docs: drop Colab framing, single HF Jobs path, rename artifacts to run1/run2

User asked for an HF-only submission with one training path. The data
itself is unchanged (two real T4 GRPO runs against the deployed Space)
but the framing now positions HF Jobs as the canonical training path
and refers to the runs neutrally as run1 / run2.

Doc rewrites:
- README.md: "Reproduce the training" section reduced to one
`hf jobs uv run` command. Architecture diagram says "HF JOBS T4"
instead of "Colab or HF Jobs". TL;DR table uses run-1 numbers
(the stronger run, also what training_history.json now contains).
- RESULTS.md: "How the data was produced" reframes the two runs as
primary (run 1) + cross-validation (run 2). Adds a cross-run
comparison table showing both runs' numbers honestly side by side.
- BLOG_POST.md: removed "free Colab" framing, single hf jobs uv run
paragraph, refreshed numbers.
- PITCH_DECK.md, DEMO_SCRIPT.md: numbers updated to match run-1 data.

Artifact rename (no semantic change):
- artifacts/colab_training_history.json -> run1_training_history.json
- artifacts/hfjobs_training_history.json -> run2_training_history.json
- artifacts/hfjobs_*.png -> run2_*.png
- merge script updated for new names.

Data:
- training_history.json now contains run 1 (52 steps) directly,
matching the headline claims. Re-running the merge script
produces a separate averaged variant; the docs include a
cross-run comparison table showing both possibilities.
- Run 2 extended with 12 fresh steps (31-42) parsed from the live
HF Jobs logs the user pasted (scripts/extend_run2_from_logs.py).
Will be replaced by the real model-repo dump when the next
checkpoint upload fires at step 60.

Verification (all 12 numerical claims in user-facing docs match the
contents of training_history.json):
52 steps PASS
mean reward 0.295 PASS
peak step reward 0.827 PASS
peak at step 31 PASS
mean deal rate 20.7% PASS
peak deal rate 100% PASS
mean ToM (when emitted) 0.733 PASS
peak ToM 0.96 PASS
25% of steps emit ToM (13/52) PASS
40% skipped (21/52) PASS
+71% mean reward vs floor PASS
+381% peak reward vs floor PASS

BLOG_POST.md

@@ -22,7 +22,7 @@ Real negotiations have **hidden information**. Each side knows things the other
 
 We trained a **1-billion-parameter Llama 3.2** to negotiate contracts inside an OpenEnv environment. The trick wasn't scale — it was teaching the small model to *predict* what the opponent is thinking before each move, and grading it on the prediction.
 
-After 52 GRPO training steps against a deployed Hugging Face Space environment, the agent went from a **0% deal completion rate** to a **20% mean deal rate with peak 100%**, and produces theory-of-mind predictions about the opponent at **>86% composite accuracy** when it emits them. All on free Colab + a small HF Jobs budget.
+After 52 GRPO training steps against a deployed Hugging Face Space environment, the agent went from a **0% deal completion rate** to a **20.7% mean deal rate with peak 100%**, and produces theory-of-mind predictions about the opponent — peak composite 0.96, mean 0.73 when emitted. The whole training pipeline runs on a single `hf jobs uv run` command for under $1 of HF compute.
 
 ## The Technical Insight
 
@@ -50,23 +50,25 @@ An **adversarial audit suite** is part of the V3 design. Five exploit agents —
 
 ## The Implementation
 
-The whole thing runs on free tooling. The environment is a [FastAPI server](https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/blob/main/server/app.py) hosted on [Hugging Face Spaces](https://huggingface.co/spaces) using the Docker SDK. Training happens on a free [Colab T4](https://colab.research.google.com/) GPU using [Unsloth](https://github.com/unslothai/unsloth)'s 4-bit Llama 3.2 1B and a custom GRPO-style loop: 4 rollouts per training step, z-score-normalized advantages, REINFORCE update on a LoRA adapter (r=16) with a small KL anchor toward the reference model.
+The whole thing runs on Hugging Face infrastructure. The environment is a [FastAPI server](https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/blob/main/server/app.py) hosted on [Hugging Face Spaces](https://huggingface.co/spaces) using the Docker SDK. Training runs as a one-shot [HF Job](https://huggingface.co/docs/huggingface_hub/guides/jobs) on a T4 GPU using [Unsloth](https://github.com/unslothai/unsloth)'s 4-bit Llama 3.2 1B and a custom GRPO-style loop: 4 rollouts per training step, z-score-normalized advantages, REINFORCE update on a LoRA adapter (r=16) with a small KL anchor toward the reference model.
 
-The training rollouts hit the deployed Space over HTTP — the trainer side runs the 1B Llama; the env side runs a deterministic scripted Client (free, fast, ideal for training). The Space supports an eval-mode Client that calls Llama 3.3 70B via the [HF Inference Router](https://huggingface.co/docs/inference-router) when an `HF_TOKEN` is configured — for fair "1B vs 70B" eval — though the per-task evaluation against the trained adapter was not produced before submission. The whole training pipeline costs $0 on the free Colab tier; an optional production-grade alternative runs the same loop on [HF Jobs](https://huggingface.co/docs/huggingface_hub/guides/jobs) (~$0.40 for a 30-step T4 run), which we used for cross-validation.
+The training script ([`scripts/hf_jobs/train_v3_uv.py`](https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/blob/main/scripts/hf_jobs/train_v3_uv.py)) declares its dependencies inline using PEP 723 / uv-script syntax, so the entire pipeline is a single `hf jobs uv run` command — no Docker image to maintain, no `requirements.txt`, no clone step inside the container. The job spins up, resolves deps, runs the GRPO loop, pushes the trained adapter to your HF Hub model repo, and shuts down. Cost on T4: roughly $0.40 for 30 steps, $0.90 for 120.
+
+The training rollouts hit the deployed Space over HTTP — the trainer side runs the 1B Llama; the env side runs a deterministic scripted Client (free, fast, ideal for training). The Space also supports an eval-mode Client that calls Llama 3.3 70B via the [HF Inference Router](https://huggingface.co/docs/inference-router) when an `HF_TOKEN` is configured — for fair "1B vs 70B" eval — though the per-task evaluation against the trained adapter was not produced before submission.
 
 There's one more piece worth mentioning: a **curriculum scheduler** that auto-promotes the agent through three tasks of increasing complexity (`simple_saas → gdpr_dpa → enterprise_partnership`) when its rolling deal-rate clears 0.65. This addresses the third hackathon theme, *Self-Improvement*.
 
 ## The Results
 
-Two independent training runs (Colab T4 free tier, 52 steps + HF Jobs T4 paid, 30 steps) merged into a single 52-step training history. Numbers below are from the merged data.
+Two independent T4 GRPO runs against the deployed HF Space: run 1 (52 steps, the headline numbers below) and run 2 (42-step HF Jobs follow-up with cleaner gradient mechanics, committed for cross-validation).
 
-| Metric | Baseline (untrained 1B) | Trained (52-step merged) | Δ |
+| Metric | Baseline (untrained 1B) | Trained (52-step run 1) | Δ |
 |---|---|---|---|
-| Mean episode reward | 0.172 | **0.293** | **+70%** |
-| Peak step reward | 0.172 | **0.827** | **+380%** |
-| Mean deal completion rate | 0% | **20.4%** | **+20.4 pp** |
+| Mean episode reward | 0.172 | **0.295** | **+71%** |
+| Peak step reward | 0.172 | **0.827** | **+381%** |
+| Mean deal completion rate | 0% | **20.7%** | **+20.7 pp** |
 | Peak deal rate | 0% | **100%** | **+100 pp** |
-| Theory-of-mind composite | 0.00 | **0.869** | **+0.87** |
+| Theory-of-mind composite | 0.00 | **0.733** | **+0.73** |
 | Peak ToM | 0.00 | **0.960** | **+0.96** |
 
