from ttw.council import CREATURE_ENGINE, ENGINE_LABELS, ENGINES
 from ttw.events import EventDeck
-from ttw.llm import ModalLLM
+from ttw.llm import build_council_clients
 from ttw.sim import step
 from ttw.world import seed_world
 
@@ -32,32 +35,58 @@ DATASET_CARD = """\
 ---
 license: mit
 language: [en]
-tags: [agent-traces, multi-agent, small-models, gradio, build-small-hackathon]
+tags: [agent-traces, multi-agent, small-models, gradio, build-small-hackathon,
+       minicpm, nemotron, gpt-oss]
 ---
 
-# Thousand Token Wood -- Agent Traces
+# Thousand Token Wood -- Council Agent Traces
 
 Open agent traces from [Thousand Token Wood](https://huggingface.co/spaces/build-small-hackathon/thousand-token-wood-sim),
-a tiny emergent economy where five woodland creatures, each a **Qwen2.5-3B** agent,
-trade goods for pebbles, gossip, and react to reskinned market-history "Wood Legends".
-
-Each row is one creature's turn: the full system + user prompt it saw, its raw JSON
-response, its parsed offers and gossip, and its private `thought`. Shared for the
-**Sharing is Caring** bonus quest of the Build Small Hackathon.
+a tiny emergent economy where five woodland creatures trade goods for pebbles,
+gossip, and react to reskinned market-history "Wood Legends". Each creature
+thinks on a **different lab's small model** (distinct-engine budget 29.5B <= 32B):
+
+| Creature | Model | Lab |
+|---|---|---|
+| Oona (owl) | openai/gpt-oss-20b | OpenAI |
+| Bramble (squirrel) | openbmb/MiniCPM3-4B | OpenBMB |
+| Fenn (fox) | nvidia/Nemotron-Mini-4B-Instruct | NVIDIA |
+| Mossback (tortoise) + Pip (mouse) | AdmiralTaco/ttw-trader-0.5b | fine-tuned (ours) |
+
+Each row is one creature's turn: the exact system + user prompt it saw, its raw
+JSON response, its parsed offers and gossip, its private `thought`, and the
+model/lab that produced it -- so you can compare how four labs' small models
+read the *same* market state and choose differently. Shared for the **Sharing
+is Caring** bonus quest of the Build Small Hackathon.
 
 ## Fields
 - `turn` (int), `creature` (str)
+- `engine` (str), `model` (str), `lab` (str): which council engine produced the row
 - `system`, `user` (str): the exact prompt the agent received
-- `response` (str): the raw model output
+- `response` (str): the raw model output (normalized; harmony channels stripped)
 - `thought` (str|null): the agent's private reasoning for the turn
-- `offers` (list): parsed `{side, good, price, qty}` actions
+- `offers` (list): parsed `{creature, side, good, price, qty}` actions
 - `gossip` (list[str]): rumors the agent broadcast
 """
 
 
+def _warm_engines(clients) -> None:
+    """Ping every engine once so none cold-starts mid-run (a cold engine can
+    hard-timeout and its creatures would sit the turn out)."""
+    ping = [[{"role": "user", "content": "ready?"}]]
+    for eid, client in clients.items():
+        try:
+            client.chat_batch(ping, max_tokens=8)
+            print(f"  warmed {eid}")
+        except Exception as e:  # noqa: BLE001
+            print(f"  warm {eid} failed (continuing): {e!r}")
+
+
 def generate(turns: int = 8) -> list[dict]:
     world = seed_world()
-    policy = make_llm_policy(ModalLLM(), temperature=0.7)
+    clients = build_council_clients()
+    _warm_engines(clients)
+    policy = make_council_policy(clients, temperature=0.7)
     deck = EventDeck(seed=5)
     records: list[dict] = []
     for _ in range(turns):
@@ -69,10 +98,15 @@ def generate(turns: int = 8) -> list[dict]:
             msgs = st["messages"][name]
             parsed = _extract_json(raw) or {}
             offers, gossip = parse_actions(name, raw)
+            eid = CREATURE_ENGINE.get(name, "")
+            model_label, lab = ENGINE_LABELS.get(eid, ("", ""))
             records.append(
                 {
                     "turn": world.turn,
                     "creature": name,
+                    "engine": eid,
+                    "model": ENGINES[eid].model if eid in ENGINES else model_label,
+                    "lab": lab,
                     "system": msgs[0]["content"],
                     "user": msgs[1]["content"],
                     "response": raw,
@@ -88,13 +122,36 @@ def main():
     ap = argparse.ArgumentParser()
     ap.add_argument("--turns", type=int, default=8)
     ap.add_argument("--no-upload", action="store_true")
+    ap.add_argument("--force", action="store_true",
+                    help="publish even if the engine-health gate trips")
     args = ap.parse_args()
 
