Revert "Merge remote main with local project"

This reverts commit ed658215c877763e70b07a4c575831c762d5daa1, reversing
changes made to 20ef5229264b826ffe0a39701c5b96d0786f0c93.

README.md

@@ -8,7 +8,6 @@ pinned: false
 license: mit
 ---
 
-<<<<<<< HEAD
 # 🛡️ SENTINEL — Self-Evolving Network for Training Intelligent Agents Under Adversarial Long-Horizon Tasks
 
 > Agents fail because they trust blindly. SENTINEL trains skepticism, recovery, and oversight.
@@ -46,36 +45,6 @@ SENTINEL turns that failure mode into a **trainable environment**. The model onl
 ## 🌍 Real-World Bridge
 
 SENTINEL is not a normal chatbot that answers one prompt. It is the training ground for the **hidden control loop** inside a long-running agent.
-=======
-# SENTINEL
-
-Self-Evolving Network for Training Intelligent Agents Under Adversarial Long-Horizon Tasks.
-
-SENTINEL is an OpenEnv-compatible RL environment for one core skill: training an orchestrator to decide who to trust, when to verify, how to recover, and how to finish long multi-agent work when specialist agents are unreliable or adversarial.
-
-## Rollout Source Of Truth
-
-The phased execution plan and presentation assets now live in-repo:
-
-- [Rollout](docs/ROLL_OUT.md)
-- [Narrative Lock](docs/presentation/NARRATIVE_LOCK.md)
-- [Visual System](docs/diagrams/VISUAL_SYSTEM.md)
-
-## Why It Matters
-
-Modern agent systems fail in the same pattern:
-
-1. A long task is decomposed into many steps.
-2. The orchestrator delegates to sub-agents or tools.
-3. One specialist returns a confident but wrong result.
-4. The system trusts it, builds on it, and drifts into failure.
-
-SENTINEL turns that failure mode into a trainable environment. The model only sees behavior: returned outcomes, confidence, stakes, history, and trust scores. It never sees hidden specialist identities.
-
-## Real-World Bridge
-
-SENTINEL is not a normal chatbot that answers one prompt. It is the training ground for the hidden control loop inside a long-running agent.
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 Example user mission:
 
@@ -86,7 +55,6 @@ fix the risky parts, and prepare it for deployment.
 
 What SENTINEL abstracts:
 
-<<<<<<< HEAD
 1. The user mission becomes a scenario with a **task graph**.
 2. The LLM orchestrator sees one subtask, current stakes, public specialist IDs, and trust scores.
 3. The model emits one control action: `delegate`, `verify`, `solve_independently`, or `skip`.
@@ -186,67 +154,6 @@ Hidden profiles:
 ---
 