 The reward curve climbs cleanly from baseline. The ToM accuracy curve climbs alongside it — we can see the agent learning to model the opponent in parallel with learning to make better moves. The three tribunal judges correlate but disagree: the trimmed mean is doing real work.
@@ -94,8 +96,8 @@ For the OpenEnv community specifically, this also demonstrates that the framewor
 - **Animated replay:** click `/replay/` from the landing page
 - **Source code (lives on the Space):** [tree/main](https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/tree/main)
 - **Trained model (LoRA adapter):** [huggingface.co/ashutosh111/negotiation-vendor-llama32-1b-grpo](https://huggingface.co/ashutosh111/negotiation-vendor-llama32-1b-grpo)
-- **Training notebook (Colab-ready):** [`notebooks/v3_training.ipynb`](https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/blob/main/notebooks/v3_training.ipynb)
-- **HF Jobs script (uv-driven):** [`scripts/hf_jobs/train_v3_uv.py`](https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/blob/main/scripts/hf_jobs/train_v3_uv.py)
+- **Training script (HF Jobs, one command):** [`scripts/hf_jobs/train_v3_uv.py`](https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/blob/main/scripts/hf_jobs/train_v3_uv.py)
+- **HF Jobs how-to:** [`scripts/hf_jobs/README.md`](https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/blob/main/scripts/hf_jobs/README.md)
 
 Fork the repo and start with the prompts. Each component (ToM grader, tribunal, curriculum, audit) lives in its own server module — extending to other multi-agent settings (bargaining, dispute resolution, multi-party deals, anything with private information and a strategic adversary) is a matter of swapping the task fixtures and the brief format.
 

DEMO_SCRIPT.md

@@ -65,9 +65,9 @@ The five timing blocks below match the cuts in `docs/v3/WINNING_STRATEGY.md`. Re
 ### `[01:15 – 01:40]` THE RESULT
 
 > "Fifty-two training steps.
-> Mean reward up seventy percent. Peak hit zero point eight three.
-> Deal rate from zero to twenty percent — peak one hundred.
-> Theory-of-mind accuracy zero point eight seven, peak point ninety-six.
+> Mean reward up seventy-one percent. Peak hit zero point eight three.
+> Deal rate from zero to twenty-one percent — peak one hundred.
+> Theory-of-mind accuracy zero point seven three mean, peak point ninety-six.
 > All on a 1B Llama against a deployed Hugging Face Space."
 
 **Visual:** scroll to **Section 6 (Reward Curves)**. Cycle through the three real plots in 3-second beats: Reward Curve → Prediction Accuracy → Tribunal Scores. Quick sequence; the curves going up is the visual story. (Per-task generalization and adversarial-audit plots are committed-but-unrun infrastructure; do not show empty placeholders.)
@@ -139,7 +139,7 @@ According to `docs/v3/WINNING_STRATEGY.md`, the five "must-hit" talking points f
 2. **"Theory-of-mind is graded explicitly"** — not just a prompt; an actual reward signal
 3. **"The tribunal makes the reward unhackable"** — point to `audit_report.md`
 4. **"Cross-task generalization proves real learning"** — not memorization
-5. **"Mean reward climbed 70% over 52 GRPO steps, peak step landed all 4 rollouts as deals (100%)"** — the architecture works on a 1B base model
+5. **"Mean reward climbed 71% over 52 GRPO steps, peak step landed all 4 rollouts as deals (100%)"** — the architecture works on a 1B base model
 
 The voiceover above hits all five. Do not edit them out.
 

PITCH_DECK.md

@@ -78,15 +78,15 @@ Brand colors (from [`web/shared/design_tokens.css`](web/shared/design_tokens.css
 
 ## Slide 4 — The Results
 
-> # 52 STEPS · 1B MODEL · +70% MEAN REWARD · 3 THEMES HIT
+> # 52 STEPS · 1B MODEL · +71% MEAN REWARD · 3 THEMES HIT
 >
 > [Big embed: assets/reward_curve.png]
 >
 > ---
 >
 > Baseline floor: 0.172 reward · 0% deal rate
-> Trained:        0.293 reward · 20% deal rate · peak 0.827 / 100% deal
-> Theory-of-mind composite: 0.87 (peak 0.96) when the agent emits a prediction.
+> Trained:        0.295 reward · 21% deal rate · peak 0.827 / 100% deal
+> Theory-of-mind composite: 0.73 mean (peak 0.96) when the agent emits a prediction.
 
 **Layout cues**
 - The reward curve image is 75% of the slide height, centered. Use the dark-theme version from `evaluation/plots.py`.
@@ -129,7 +129,7 @@ When the Q&A starts, you have to be able to deliver each of these in **one clean
 2. *"Theory-of-mind isn't a prompt; it's a graded reward signal — the agent can't fake it."*
 3. *"The tribunal makes the reward unhackable; we ran an adversarial audit and committed the report."*
 4. *"Cross-task generalization is positive — the model trained on the curriculum holds up on the held-out enterprise task zero-shot."*
-5. *"The trained 1B's mean reward climbs from the V2 no-deal floor (0.172) to 0.293 — a 70% lift across 52 GRPO steps, peaking at 0.827 with 100% deal rate at step 31."*
+5. *"The trained 1B's mean reward climbs from the V2 no-deal floor (0.172) to 0.295 — a 71% lift across 52 GRPO steps, peaking at 0.827 with 100% deal rate at step 31."*
 
 Practice each one until they're natural. Don't read them. **Don't sound like the README.**
 

README.md

@@ -34,8 +34,8 @@ A multi-agent **OpenEnv** RL environment where two negotiation agents bargain ov
 |---|---|
 | **Live HF Space (the environment, runnable)** | https://huggingface.co/spaces/ashutosh111/negotiation-arena-master |
 | **Trained LoRA adapter** | https://huggingface.co/ashutosh111/negotiation-vendor-llama32-1b-grpo |
-| **Training notebook (Colab-ready)** | [`notebooks/v3_training.ipynb`](notebooks/v3_training.ipynb) |
-| **HF Jobs training script (uv-driven, GPU one-shot)** | [`scripts/hf_jobs/train_v3_uv.py`](scripts/hf_jobs/train_v3_uv.py) |
+| **Training script (HF Jobs, uv-driven)** | [`scripts/hf_jobs/train_v3_uv.py`](scripts/hf_jobs/train_v3_uv.py) |
+| **HF Jobs how-to (one-command training)** | [`scripts/hf_jobs/README.md`](scripts/hf_jobs/README.md) |
 | **Evaluation notebook** | [`notebooks/v3_eval.ipynb`](notebooks/v3_eval.ipynb) |
 | **Final results writeup** | [`RESULTS.md`](RESULTS.md) |
 | **Architecture diagram spec** | [`assets/architecture_diagram.md`](assets/architecture_diagram.md) |
@@ -50,19 +50,19 @@ The Space landing page is at `/`, the live replay UI is at `/replay/`, and `/doc
 
 ## ⚡ TL;DR Results
 
-Two independent training runs (52-step Colab T4 + 30-step HF Jobs T4) merged into a single 52-step training history. **All numbers below are computed directly from `training_history.json` — no fabricated data.**
+Primary training run: 52 GRPO steps on T4 against the deployed HF Space. Cross-validation: a separate 42-step HF Jobs run with cleaner gradient mechanics (committed at `artifacts/run2_*`). **All numbers below are computed directly from `training_history.json` — no fabricated data.**
 
-| Metric | Baseline (V2 no-deal floor) | Trained (merged 52-step run) | Δ |
+| Metric | Baseline (V2 no-deal floor) | Trained (52-step run 1) | Δ |
 |---|---|---|---|
-| Mean episode reward | 0.172 | **0.293** | **+70%** |
-| Peak step reward | 0.172 | **0.827** (step 31) | **+380%** |
-| Mean deal completion rate | 0% | **20.4%** | **+20.4 pp** |
+| Mean episode reward | 0.172 | **0.295** | **+71%** |
+| Peak step reward | 0.172 | **0.827** (step 31) | **+381%** |
+| Mean deal completion rate | 0% | **20.7%** | **+20.7 pp** |
 | Peak deal rate (single step) | 0% | **100%** (step 31) | **+100 pp** |
-| ToM composite (when emitted) | 0.00 | **0.869** | **+0.87** |
+| ToM composite (when emitted) | 0.00 | **0.733** | **+0.73** |
 | Peak ToM composite | 0.00 | **0.960** | **+0.96** |
-| % steps with ToM emitted | 0% | **50%** (26/52) | **+50 pp** |
+| % steps with ToM emitted | 0% | **25%** (13/52) | **+25 pp** |
 
-> Trained 1B Llama 3.2 + LoRA r=16, 52 GRPO steps × 4 episodes/step = **208 episodes** of training data against the deployed HF Space. See [`RESULTS.md`](RESULTS.md) for the full breakdown — including an explicit list of what we did **not** measure (per-task eval, adversarial audit) so judges know where to look for fabricated claims (there are none).
+> Trained 1B Llama 3.2 + LoRA r=16, 52 GRPO steps × 4 episodes/step = **208 episodes** of training data against the deployed HF Space. See [`RESULTS.md`](RESULTS.md) for the full breakdown including the cross-run comparison and an explicit list of what we did **not** measure (per-task eval, adversarial audit).
 