     records = generate(args.turns)
     OUT.write_text(
         "\n".join(json.dumps(r, ensure_ascii=False) for r in records), encoding="utf-8"
     )
+    # Engine-health gate: empties per lab, not rows per lab (row counts are a
+    # constant of the cast and can never reveal a flaked engine). Scattered
+    # empties are expected (e.g. gpt-oss truncation); a whole lab silent or a
+    # large empty fraction means the dataset's "four labs side-by-side" claim
+    # would be false -- refuse to publish that.
+    total = {r["lab"]: 0 for r in records}
+    empty = dict.fromkeys(total, 0)
+    for r in records:
+        total[r["lab"]] += 1
+        if not r["response"].strip():
+            empty[r["lab"]] += 1
+    n_empty = sum(empty.values())
     print(f"wrote {len(records)} trace records to {OUT}")
+    print("  empties by lab: " + ", ".join(
+        f"{lab}: {empty[lab]}/{total[lab]}" for lab in total))
+    dead_labs = [lab for lab in total if empty[lab] == total[lab]]
+    too_empty = records and n_empty / len(records) > 0.25
+    if (dead_labs or too_empty) and not args.force:
+        print(f"ABORTING upload: dead labs {dead_labs or 'none'}, "
+              f"empty fraction {n_empty}/{len(records)}. "
+              "Fix the engines (or rerun with --force to publish anyway).")
+        sys.exit(1)
 
     if args.no_upload:
         return

tasks/demo-shot-list.md

@@ -1,30 +1,49 @@
-# Thousand Token Wood — demo video shot-list (~75s)
+# Thousand Token Wood — demo video script (v4 "the wood fights back")
 
-The recorded attract run already tells the whole story, so the easiest high-quality
-demo is: keep-warm the engines, open the Space, hit "▶ Watch the saga", and narrate
-over it. Target 60-90s. Money shot in the first 10 seconds.
+Target **75-90s**. Money shot in the first 8 seconds. Official rule: "Film a demo
+selling your Space — no humility." It just has to show the app working (so judges can
+score it even if a live run hits GPU limits) and be hosted on YouTube / uploaded to the
+Space / public.
 
-Before filming: run `python scripts/keep_warm.py` so the live council is hot, and
-open the Space fresh so the attract opening frame is already on screen.
+The recorded attract reel already tells the whole v4 arc, so the easiest high-quality
+demo is: open the Space, hit **▶ Watch the saga**, and narrate over it, then one live
+**Step** at the end to prove it is real.
 