 ## 💰 Reward Model
-=======
-1. The user mission becomes a scenario with a task graph.
-2. The LLM orchestrator sees one subtask, current stakes, public specialist ids, and trust scores.
-3. The model emits one control action: `delegate`, `verify`, `solve_independently`, or `skip`.
-4. A hidden specialist profile responds: accurate, overconfident, domain-bound, adversarial, or degrading.
-5. The reward engine scores the action and the trust ledger updates.
-6. GRPO/TRL uses that reward to train better orchestration behavior.
-
-This is why the project matters for real agents: after many long user requests, the failure is often not "the LLM cannot speak." The failure is that the system trusted the wrong intermediate result and kept building on it. SENTINEL trains the agent to catch that failure while it is still recoverable.
-
-Judge-readable endpoints:
-
-```bash
-curl http://localhost:7860/problem
-curl "http://localhost:7860/mission?task_type=task3"
-```
-
-## Environment Shape
-
-- API: `reset()`, `step(action)`, `state()`
-- Runtime: FastAPI on port `7860`
-- Tasks: `task1`, `task2`, `task3`
-- Specialists: 5 scripted FSM agents with shuffled hidden profiles
-- Rewards: per-step reward plus terminal score, normalized to `0.0-1.0`
-- Dataset: 120 abstract multi-agent scenarios
-- Session store: single-process memory with TTL/LRU cleanup
-- Optional adaptive curriculum: pass `adaptive=true` on `/reset` for Theme 4 demos
-- Live trust stream: `/stream?session_id=...` feeds the `/trust-dashboard` bars
-
-Deployment contract: run one server worker for the submitted Space. Active `SentinelEnv` objects live in process memory, so multi-worker deployments need sticky sessions or a shared store such as Redis. The Dockerfile intentionally starts uvicorn with `--workers 1`.
-
-## Live Submission Targets
-
-- GitHub: `https://github.com/ADITYAGABA1322/sentinel-env`
-- Hugging Face Space: `https://xcodeaddy-sentinel-env.hf.space`
-- OpenEnv base URL: `https://xcodeaddy-sentinel-env.hf.space`
-
-## Specialist Behaviors
-
-| Public Slot | Hidden Behavior |
-| --- | --- |
-| S0-S4 | Public ids are shuffled every episode |
-
-Hidden profiles:
-
-- `AccurateSlow`: 90 percent accurate, costs 2 steps.
-- `OverconfidentFast`: quick and confident, wrong 40 percent of the time.
-- `DomainBound`: strong on analysis/verification, weak elsewhere.
-- `Adversarial`: benign below stakes `0.70`, poisons at stakes `>=0.70`.
-- `Degrading`: strong early, gets worse with reuse.
-
-## Tasks
-
-| Task | Difficulty | Goal |
-| --- | --- | --- |
-| `task1` | Easy | Learn basic delegation over a 10-subtask chain. |
-| `task2` | Medium | Complete a 15-subtask DAG and build calibrated trust. |
-| `task3` | Hard | Complete a 20-subtask adversarial mission under high-stakes poison attempts. |
-
-## Reward Model
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 Rewards are deterministic and boundary-exclusive per step: `(0.01, 0.99)`.
 
@@ -255,16 +162,11 @@ Shared threshold/cost constants live in `sentinel_config.py`:
 - `ADVERSARIAL_TRIGGER_STAKES = 0.70`: adversarial specialist starts poisoning.
 - `ADVERSARIAL_AWARENESS_STAKES = 0.70`: graders reward verification/evasion.
 - `CRITICAL_POISON_STAKES = 0.85`: unchecked poison ends the episode.
-<<<<<<< HEAD
 - `VERIFY_EXTRA_STEP_COST = 1`: verify cost = specialist step cost + 1.
-=======
-- `VERIFY_EXTRA_STEP_COST = 1`: verify cost is specialist step cost plus one.
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 Task 3 terminal score:
 
 ```text
-<<<<<<< HEAD
 0.35 × completion_rate
 + 0.30 × adversarial_detection_rate
 + 0.25 × trust_calibration
@@ -272,17 +174,6 @@ Task 3 terminal score:
 ```
 
 **Reward Engine v2** adds process-aware signals on top of outcome scoring:
-=======
-0.35 * completion_rate
-+ 0.30 * adversarial_detection_rate
-+ 0.25 * trust_calibration
-+ 0.10 * efficiency
-```
-
-The episode `score` exposed in `info` and inference logs is the mean reward over emitted grading events, normalized to `0.0-1.0`. It is intentionally not raw cumulative return; terminal reward and efficiency terms carry the penalty for unfinished or wasteful episodes while keeping scores comparable across tasks with different horizons.
-
-Reward Engine v2 adds process-aware signals on top of outcome scoring:
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 - `confidence_alignment`: penalizes confident wrong outputs.
 - `domain_routing`: rewards domain-bound behavior only when it is actually in-domain.
@@ -290,7 +181,6 @@ Reward Engine v2 adds process-aware signals on top of outcome scoring:
 
 The active step formulas are exposed at `/grader`, and each active episode exposes a full component trace at `/reward-report?session_id=<id>`.
 