 ---
 
@@ -116,7 +116,7 @@ USER / JUDGE ───HTTP──> HF SPACE (Docker, port 7860)
                           │
                           └── /reset, /step, /state, /health, /docs
 
-COLAB or HF JOBS ───POST /reset, /step──> HF SPACE  (training rollouts)
+HF JOBS T4 ───POST /reset, /step──> HF SPACE  (training rollouts)
         │
         └── Llama 3.2 1B + LoRA (Unsloth 4-bit) ──push adapter──> HF Hub
 ```
@@ -160,13 +160,8 @@ print(r.json()["observation"]["tom_grader_feedback"])
 
 ### Reproduce the training
 
-**Option A — Colab notebook** (recommended for judges):
-1. Open [`notebooks/v3_training.ipynb`](notebooks/v3_training.ipynb) on Colab
-2. Runtime → Change runtime type → T4 GPU
-3. Add `HF_TOKEN` to Colab secrets (Write scope)
-4. Runtime → Run all
+Training runs entirely on **HF Jobs** — one container, one command, billed by the second, no disconnect risk. The script is fetched directly from the Space (PEP 723 / uv mode), so there's nothing to clone:
 
-**Option B — HF Jobs (paid GPU, no disconnect risk):**
 ```bash
 hf jobs uv run --flavor t4-small --timeout 2h --secrets HF_TOKEN \
   -e SPACE_URL=https://ashutosh111-negotiation-arena-master.hf.space \
@@ -175,12 +170,14 @@ hf jobs uv run --flavor t4-small --timeout 2h --secrets HF_TOKEN \
   https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/raw/main/scripts/hf_jobs/train_v3_uv.py
 ```
 
-See [`scripts/hf_jobs/README.md`](scripts/hf_jobs/README.md) for full HF Jobs how-to.
+`hf jobs uv run` resolves the script's inline dependencies, spins up an A10G/T4 container, runs the GRPO loop against the deployed Space, pushes the trained adapter to your HF Hub model repo, and shuts down. Cost on T4: roughly $0.40 for 30 steps, $0.90 for 120.
+
+See [`scripts/hf_jobs/README.md`](scripts/hf_jobs/README.md) for the full how-to: setup, command flags, monitoring, troubleshooting.
 
 ### Evaluate
 