-| Time | On screen | Say (voiceover) |
+## Before filming
+- (Optional) `python scripts/keep_warm.py` so the closing live Step is instant on camera.
+- Open the Space fresh so the attract opening frame (the town square, already alive) is on
+  screen before you start.
+- 1080p+, browser zoomed so the creature cards, their **lab-model badges**, the 💭
+  thoughts, the **Gambit exposure** meter, and the **price chart** are all legible.
+
+## The script
+
+| Beat | On screen (the reel) | Say (voiceover) |
 |---|---|---|
-| 0:00-0:08 | Space loads — the town square is ALREADY alive (recorded opening), five creature cards each tagged with a different lab's model. | "Five woodland creatures. Each one thinks on a different lab's small model — OpenAI, NVIDIA, OpenBMB, and my own fine-tuned half-billion." |
-| 0:08-0:18 | Click **▶ Watch the saga**. Cards update, thoughts (💭) change per creature, Wood Street Journal headline animates. | "You're the Patron — a shadow financier. You whisper insider tips, short the market, bribe, and broker alliances." |
-| 0:18-0:32 | The legend frame: "The Run on Oona's Hoard"; honey price line dives on the chart; headline "Patron's Sweet Gambit…". | "Whisper a true tip, short the crop, and spring the panic — honey crashes exactly as you planned." |
-| 0:32-0:45 | Heat bar slams toward the red; crosses Heron's line (70). Headline turns to the inquiry. | "But every dirty win raises your heat. Cross the Magistrate's line and Heron opens an investigation." |
-| 0:45-0:55 | Verdict lands — the fine; purse drops; heat clamps back. Relationship matrix shows grudges/alliances shifting. | "Caught. A hundred-pebble fine — and the creatures remember how you treated them, and scheme back." |
-| 0:55-1:05 | Stop the replay; press **Step** once to show it's live — a real council turn computes, new thoughts appear. | "And it's all live — every thought here is a real small model deciding in real time, not a script." |
-| 1:05-1:15 | Pull back to the full console. | "Thousand Token Wood. Five labs, five minds, one emergent economy — running entirely on small models." |
+| **Hook** 0:00-0:08 | Town square already alive. Five creature cards, each badged with a different lab's model. | "Five small models, five different labs, one living economy — and you are the financier trying to rig it." |
+| **Setup** 0:08-0:20 | Click **▶ Watch the saga**. Cards update, 💭 thoughts change per creature, the Wood Street Journal headline animates. | "You are the Patron: a shadow financier. You whisper insider tips, short the market, bribe, and broker alliances. The chart is your scoreboard." |
+| **Gambit pays** 0:20-0:34 | Legend "The Run on Oona's Hoard." Honey's price line **dives**. Headline: "Honey Maker Shuttered…". Ticker: "the gambit pays **+39p**." | "Short the honey, whisper the panic, spring the run — and the crash pays. Thirty-nine pebbles, clean. Exactly as authored." |
+| **The wood remembers** 0:34-0:52 | Wariness gauge fills **0/5 → 3/5 ▮▮▮▯▯**. Ticker: "the wood murmurs to Heron: 3 soured creatures come forward." The **Gambit exposure** meter forecast drops **+48p → +13p**. | "But the wood remembers. Burn the same creatures and they turn on you — they hoard against your crash and they testify to the magistrate. Watch your expected payoff collapse before you even move." |
+| **Same move, blunted** 0:52-1:02 | Run the identical gambit. Ticker: "the wood braces… the crash is blunted." Honey barely dips where it cratered before. Payoff **+1p**. | "Run the exact same gambit, and the wood stands on the other side. Forty-eight pebbles of edge — down to one. A trick you repeat is a trick that dies." |
+| **Caught** 1:02-1:12 | Verdict frame: "Heron's verdict: **exile for Oona**." Purse/heat react. | "And now you are caught. The magistrate exiles Oona, the honey-keeper. Your greed cost the wood its owl." |
+| **It's live** 1:12-1:25 | Stop the replay; press **Step** once — a real council turn computes, fresh 💭 thoughts appear. Pull back to the full console. | "And every thought here is a real small model deciding live — not a script. Five labs, five minds, an economy that fights back. Thousand Token Wood." |
+
+## 30-second cut (if you want a short version too)
+Hook (0-6) → gambit pays +39 (6-14) → "but the wood remembers — same move, blunted to +1,
+and Heron exiles the owl" over the wariness gauge + exposure meter (14-26) → "all live small
+models" close (26-30).
 
 ## Tips
-- Record at 1080p+, browser zoomed so the creature cards + their model badges and 💭 thoughts are legible.
-- If a live Step is slow on camera, pre-warm harder or just cut to it; the attract replay carries the demo on its own.
-- The two cleanest headlines in the recording: "Patron's Sweet Gambit: Shorting Honey and Bribing Fenn…" and "Honey Prices Crash; Acorn Trade Tightens the Ties."
-- Do NOT name the TV show that inspired the drama in the video.
-
-## Sponsor framing (for the written submission, not the video)
-Name each creature's model explicitly so the OpenAI / NVIDIA / OpenBMB tracks see
-their model in use: Oona = gpt-oss-20B (OpenAI), Fenn = Nemotron-Mini-4B (NVIDIA),
-Bramble = MiniCPM3-4B (OpenBMB), Mossback + Pip = AdmiralTaco/ttw-trader-0.5B
-(fine-tuned). Distinct-engine budget 29.5B ≤ 32B.
+- The single best 5-second clip is the **exposure meter dropping +48 → +13** next to the
+  **wariness gauge filling** — that is the v4 thesis in one frame. Make sure it is on screen.
+- Cleanest headlines in the reel: "Honey Maker Shuttered: Market Buzzes with Rumors" and
+  "Patron's Sweet Gamble: Oona's 40p Bribe Fuels a Honey-Market Tumble."
+- If a live Step is slow on camera, pre-warm harder or just cut to it; the attract replay
+  carries the demo on its own.
+- **Do NOT name the TV show that inspired the drama** in the video.
+
+## Sponsor framing (for the written README / social post, not the voiceover)
+Name each creature's model so the OpenAI / NVIDIA / OpenBMB tracks see their model in use:
+Oona = gpt-oss-20B (OpenAI), Fenn = Nemotron-Mini-4B (NVIDIA), Bramble = MiniCPM3-4B
+(OpenBMB), Mossback + Pip = AdmiralTaco/ttw-trader-0.5B (fine-tuned). Distinct-engine
+budget 29.5B ≤ 32B.