-<<<<<<< HEAD
 ---
 
 ## ✨ WOW Factor Features
@@ -316,32 +206,6 @@ The active step formulas are exposed at `/grader`, and each active episode expos
 ---
 
 ## 🌐 API
-=======
-## WOW Factor Features
-
-SENTINEL now includes three judge-facing upgrades:
-
-1. **Adaptive difficulty engine**: `DifficultyController` watches rolling adversarial detection rate. Strong agents get earlier adversarial triggers, more high-stakes nodes, and a tighter step budget. Struggling agents get easier episodes. Enable it with:
-
-```bash
-curl -X POST http://localhost:7860/reset \
-  -H "Content-Type: application/json" \
-  -d '{"task_type":"task3","seed":42,"adaptive":true}'
-```
-
-2. **Behavioral fingerprints**: every observation includes `behavioral_fingerprints` for S0-S4:
-
-- `confidence_accuracy_gap`
-- `domain_hit_rate`
-- `stakes_volatility`
-- low/high stakes accuracy
-
-These are public behavioral signals only. They do not leak the hidden specialist identity.
-
-3. **Live trust stream**: `/stream?session_id=<id>` emits server-sent events with trust updates, fingerprints, and difficulty profile. Open `/trust-dashboard?session_id=<id>` during a demo to watch the trust bars update live.
-
-## API
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 curl http://localhost:7860/health
@@ -356,19 +220,11 @@ curl "http://localhost:7860/reward-report?session_id=<session_id>"
 curl http://localhost:7860/difficulty
 ```
 
-<<<<<<< HEAD
 The root route `/` serves the live **SENTINEL dashboard** on Hugging Face Spaces.
 Use `/api` for the JSON route index.
 Use `/assets/baseline_comparison.png` for the committed baseline chart used in the dashboard.
 
 ### Live Stream Demo
-=======
-The root route `/` serves the live SENTINEL dashboard on Hugging Face Spaces.
-Use `/api` for the JSON route index.
-Use `/assets/baseline_comparison.png` for the committed baseline chart used in the dashboard.
-
-Live stream demo:
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 # Terminal 1
@@ -383,7 +239,6 @@ curl -s -X POST http://localhost:7860/reset \
 open "http://localhost:7860/trust-dashboard?session_id=<session_id>"
 ```
 
-<<<<<<< HEAD
 ### Start an Episode
 
 ```bash
@@ -409,9 +264,6 @@ curl -X POST "http://localhost:7860/step?session_id=<SESSION_ID>" \
 ---
 
 ## 🧪 Backend Walkthrough
-=======
-## Backend Walkthrough
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 For terminal-first debugging and pitch clarity, run:
 
@@ -429,18 +281,13 @@ This prints the full backend story:
 
 The key scenario to understand is `task3, seed=42`: public slot `S0` is secretly adversarial. It behaves correctly at low stakes, gains trust, then starts poisoning high-stakes nodes. SENTINEL exists to train the orchestrator to catch that shift.
 
-<<<<<<< HEAD
 ### Adaptive Evaluation
-=======
-Adaptive evaluation:
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 python training/evaluate.py --episodes 100 --task task3 --adaptive --reset-difficulty \
   --plot outputs/task3_adaptive_comparison.png
 ```
 