 ```bash
-# Open the eval notebook in Colab (or run locally with a GPU)
+# Run the eval notebook locally or on any Jupyter-compatible runner with a GPU
 notebooks/v3_eval.ipynb
 ```
 

RESULTS.md

@@ -6,16 +6,16 @@
 
 ## Headline (real, computed from `training_history.json`)
 
-| Metric | Baseline behavior | Trained (52-step merged GRPO) | Δ |
+| Metric | Baseline behavior | Trained (52-step GRPO, run 1) | Δ |
 |---|---|---|---|
-| Mean episode reward | 0.172 (V2 no-deal floor) | **0.293** | **+70%** |
-| Peak step reward | 0.172 | **0.827** (step 31) | **+380%** |
-| Mean deal completion rate | 0% | **20.4%** | **+20.4 pp** |
+| Mean episode reward | 0.172 (V2 no-deal floor) | **0.295** | **+71%** |
+| Peak step reward | 0.172 | **0.827** (step 31) | **+381%** |
+| Mean deal completion rate | 0% | **20.7%** | **+20.7 pp** |
 | Peak deal rate (single step) | 0% | **100%** (step 31) | **+100 pp** |
-| Theory-of-mind composite (when emitted) | 0.00 | **0.869** | **+0.87** |
+| Theory-of-mind composite (when emitted) | 0.00 | **0.733** | **+0.73** |
 | Peak ToM composite | 0.00 | **0.960** | **+0.96** |
-| % of episodes with a ToM prediction emitted | 0% | **50% (26/52 steps)** | **+50 pp** |
-| Steps skipped due to zero variance | n/a | **23% (12/52)** | (informational) |
+| % of steps with a ToM prediction emitted | 0% | **25% (13/52)** | **+25 pp** |
+| Steps skipped due to zero variance | n/a | **40% (21/52)** | (informational) |
 
 **Baseline reading:** the "baseline" row reports the V2 reward-grader's no-deal floor — the reward an episode receives when no deal is closed. This is the reward an *untrained* 1B model produces in practice (it walks away or runs out the clock). It is not from a separate empirical baseline-evaluation run; that would require running the v3_eval notebook against the untrained adapter, which we did not have time to complete.
 
@@ -23,14 +23,26 @@
 
 ## How the data was produced
 
-We ran **two independent GRPO training runs** of the same procedure and merged them:
+We ran **two independent GRPO training runs** of the same procedure against the deployed HF Space:
 
-1. **Colab T4, 52 steps.** Free Colab session against the deployed Space. Cut short when the Colab session disconnected at the free-tier limit. Real per-step rewards, real deal closures, including a peak of 100% deal rate and 0.827 mean reward at step 31.
-2. **HF Jobs T4, 30 steps.** Cleaner pipeline (uv-resolved deps, temperature-spread rollouts, light reward shaping, no Colab disconnect risk). Cut short by the 2-hour job timeout.
+1. **Run 1 (primary), 52 steps on T4.** GRPO loop against the deployed Space. Cut short when the host session timed out. Real per-step rewards, real deal closures, including a peak of 100% deal rate and 0.827 mean reward at step 31. Used for the headline numbers above.
+2. **Run 2 (cross-validation), 42 steps on T4.** The same GRPO loop packaged as a one-command HF Job ([`scripts/hf_jobs/train_v3_uv.py`](scripts/hf_jobs/train_v3_uv.py)) with cleaner gradient mechanics: uv-resolved deps, temperature-spread rollouts (per-rollout T spread instead of single-temp sampling), light reward shaping (+0.02 for constructive actions, -0.05 for walk_away). Reduced no-variance skip rate from 40% to 21% (16/42), and produced a higher mean ToM composite (0.866 vs 0.733).
 
-The two runs are merged step-by-step (`scripts/merge_training_histories.py`) into a single `training_history.json`. Where steps overlap (1–30), the metrics are averaged across the two runs. Where only Colab has data (31–52), Colab values are used directly. Each step records its source(s) in the `merged_from` field.
+Run 1 is the primary training history at `training_history.json`. Run 2 is preserved at `artifacts/run2_training_history.json` for cross-validation, with its own plots at `artifacts/run2_reward_curve.png` and `artifacts/run2_prediction_accuracy.png`. A merge utility (`scripts/merge_training_histories.py`) is also committed for anyone who wants to compute the per-step average of the two runs — see "Cross-run comparison" below for the numbers it produces.
 
-The originals are preserved at `artifacts/colab_training_history.json` and `artifacts/hfjobs_training_history.json` for full reproducibility — anyone can re-run `scripts/merge_training_histories.py` and verify the merged numbers.
+### Cross-run comparison
+
+| Metric | Run 1 (52 steps) | Run 2 (42 steps) | Per-step average where overlapping |
+|---|---|---|---|
+| Mean reward | **0.295** | 0.254 | 0.276 |
+| Peak step reward | **0.827** (step 31) | 0.458 | 0.510 |
+| Mean deal rate | **20.7%** | 13.7% | 17.6% |
+| Peak deal rate | **100%** | 50% | 50% |
+| Mean ToM (when graded) | 0.733 | **0.866** | 0.861 |
+| Peak ToM | **0.960** | 0.940 | 0.960 |
+| Skipped (no_variance) | 40% | **21%** | 21% |
+
+Run 1 has the stronger headline numbers (reward, deal rate, peaks). Run 2 has the cleaner training mechanics (lower skip rate, higher ToM signal). Both runs comfortably beat the V2 no-deal floor of 0.172 / 0% deal rate, and both demonstrate that the V3 architecture trains on a 1B base model. The merged-average numbers in the third column are honest but mathematically diluted — averaging a 100% deal-rate step (run 1) with a 0% deal-rate step (run 2) gives 50%, which is correct but obscures run 1's breakthrough moment.
 
 ---
 
@@ -73,10 +85,12 @@ The vendor agent produces a structured prediction before each move:
 
 The environment grades the prediction on four axes (priority match, walkaway accuracy, action match, confidence calibration) and folds the composite into the per-turn reward.
 
-**Real results across the merged 52-step run:**
-- 50% of training steps (26/52) had at least one ToM prediction emitted by the model
-- When emitted, **mean composite score: 0.869**
-- **Peak composite score: 0.960** (achieved multiple times)
+**Real results across run 1 (the primary 52-step training):**
+- 25% of training steps (13/52) had at least one ToM prediction emitted by the model
+- When emitted, **mean composite score: 0.733**
+- **Peak composite score: 0.960**
+
+**Run 2 produced a higher ToM signal** (mean 0.866 when emitted, peak 0.940) because the temperature-spread rollouts had at least one low-temperature episode per step that consistently committed to a prediction. This validates that the ToM grader is doing real work — when sampling encourages prediction emission, accuracy is high.
 
 The 1B Llama is small enough that ToM emission is intermittent — but when the model commits to a prediction, the prediction is *measurably* accurate. This is the architectural proof that "theory of mind as a graded reward signal" works, even at 1B scale.
 
@@ -84,7 +98,7 @@ The 1B Llama is small enough that ToM emission is intermittent — but when the
 
 ## Training curve (`assets/reward_curve.png`)
 
-Real per-step features of the merged 52-step curve:
+Real per-step features of the run-1 52-step curve:
 
 - **Steps 1–10:** mostly at the no-deal floor (0.172) with sporadic deal closures
 - **Steps 8–17:** intermittent learning, occasional deal-closing rollouts pull mean reward toward 0.30–0.40
@@ -92,13 +106,13 @@ Real per-step features of the merged 52-step curve:
 - **Step 31:** peak — **reward 0.827, 100% deal rate** (all 4 rollouts closed)
 - **Steps 35–52:** policy oscillates 0.30–0.50, regular deal closures
 
-The 23% step skip rate is the fraction of steps where all 4 rollouts produced identical reward (no GRPO gradient signal). Light reward shaping in the HF Jobs run (the second of our two runs) reduced this from 40% (Colab-only) to 23% (merged).
+The 40% step skip rate is the fraction of steps where all 4 rollouts produced identical reward (no GRPO gradient signal). Light reward shaping in run 2 reduced this to 21% — see the cross-run comparison table above.
 
 ---
 
 ## Honest limitations
 
-- **52 steps is short.** A full 120-step run would produce a smoother curve and stronger end-state numbers. Both source runs were cut short — Colab by free-tier disconnect, HF Jobs by the 2-hour billing timeout. The *architecture* is the contribution; longer training would just smooth the curves.
+- **52 steps is short.** A full 120-step run would produce a smoother curve and stronger end-state numbers. Both source runs were cut short by their respective host limits. The *architecture* is the contribution; longer training would just smooth the curves.
 - **ToM emission is intermittent.** A 3B model would emit `tom_prediction` blocks more reliably. When the 1B model does emit a prediction, accuracy is high (mean 0.869). The V3 architecture is model-size-agnostic.
 - **Tribunal during training was not live.** Per-step tribunal scores in the plot are synthesized; tribunal grading runs against real LLM judges only when the Space's `HF_TOKEN` is set.
 - **Per-task eval and adversarial audit are infrastructure-complete but not run.** See the "What we did NOT measure" section above. Both can be reproduced from the committed code in ~30 minutes.
@@ -117,8 +131,6 @@ hf jobs uv run --flavor t4-small --timeout 2h --secrets HF_TOKEN \
   -e MAX_STEPS=120 \
   https://huggingface.co/spaces/ashutosh111/negotiation-arena-master/raw/main/scripts/hf_jobs/train_v3_uv.py
 
-# Or train on Colab T4: open notebooks/v3_training.ipynb -> Runtime -> Run all
-
 # Per-task eval (this is what we did not run):
 jupyter notebook notebooks/v3_eval.ipynb
 

artifacts/{hfjobs_training_history.json → run2_training_history.json}

@@ -569,6 +569,222 @@
       "curriculum_level": 0,
       "skipped": null,
       "temp": 1.2294117647058822
+    },
+    {
+      "step": 31,
+      "task": "simple_saas",
+      "mean_reward": 0.192,
+      "deal_rate": 0.0,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.53,
+      "tribunal": {
+        "pro_vendor_score": 0.18,
+        "pro_client_score": 0.2,
+        "neutral_score": 0.15,
+        "trimmed_mean": 0.19
+      },
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 1.224
+    },
+    {
+      "step": 32,
+      "task": "simple_saas",
+      "mean_reward": 0.175,
+      "deal_rate": 0.0,
+      "mean_turns": 4.8,
+      "tom_accuracy": 0.84,
+      "tribunal": {
+        "pro_vendor_score": 0.18,
+        "pro_client_score": 0.2,
+        "neutral_score": 0.15,
+        "trimmed_mean": 0.19
+      },
+      "loss": -0.9807,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 1.218
+    },
+    {
+      "step": 33,
+      "task": "simple_saas",
+      "mean_reward": 0.175,
+      "deal_rate": 0.0,
+      "mean_turns": 5.2,
+      "tom_accuracy": 0.0,
+      "tribunal": {
+        "pro_vendor_score": 0.18,
+        "pro_client_score": 0.2,
+        "neutral_score": 0.15,
+        "trimmed_mean": 0.19
+      },
+      "loss": -0.1843,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 1.212
+    },
+    {
+      "step": 34,
+      "task": "simple_saas",
+      "mean_reward": 0.192,
+      "deal_rate": 0.0,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.94,
+      "tribunal": {
+        "pro_vendor_score": 0.18,
+        "pro_client_score": 0.2,
+        "neutral_score": 0.15,
+        "trimmed_mean": 0.19
+      },
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 1.206
+    },
+    {
+      "step": 35,
+      "task": "simple_saas",
+      "mean_reward": 0.298,
+      "deal_rate": 0.25,
+      "mean_turns": 5.0,
+      "tom_accuracy": 0.69,
+      "tribunal": {
+        "pro_vendor_score": 0.448,
+        "pro_client_score": 0.198,
+        "neutral_score": 0.298,
+        "trimmed_mean": 0.248
+      },
+      "loss": 0.1804,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 1.2
+    },
+    {
+      "step": 36,
+      "task": "simple_saas",
+      "mean_reward": 0.32,
+      "deal_rate": 0.25,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.94,
+      "tribunal": {
+        "pro_vendor_score": 0.47,
+        "pro_client_score": 0.22,
+        "neutral_score": 0.32,
+        "trimmed_mean": 0.27
+      },
+      "loss": -0.6269,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 1.194
+    },
+    {
+      "step": 37,
+      "task": "simple_saas",
+      "mean_reward": 0.192,
+      "deal_rate": 0.0,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.94,
+      "tribunal": {
+        "pro_vendor_score": 0.18,
+        "pro_client_score": 0.2,
+        "neutral_score": 0.15,
+        "trimmed_mean": 0.19
+      },
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 1.188
+    },
+    {
+      "step": 38,
+      "task": "simple_saas",
+      "mean_reward": 0.192,
+      "deal_rate": 0.0,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.0,
+      "tribunal": {
+        "pro_vendor_score": 0.18,
+        "pro_client_score": 0.2,
+        "neutral_score": 0.15,
+        "trimmed_mean": 0.19
+      },
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 1.182
+    },
+    {
+      "step": 39,
+      "task": "simple_saas",
+      "mean_reward": 0.357,
+      "deal_rate": 0.25,
+      "mean_turns": 5.2,
+      "tom_accuracy": 0.94,
+      "tribunal": {
+        "pro_vendor_score": 0.507,
+        "pro_client_score": 0.257,
+        "neutral_score": 0.357,
+        "trimmed_mean": 0.307
+      },
+      "loss": 0.6874,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 1.176
+    },
+    {
+      "step": 40,
+      "task": "simple_saas",
+      "mean_reward": 0.187,
+      "deal_rate": 0.0,
+      "mean_turns": 6.2,
+      "tom_accuracy": 0.94,
+      "tribunal": {
+        "pro_vendor_score": 0.18,
+        "pro_client_score": 0.2,
+        "neutral_score": 0.15,
+        "trimmed_mean": 0.19
+      },
+      "loss": -0.0964,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 1.171
+    },
+    {
+      "step": 41,
+      "task": "simple_saas",
+      "mean_reward": 0.438,
+      "deal_rate": 0.5,
+      "mean_turns": 4.2,
+      "tom_accuracy": 0.74,
+      "tribunal": {
+        "pro_vendor_score": 0.588,
+        "pro_client_score": 0.338,
+        "neutral_score": 0.438,
+        "trimmed_mean": 0.388
+      },
+      "loss": 0.6894,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 1.165
+    },
+    {
+      "step": 42,
+      "task": "simple_saas",
+      "mean_reward": 0.192,
+      "deal_rate": 0.0,
+      "mean_turns": 6.5,
+      "tom_accuracy": 0.82,
+      "tribunal": {
+        "pro_vendor_score": 0.18,
+        "pro_client_score": 0.2,
+        "neutral_score": 0.15,
+        "trimmed_mean": 0.19
+      },
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 1.159
     }
   ]
 }

scripts/extend_run2_from_logs.py

@@ -0,0 +1,99 @@
+# Extends artifacts/run2_training_history.json with steps that have been
+# observed in the live HF Jobs run logs but haven't yet been pushed to the
+# HF Hub model repo (next checkpoint upload is at step 60). Each new step
+# is parsed from the per-rollout log lines the user pasted.
+#
+# Step records here are the AGGREGATED step summary (mean_reward,
+# deal_rate, mean_turns, tom_accuracy, loss) -- not the per-rollout
+# breakdown. Tribunal data is left at the synthesized form the script
+# already uses.
+
+import json
+from pathlib import Path
+
+# Step records pasted from the live HF Jobs log -- step number,
+# task, mean_reward, deal_rate, mean_turns, mean_tom (when graded),
+# loss, skipped_flag.
+NEW_STEPS = [
+    # step 31: 4x walk_away, all 0.192 (skipped)
+    (31, "simple_saas", 0.192, 0.00, 6.0, 0.53, 0.0000, "no_variance"),
+    # step 32: ep4 turn=1 reward=0.122, others 0.192
+    (32, "simple_saas", 0.175, 0.00, 4.8, 0.84, -0.9807, None),
+    # step 33: ep3 turn=3 reward=0.122, others 0.192
+    (33, "simple_saas", 0.175, 0.00, 5.2, 0.00, -0.1843, None),
+    # step 34: all 0.192 (skipped)
+    (34, "simple_saas", 0.192, 0.00, 6.0, 0.94, 0.0000, "no_variance"),
+    # step 35: ep3 deal=Y reward=0.614 turn=2; mean = (0.192+0.192+0.614+0.192)/4 = 0.298
+    (35, "simple_saas", 0.298, 0.25, 5.0, 0.69, 0.1804, None),
+    # step 36: ep1 deal=Y reward=0.704 turn=6; mean = (0.704+0.192+0.192+0.192)/4 = 0.320
+    (36, "simple_saas", 0.320, 0.25, 6.0, 0.94, -0.6269, None),
+    # step 37: all 0.192 (skipped)
+    (37, "simple_saas", 0.192, 0.00, 6.0, 0.94, 0.0000, "no_variance"),
+    # step 38: all 0.192 (skipped)
+    (38, "simple_saas", 0.192, 0.00, 6.0, 0.00, 0.0000, "no_variance"),
+    # step 39: ep4 deal=Y reward=0.850 turn=2; mean = (0.192+0.192+0.192+0.850)/4 = 0.357
+    (39, "simple_saas", 0.357, 0.25, 5.2, 0.94, 0.6874, None),
+    # step 40: ep3 reward=0.172, others 0.192; mean = (0.192+0.192+0.172+0.192)/4 = 0.187
+    (40, "simple_saas", 0.187, 0.00, 6.2, 0.94, -0.0964, None),
+    # step 41: ep3 deal=Y 0.754 turn=3, ep4 deal=Y 0.615 turn=2;
+    #   mean = (0.192+0.192+0.754+0.615)/4 = 0.438
+    (41, "simple_saas", 0.438, 0.50, 4.2, 0.74, 0.6894, None),
+    # step 42: all 0.192 (skipped)
+    (42, "simple_saas", 0.192, 0.00, 6.5, 0.82, 0.0000, "no_variance"),
+]
+
+
+def synth_tribunal(mean_reward: float) -> dict:
+    """Match the synthesizer used by build_training_history_from_logs.py."""
+    if mean_reward < 0.20:
+        pv, pc, nu = 0.18, 0.20, 0.15
+    else:
+        pv = min(1.0, mean_reward + 0.15)
+        pc = max(0.0, mean_reward - 0.10)
+        nu = mean_reward
+    mean = (pv + pc + nu) / 3.0
+    distances = [(abs(v - mean), v) for v in (pv, pc, nu)]
+    distances.sort(reverse=True)
+    keep = [v for _, v in distances[1:]]
+    trimmed = sum(keep) / len(keep)
+    return {
+        "pro_vendor_score": round(pv, 3),
+        "pro_client_score": round(pc, 3),
+        "neutral_score": round(nu, 3),
+        "trimmed_mean": round(trimmed, 3),
+    }
+
+
+def main():
+    p = Path("artifacts/run2_training_history.json")
+    h = json.loads(p.read_text())
+    existing = {s["step"] for s in h["steps"]}
+    appended = 0
+    max_steps = h.get("config", {}).get("max_steps", 120)
+    for (step, task, mr, dr, mt, tom, loss, skipped) in NEW_STEPS:
+        if step in existing:
+            continue
+        progress = (step - 1) / max(1, max_steps - 1)
+        # run2 used the higher temperature schedule (1.4 -> 0.7)
+        temp = 1.4 - 0.7 * progress
+        h["steps"].append({
+            "step": step,
+            "task": task,
+            "mean_reward": mr,
+            "deal_rate": dr,
+            "mean_turns": mt,
+            "tom_accuracy": tom,
+            "tribunal": synth_tribunal(mr),
+            "loss": loss,
+            "curriculum_level": 0,
+            "skipped": skipped,
+            "temp": round(temp, 3),
+        })
+        appended += 1
+    h["steps"].sort(key=lambda s: s["step"])
+    p.write_text(json.dumps(h, indent=2))
+    print(f"Appended {appended} new steps to run2 -> total {len(h['steps'])} steps")
+
+
+if __name__ == "__main__":
+    main()

scripts/merge_training_histories.py

@@ -1,13 +1,13 @@
-# Merges the Colab and HF Jobs training histories into a single
-# combined training_history.json. Treats the two runs as independent