-<<<<<<< HEAD
 ---
 
 ## 🖥️ Live Dashboard
@@ -503,89 +350,6 @@ sentinel-env/
 ---
 
 ## ⚡ Local Setup
-=======
-## Live Dashboard
-
-The Space opens directly into **SENTINEL Trust Mission Control**, a judge-demo dashboard:
-
-- live task progress and score
-- S0-S4 network theater with trust state per public slot
-- manual `delegate`, `verify`, `solve_independently`, and `skip` controls
-- heuristic auto-policy and one-click recommended move
-- API playground showing raw request and response payloads
-- profile reshuffle demo via seed swap
-- before-and-after story lane for judge presentation
-- hackathon readiness panel for what is done vs still pending
-- risk gate for high-stakes subtasks
-- flight recorder of step rewards and decisions
-- code-flow map from `reset()` to reward
-- hackathon theme coverage map
-- adversarial detection and poisoning counters
-- baseline proof table and chart for random, heuristic, and oracle-lite policies
-
-Current status as of April 22, 2026:
-
-| Requirement | Status |
-| --- | --- |
-| Hugging Face Space | Live |
-| Docker build | Passing |
-| OpenEnv validation | Passing |
-| Baseline chart | Committed |
-| Live trust UI | Deployed |
-| Mini-blog/video | Still required before finale |
-| Onsite GRPO curve | Still required during finale |
-
-Start an episode:
-
-```bash
-curl -X POST http://localhost:7860/reset \
-  -H "Content-Type: application/json" \
-  -d '{"task_type":"task3","seed":42}'
-```
-
-Step:
-
-```bash
-curl -X POST "http://localhost:7860/step?session_id=<SESSION_ID>" \
-  -H "Content-Type: application/json" \
-  -d '{
-    "session_id":"<SESSION_ID>",
-    "task_type":"task3",
-    "action_type":"delegate",
-    "specialist_id":"S2",
-    "reasoning":"S2 has the best observed trust score"
-  }'
-```
-
-## Project Structure
-
-```text
-sentinel-env/
-|-- app.py
-|-- environment.py
-|-- models.py
-|-- graders.py
-|-- specialists.py
-|-- trust_ledger.py
-|-- task_graph.py
-|-- comms_bus.py
-|-- scenarios.py
-|-- inference.py
-|-- openenv.yaml
-|-- Dockerfile
-|-- requirements.txt
-|-- training/
-|   |-- train.py
-|   |-- evaluate.py
-|   `-- colab_notebook.ipynb
-`-- tests/
-    |-- test_environment.py
-    |-- test_graders.py
-    `-- test_specialists.py
-```
-
-## Local Setup
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 python3 -m venv .venv
@@ -595,11 +359,7 @@ pip install -r requirements.txt
 pip install pytest
 ```
 
-<<<<<<< HEAD
 ### Run Checks
-=======
-Run checks:
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 python -m py_compile app.py server/app.py environment.py models.py graders.py specialists.py trust_ledger.py task_graph.py scenarios.py inference.py comms_bus.py mission_context.py sentinel_config.py training/evaluate.py training/train.py scripts/backend_walkthrough.py
@@ -610,45 +370,29 @@ python training/train.py --dry-run --episodes 5
 python scripts/backend_walkthrough.py --task task3 --seed 42 --policy heuristic --compare --max-rows 14
 ```
 
-<<<<<<< HEAD
 ### Run the Server
-=======
-Run the server:
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 uvicorn app:app --host 0.0.0.0 --port 7860
 ```
 
-<<<<<<< HEAD
 ### Validate with OpenEnv
-=======
-Validate with OpenEnv:
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 pip install openenv-core==0.2.3
 openenv validate . --json
 ```
 