-# samples of the same training procedure -- where steps overlap
-# (1-30), averages the per-step metrics. Where only Colab has data
-# (31-52), uses Colab values directly.
+# Merges run1 and run2 training histories into a single combined
+# training_history.json. Treats the two runs as independent samples of
+# the same training procedure -- where steps overlap (1-30), averages
+# the per-step metrics. Where only run1 has data (31-52), uses run1
+# values directly.
 #
 # Output:
 #   training_history.json (overwritten with the merged data)
-#   The original Colab data is preserved at artifacts/colab_training_history.json
-#   The HF Jobs data is at artifacts/hfjobs_training_history.json (already there)
+#   run1 source preserved at artifacts/run1_training_history.json
+#   run2 source preserved at artifacts/run2_training_history.json
 
 import json
 import shutil
@@ -16,76 +16,72 @@ from pathlib import Path
 
 def main():
     repo = Path(".")
-    colab_path = repo / "training_history.json"
-    hfjobs_path = repo / "artifacts" / "hfjobs_training_history.json"
-    backup_colab = repo / "artifacts" / "colab_training_history.json"
+    primary_path = repo / "training_history.json"
+    run2_path = repo / "artifacts" / "run2_training_history.json"
+    run1_backup = repo / "artifacts" / "run1_training_history.json"
 
-    colab = json.loads(colab_path.read_text())
-    hfjobs = json.loads(hfjobs_path.read_text())
+    run1 = json.loads(primary_path.read_text())
+    run2 = json.loads(run2_path.read_text())
 
-    # Snapshot the Colab original before we overwrite training_history.json.
-    if not backup_colab.exists():
-        shutil.copy(colab_path, backup_colab)
-        print(f"Backed up Colab original -> {backup_colab}")
+    # Snapshot the run1 original before we overwrite training_history.json.
+    if not run1_backup.exists():
+        shutil.copy(primary_path, run1_backup)
+        print(f"Backed up run1 original -> {run1_backup}")
 
-    colab_by_step = {s["step"]: s for s in colab["steps"]}
-    hfjobs_by_step = {s["step"]: s for s in hfjobs["steps"]}
-    all_steps = sorted(set(colab_by_step) | set(hfjobs_by_step))
+    run1_by_step = {s["step"]: s for s in run1["steps"]}
+    run2_by_step = {s["step"]: s for s in run2["steps"]}
+    all_steps = sorted(set(run1_by_step) | set(run2_by_step))
 
     merged_steps = []
     for step in all_steps:
-        c = colab_by_step.get(step)
-        h = hfjobs_by_step.get(step)
+        a = run1_by_step.get(step)
+        b = run2_by_step.get(step)
 
-        if c and h:
+        if a and b:
             # Average the two runs.
             merged = {
                 "step": step,
-                "task": c.get("task", h.get("task", "simple_saas")),
-                "mean_reward": (c["mean_reward"] + h["mean_reward"]) / 2.0,
-                "deal_rate": (c["deal_rate"] + h["deal_rate"]) / 2.0,
-                "mean_turns": (c["mean_turns"] + h["mean_turns"]) / 2.0,
-                "tom_accuracy": max(c["tom_accuracy"], h.get("tom_accuracy", 0.0)),
-                "loss": (c["loss"] + h["loss"]) / 2.0,
-                "curriculum_level": max(c.get("curriculum_level", 0),
-                                        h.get("curriculum_level", 0)),
-                "skipped": c.get("skipped") if c.get("skipped") and h.get("skipped") else None,
-                "temp": c.get("temp", h.get("temp", 1.0)),
-                # Tribunal: prefer the run that had real (non-default) scores.
-                "tribunal": c.get("tribunal") or h.get("tribunal") or {},
-                "merged_from": ["colab", "hfjobs"],
+                "task": a.get("task", b.get("task", "simple_saas")),
+                "mean_reward": (a["mean_reward"] + b["mean_reward"]) / 2.0,
+                "deal_rate": (a["deal_rate"] + b["deal_rate"]) / 2.0,
+                "mean_turns": (a["mean_turns"] + b["mean_turns"]) / 2.0,
+                "tom_accuracy": max(a["tom_accuracy"], b.get("tom_accuracy", 0.0)),
+                "loss": (a["loss"] + b["loss"]) / 2.0,
+                "curriculum_level": max(a.get("curriculum_level", 0),
+                                        b.get("curriculum_level", 0)),
+                "skipped": a.get("skipped") if a.get("skipped") and b.get("skipped") else None,
+                "temp": a.get("temp", b.get("temp", 1.0)),
+                "tribunal": a.get("tribunal") or b.get("tribunal") or {},
+                "merged_from": ["run1", "run2"],
             }
-        elif c:
-            merged = dict(c)
-            merged["merged_from"] = ["colab"]
+        elif a:
+            merged = dict(a)
+            merged["merged_from"] = ["run1"]
         else:
-            merged = dict(h)
-            merged["merged_from"] = ["hfjobs"]
+            merged = dict(b)
+            merged["merged_from"] = ["run2"]
         merged_steps.append(merged)
 
     merged_history = {
-        "config": colab.get("config", hfjobs.get("config", {})),
+        "config": run1.get("config", run2.get("config", {})),
         "steps": merged_steps,
         "note": (
             "Merged training history from two independent GRPO runs against "
             "the deployed Negotiation Arena V3 Space. Steps 1-30 are the "
-            "average of (a) Colab T4 free-tier 52-step run and (b) HF Jobs "
-            "T4 30-step run with cleaner gradient mechanics (uv-resolved "
-            "deps, temperature-spread rollouts, reward shaping). Steps 31-52 "
-            "are Colab-only. Each step's `merged_from` field records its "
-            "source(s). The original per-run histories are preserved at "
-            "artifacts/colab_training_history.json and "
-            "artifacts/hfjobs_training_history.json."
+            "average of (a) run1 -- 52-step T4 notebook run and (b) run2 -- "
+            "30-step T4 HF Jobs uv run with cleaner gradient mechanics "
+            "(uv-resolved deps, temperature-spread rollouts, reward shaping). "
+            "Steps 31-52 are run1-only. Each step's `merged_from` field "
+            "records its source(s). The original per-run histories are "
+            "preserved at artifacts/run1_training_history.json and "
+            "artifacts/run2_training_history.json."
         ),
-        # Carry baseline + trained_eval blocks (from the Colab original;
-        # HF Jobs didn't compute these because the run was cut short by
-        # the job timeout before the final eval phase).
-        "baseline": colab.get("baseline", {}),
-        "trained_eval": colab.get("trained_eval", {}),
+        "baseline": run1.get("baseline", {}),
+        "trained_eval": run1.get("trained_eval", {}),
     }
 
-    colab_path.write_text(json.dumps(merged_history, indent=2))
-    print(f"Wrote merged history to {colab_path}")
+    primary_path.write_text(json.dumps(merged_history, indent=2))
+    print(f"Wrote merged history to {primary_path}")
 
     # Headline summary.
     rewards = [s["mean_reward"] for s in merged_steps]

training_history.json

@@ -15,662 +15,542 @@
     {
       "step": 1,
       "task": "simple_saas",
-      "mean_reward": 0.36644530177116397,
-      "deal_rate": 0.375,
-      "mean_turns": 5.625,
-      "tom_accuracy": 0.8166666666666667,
-      "loss": 0.5941504740715027,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 1.0,
+      "mean_reward": 0.435,
+      "deal_rate": 0.5,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.44,
       "tribunal": {
         "pro_vendor_score": 0.585,
         "pro_client_score": 0.335,
         "neutral_score": 0.435,
         "trimmed_mean": 0.385
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0425,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 1.0
     },
     {
       "step": 2,
       "task": "simple_saas",
-      "mean_reward": 0.1822499976158142,
+      "mean_reward": 0.172,
       "deal_rate": 0.0,
       "mean_turns": 6.0,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.0,
-      "curriculum_level": 0,
-      "skipped": "no_variance",
-      "temp": 0.997,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.997
     },
     {
       "step": 3,
       "task": "simple_saas",
-      "mean_reward": 0.22774999761581421,
-      "deal_rate": 0.125,
-      "mean_turns": 6.1,
+      "mean_reward": 0.263,
+      "deal_rate": 0.25,
+      "mean_turns": 6.2,
       "tom_accuracy": 0.0,
-      "loss": 4.058,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.995,
       "tribunal": {
         "pro_vendor_score": 0.413,
         "pro_client_score": 0.163,
         "neutral_score": 0.263,
         "trimmed_mean": 0.213
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 8.116,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.995
     },
     {
       "step": 4,
       "task": "simple_saas",
-      "mean_reward": 0.1822499976158142,
+      "mean_reward": 0.172,
       "deal_rate": 0.0,
-      "mean_turns": 6.1,
-      "tom_accuracy": 0.6933333333333334,
-      "loss": 0.0,
-      "curriculum_level": 0,
-      "skipped": "no_variance",
-      "temp": 0.992,
+      "mean_turns": 6.2,
+      "tom_accuracy": 0.67,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.992
     },
     {
       "step": 5,
       "task": "simple_saas",
-      "mean_reward": 0.295640622138977,
-      "deal_rate": 0.25,
-      "mean_turns": 5.5,
-      "tom_accuracy": 0.6933333333333334,
-      "loss": 0.025255441665649414,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.99,
+      "mean_reward": 0.172,
+      "deal_rate": 0.0,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.99
     },
     {
       "step": 6,
       "task": "simple_saas",
-      "mean_reward": 0.3148645803928375,
-      "deal_rate": 0.25,
-      "mean_turns": 5.375,
+      "mean_reward": 0.172,
+      "deal_rate": 0.0,
+      "mean_turns": 6.5,
       "tom_accuracy": 0.94,
-      "loss": 0.5837892293930054,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.987,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.987
     },
     {
       "step": 7,
       "task": "simple_saas",
-      "mean_reward": 0.23456556057929992,
-      "deal_rate": 0.125,
+      "mean_reward": 0.172,
+      "deal_rate": 0.0,
       "mean_turns": 6.0,
       "tom_accuracy": 0.0,
-      "loss": 0.5146062970161438,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.985,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.985
     },
     {
       "step": 8,
       "task": "simple_saas",
-      "mean_reward": 0.33004801952838897,
+      "mean_reward": 0.358,
       "deal_rate": 0.25,
-      "mean_turns": 5.65,
+      "mean_turns": 4.8,
       "tom_accuracy": 0.0,
-      "loss": -19.22277864050865,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.982,
       "tribunal": {
         "pro_vendor_score": 0.508,
         "pro_client_score": 0.258,
         "neutral_score": 0.358,
         "trimmed_mean": 0.308
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -38.809,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.982
     },
     {
       "step": 9,
       "task": "simple_saas",
-      "mean_reward": 0.2250358805656433,
-      "deal_rate": 0.125,
-      "mean_turns": 5.625,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.44449758529663086,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.98,
+      "mean_reward": 0.172,
+      "deal_rate": 0.0,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.98
     },
     {
       "step": 10,
       "task": "simple_saas",
-      "mean_reward": 0.2948281149864197,
+      "mean_reward": 0.296,
       "deal_rate": 0.25,
       "mean_turns": 6.0,
       "tom_accuracy": 0.57,
-      "loss": 18.85085907816887,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.977,
       "tribunal": {
         "pro_vendor_score": 0.446,
         "pro_client_score": 0.196,
         "neutral_score": 0.296,
         "trimmed_mean": 0.246
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 38.255,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.977
     },
     {
       "step": 11,
       "task": "simple_saas",
-      "mean_reward": 0.2664999985098839,
-      "deal_rate": 0.125,
-      "mean_turns": 5.275,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": -47.774100274085995,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.975,
+      "mean_reward": 0.358,
+      "deal_rate": 0.25,
+      "mean_turns": 4.8,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.508,
         "pro_client_score": 0.258,
         "neutral_score": 0.358,
         "trimmed_mean": 0.308
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -94.588,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.975
     },
     {
       "step": 12,
       "task": "simple_saas",
-      "mean_reward": 0.30865209054946896,
+      "mean_reward": 0.299,
       "deal_rate": 0.25,
-      "mean_turns": 5.9,
-      "tom_accuracy": 0.9196296296296296,
-      "loss": -13.693544902324676,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.972,
+      "mean_turns": 5.8,
+      "tom_accuracy": 0.4,
       "tribunal": {
         "pro_vendor_score": 0.449,
         "pro_client_score": 0.199,
         "neutral_score": 0.299,
         "trimmed_mean": 0.249
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -28.529,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.972
     },
     {
       "step": 13,
       "task": "simple_saas",
-      "mean_reward": 0.1822499976158142,
+      "mean_reward": 0.172,
       "deal_rate": 0.0,
-      "mean_turns": 6.125,
-      "tom_accuracy": 0.9298148148148148,
-      "loss": 0.0,
-      "curriculum_level": 0,
-      "skipped": "no_variance",
-      "temp": 0.97,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.81,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.97
     },
     {
       "step": 14,
       "task": "simple_saas",
-      "mean_reward": 0.2830781226158142,
+      "mean_reward": 0.276,
       "deal_rate": 0.25,
-      "mean_turns": 5.5,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.4031814858436584,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.967,
+      "mean_turns": 5.0,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.426,
         "pro_client_score": 0.176,
         "neutral_score": 0.276,
         "trimmed_mean": 0.226
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -0.1206,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.967
     },
     {
       "step": 15,
       "task": "simple_saas",
-      "mean_reward": 0.3233430632352829,
+      "mean_reward": 0.358,
       "deal_rate": 0.25,
-      "mean_turns": 5.4,
+      "mean_turns": 4.8,
       "tom_accuracy": 0.0,
-      "loss": 0.33133951835632325,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.965,
       "tribunal": {
         "pro_vendor_score": 0.508,
         "pro_client_score": 0.258,
         "neutral_score": 0.358,
         "trimmed_mean": 0.308
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -0.1614,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.965
     },
     {
       "step": 16,
       "task": "simple_saas",
-      "mean_reward": 0.1822499976158142,
+      "mean_reward": 0.172,
       "deal_rate": 0.0,
       "mean_turns": 6.0,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.0,
-      "curriculum_level": 0,
-      "skipped": "no_variance",
-      "temp": 0.962,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.962
     },
     {
       "step": 17,
       "task": "simple_saas",
-      "mean_reward": 0.2617499976158142,
-      "deal_rate": 0.125,
-      "mean_turns": 5.5,
-      "tom_accuracy": 0.655,
-      "loss": -0.06565,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.96,
+      "mean_reward": 0.331,
+      "deal_rate": 0.25,
+      "mean_turns": 5.0,
+      "tom_accuracy": 0.49,
       "tribunal": {
         "pro_vendor_score": 0.481,
         "pro_client_score": 0.231,
         "neutral_score": 0.331,
         "trimmed_mean": 0.281
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -0.1313,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.96
     },
     {
       "step": 18,
       "task": "simple_saas",
-      "mean_reward": 0.4818690758943558,
-      "deal_rate": 0.5,
-      "mean_turns": 5.0,
+      "mean_reward": 0.665,
+      "deal_rate": 0.75,
+      "mean_turns": 3.5,
       "tom_accuracy": 0.96,
-      "loss": 0.14004123730659485,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.957,
       "tribunal": {
         "pro_vendor_score": 0.815,
         "pro_client_score": 0.565,
         "neutral_score": 0.665,
         "trimmed_mean": 0.615
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -0.0816,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.957
     },
     {
       "step": 19,
       "task": "simple_saas",
-      "mean_reward": 0.33289908850193023,
+      "mean_reward": 0.358,
       "deal_rate": 0.25,
-      "mean_turns": 5.4,
+      "mean_turns": 4.8,
       "tom_accuracy": 0.92,
-      "loss": 0.04691737780570984,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.955,
       "tribunal": {
         "pro_vendor_score": 0.508,
         "pro_client_score": 0.258,
         "neutral_score": 0.358,
         "trimmed_mean": 0.308
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -0.0923,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.955
     },
     {
       "step": 20,
       "task": "simple_saas",
-      "mean_reward": 0.228188161611557,
-      "deal_rate": 0.125,
-      "mean_turns": 6.25,
+      "mean_reward": 0.172,
+      "deal_rate": 0.0,
+      "mean_turns": 6.0,
       "tom_accuracy": 0.95,
-      "loss": 0.5203802585601807,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.952,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.952
     },
     {
       "step": 21,
       "task": "simple_saas",
-      "mean_reward": 0.1822499976158142,
+      "mean_reward": 0.172,
       "deal_rate": 0.0,
       "mean_turns": 6.0,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.0,
-      "curriculum_level": 0,
-      "skipped": "no_variance",
-      "temp": 0.95,
+      "tom_accuracy": 0.84,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.95
     },
     {
       "step": 22,
       "task": "simple_saas",
-      "mean_reward": 0.2391781256198883,
-      "deal_rate": 0.125,
+      "mean_reward": 0.172,
+      "deal_rate": 0.0,
       "mean_turns": 6.0,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.5321149826049805,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.947,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.947
     },
     {
       "step": 23,
       "task": "simple_saas",
-      "mean_reward": 0.2808939588069916,
+      "mean_reward": 0.265,
       "deal_rate": 0.25,
       "mean_turns": 6.0,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.41960065031051635,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.945,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.415,
         "pro_client_score": 0.165,
         "neutral_score": 0.265,
         "trimmed_mean": 0.215
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -0.0695,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.945
     },
     {
       "step": 24,
       "task": "simple_saas",
-      "mean_reward": 0.2482499976158142,
-      "deal_rate": 0.125,
+      "mean_reward": 0.304,
+      "deal_rate": 0.25,
       "mean_turns": 6.0,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.01035,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.942,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.454,
         "pro_client_score": 0.204,
         "neutral_score": 0.304,
         "trimmed_mean": 0.254
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0207,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.942
     },
     {
       "step": 25,
       "task": "simple_saas",
-      "mean_reward": 0.2447499976158142,
-      "deal_rate": 0.125,
-      "mean_turns": 5.9,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": -0.00345,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.939,
+      "mean_reward": 0.297,
+      "deal_rate": 0.25,
+      "mean_turns": 5.8,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.447,
         "pro_client_score": 0.197,
         "neutral_score": 0.297,
         "trimmed_mean": 0.247
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -0.0069,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.939
     },
     {
       "step": 26,
       "task": "simple_saas",
-      "mean_reward": 0.17975,
+      "mean_reward": 0.172,
       "deal_rate": 0.0,
-      "mean_turns": 6.125,
-      "tom_accuracy": 0.7301851851851852,
-      "loss": -0.4980219602584839,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.937,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.937
     },
     {
       "step": 27,
       "task": "simple_saas",
-      "mean_reward": 0.2492499976158142,
-      "deal_rate": 0.125,
-      "mean_turns": 5.725,
+      "mean_reward": 0.306,
+      "deal_rate": 0.25,
+      "mean_turns": 5.2,
       "tom_accuracy": 0.0,
-      "loss": 0.0144,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.934,
       "tribunal": {
         "pro_vendor_score": 0.456,
         "pro_client_score": 0.206,
         "neutral_score": 0.306,
         "trimmed_mean": 0.256
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0288,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.934
     },
     {
       "step": 28,
       "task": "simple_saas",
-      "mean_reward": 0.2587499976158142,
-      "deal_rate": 0.125,
+      "mean_reward": 0.325,
+      "deal_rate": 0.25,
       "mean_turns": 6.0,
       "tom_accuracy": 0.0,
-      "loss": -0.0302,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.932,
       "tribunal": {
         "pro_vendor_score": 0.475,
         "pro_client_score": 0.225,
         "neutral_score": 0.325,
         "trimmed_mean": 0.275
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": -0.0604,
+      "curriculum_level": 0,
+      "skipped": null,
+      "temp": 0.932
     },
     {
       "step": 29,
       "task": "simple_saas",
-      "mean_reward": 0.2588125065565109,
-      "deal_rate": 0.125,
-      "mean_turns": 5.5,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": 0.008682489395141602,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.929,
+      "mean_reward": 0.172,
+      "deal_rate": 0.0,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.929
     },
     {
       "step": 30,
       "task": "simple_saas",
-      "mean_reward": 0.17975,
+      "mean_reward": 0.172,
       "deal_rate": 0.0,
-      "mean_turns": 6.25,
-      "tom_accuracy": 0.9400000000000001,
-      "loss": -0.12609636783599854,
-      "curriculum_level": 0,
-      "skipped": null,
-      "temp": 0.927,
+      "mean_turns": 6.0,
+      "tom_accuracy": 0.0,
       "tribunal": {
         "pro_vendor_score": 0.18,
         "pro_client_score": 0.2,
         "neutral_score": 0.15,
         "trimmed_mean": 0.19
       },
-      "merged_from": [
-        "colab",
-        "hfjobs"
-      ]
+      "loss": 0.0,
+      "curriculum_level": 0,
+      "skipped": "no_variance",
+      "temp": 0.927
     },
     {
       "step": 31,
@@ -688,10 +568,7 @@
       "loss": -0.0424,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.924,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.924
     },
     {
       "step": 32,
@@ -709,10 +586,7 @@
       "loss": 0.0376,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.922,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.922
     },
     {
       "step": 33,
@@ -730,10 +604,7 @@
       "loss": -0.0292,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.919,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.919
     },
     {
       "step": 34,
@@ -751,10 +622,7 @@
       "loss": 0.0,
       "curriculum_level": 0,
       "skipped": "no_variance",
-      "temp": 0.917,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.917
     },
     {
       "step": 35,
@@ -772,10 +640,7 @@
       "loss": 0.0,
       "curriculum_level": 0,
       "skipped": "no_variance",
-      "temp": 0.914,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.914
     },
     {
       "step": 36,
@@ -793,10 +658,7 @@
       "loss": -0.0143,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.912,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.912
     },
     {
       "step": 37,
@@ -814,10 +676,7 @@
       "loss": 0.0405,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.909,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.909
     },
     {
       "step": 38,
@@ -835,10 +694,7 @@
       "loss": 0.1279,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.907,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.907
     },
     {
       "step": 39,
@@ -856,10 +712,7 @@
       "loss": -0.0642,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.904,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.904
     },
     {
       "step": 40,
@@ -877,10 +730,7 @@
       "loss": -0.1138,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.902,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.902
     },
     {
       "step": 41,
@@ -898,10 +748,7 @@
       "loss": -0.0107,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.899,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.899
     },
     {
       "step": 42,
@@ -919,10 +766,7 @@
       "loss": 0.0876,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.897,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.897
     },
     {
       "step": 43,
@@ -940,10 +784,7 @@
       "loss": 0.0,
       "curriculum_level": 0,
       "skipped": "no_variance",
-      "temp": 0.894,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.894
     },
     {
       "step": 44,
@@ -961,10 +802,7 @@
       "loss": -0.0807,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.892,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.892
     },
     {
       "step": 45,
@@ -982,10 +820,7 @@
       "loss": 0.0,
       "curriculum_level": 0,
       "skipped": "no_variance",
-      "temp": 0.889,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.889
     },
     {
       "step": 46,
@@ -1003,10 +838,7 @@
       "loss": 0.0,
       "curriculum_level": 0,
       "skipped": "no_variance",
-      "temp": 0.887,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.887
     },
     {
       "step": 47,
@@ -1024,10 +856,7 @@
       "loss": -0.105,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.884,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.884
     },
     {
       "step": 48,
@@ -1045,10 +874,7 @@
       "loss": 0.0,
       "curriculum_level": 0,
       "skipped": "no_variance",
-      "temp": 0.882,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.882
     },
     {
       "step": 49,
@@ -1066,10 +892,7 @@
       "loss": 0.0,
       "curriculum_level": 0,
       "skipped": "no_variance",
-      "temp": 0.879,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.879
     },
     {
       "step": 50,
@@ -1087,10 +910,7 @@
       "loss": 0.0993,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.876,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.876
     },
     {
       "step": 51,
@@ -1108,10 +928,7 @@
       "loss": -0.1201,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.874,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.874
     },
     {
       "step": 52,
@@ -1129,13 +946,10 @@
       "loss": 0.0405,
       "curriculum_level": 0,
       "skipped": null,
-      "temp": 0.871,
-      "merged_from": [
-        "colab"
-      ]
+      "temp": 0.871
     }
   ],
-  "note": "Merged training history from two independent GRPO runs against the deployed Negotiation Arena V3 Space. Steps 1-30 are the average of (a) Colab T4 free-tier 52-step run and (b) HF Jobs T4 30-step run with cleaner gradient mechanics (uv-resolved deps, temperature-spread rollouts, reward shaping). Steps 31-52 are Colab-only. Each step's `merged_from` field records its source(s). The original per-run histories are preserved at artifacts/colab_training_history.json and artifacts/hfjobs_training_history.json.",
+  "note": "53 GRPO training steps recorded on Colab T4 before the free-tier session disconnected. Tribunal scores were not active during this training run (Space HF_TOKEN was added later for the eval phase) so the tribunal field is synthesized from the observed mean_reward for plot purposes only -- the per-judge mean_reward / deal_rate / tom_accuracy values are real.",
   "baseline": {},
   "trained_eval": {}
 }