-<<<<<<< HEAD
 ### Docker
-=======
-Docker:
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 docker build -t sentinel-env .
 docker run -p 7860:7860 sentinel-env
 ```
 
-<<<<<<< HEAD
 ---
 
 ## 📊 Baselines
-=======
-## Baselines
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 `inference.py` runs 30 deterministic heuristic episodes and emits only strict hackathon logs:
 
@@ -663,7 +407,6 @@ docker run -p 7860:7860 sentinel-env
 - `random`
 - `heuristic`
 - `oracle_lite`
-<<<<<<< HEAD
 - `trained`
 
 The evaluator writes `outputs/evaluation_results.json` and `outputs/baseline_comparison.png`.
@@ -671,24 +414,6 @@ The evaluator writes `outputs/evaluation_results.json` and `outputs/baseline_com
 ---
 
 ## 🚀 Hugging Face Deployment
-=======
-
-The evaluator writes `outputs/evaluation_results.json` and `outputs/baseline_comparison.png`.
-
-![Baseline Comparison](outputs/baseline_comparison.png)
-
-Latest local comparison, 20 episodes per task and policy:
-
-| Policy | Overall | Task 1 | Task 2 | Task 3 |
-| --- | ---: | ---: | ---: | ---: |
-| Random | 0.6954 | 0.7702 | 0.6505 | 0.6655 |
-| Heuristic trust-weighted | 0.7960 | 0.8690 | 0.7677 | 0.7513 |
-| Oracle-lite upper bound | 0.8553 | 0.9180 | 0.7801 | 0.8678 |
-
-The demo story is the score gap: the reward function distinguishes blind delegation from trust-aware routing, and the oracle-lite upper bound shows room for onsite RL training.
-
-## Hugging Face Deployment
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ```bash
 huggingface-cli login
@@ -708,7 +433,6 @@ curl -X POST https://xcodeaddy-sentinel-env.hf.space/reset \
 openenv validate . --json
 ```
 
-<<<<<<< HEAD
 ---
 
 ## 🏆 Hackathon Alignment
@@ -743,24 +467,3 @@ A detailed mini-blog explaining what SENTINEL does and what we trained is publis
 ## 📜 License
 
 MIT
-=======
-## Mini-Blog Draft
-
-Title: `SENTINEL: Training AI to Trust Wisely in Multi-Agent Systems`
-
-SENTINEL is an OpenEnv RL environment for one failure mode: multi-agent systems delegate blindly. One orchestrator must complete long tasks by routing work across five specialist agents whose reliability profiles are hidden and reshuffled every episode. The orchestrator only sees behavior, confidence, stakes, and history, so it must learn skepticism, verification, recovery, and calibrated trust.
-
-The specialists are deterministic FSMs on purpose: they give stable reward signals while the orchestrator remains the trainable target. Under Reward Engine v2, random routing scores `0.6954`, trust-weighted routing scores `0.7960`, and oracle-lite reaches `0.8553`, showing the environment has a meaningful learning signal before onsite GRPO training.
-
-## Hackathon Alignment
-
-- Theme 1: multi-agent interaction, partial observability, adversarial specialist, trust calibration.
-- Theme 2: long-horizon task graphs with delayed terminal reward and failure recovery.
-- Theme 3.1: professional agent orchestration workflow with API-style actions.
-- Theme 4: profile shuffle creates a self-resetting curriculum.
-- Theme 5: targets a real AI systems failure: blind trust inside agent pipelines.
-
-Winning demo line:
-
-> Agents fail because they trust blindly. SENTINEL trains skepticism, recovery, and oversight.
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387

docs/TRAINING_RUNBOOK.md

@@ -148,7 +148,6 @@ Use a Hugging Face token in Colab for:
 
 The Space itself does not need GPU to run the replay demo.
 
-<<<<<<< HEAD
 ## Hugging Face App URLs
 
 Use these two Hugging Face URLs for different jobs:
@@ -172,8 +171,6 @@ When running locally, start uvicorn with `--host 0.0.0.0`, but open the browser
 at `http://127.0.0.1:7860/` or `http://localhost:7860/`. Do not browse to
 `http://0.0.0.0:7860/`; `0.0.0.0` is only a bind address.
 
-=======
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 ## Hugging Face Credits
 
 Best use:
@@ -181,7 +178,6 @@ Best use:
 - keep the Space on CPU for normal judging,
 - optionally upgrade the Space to T4 only during the final live demo if the UI
   needs extra responsiveness,
-<<<<<<< HEAD
 - avoid doing full training inside the Space,
 - use Hugging Face Jobs or Colab for the actual GRPO run.
 
@@ -212,12 +208,6 @@ If `import-smoke` passes, run the full job:
 
 The launcher uses `pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel` because the
 current Unsloth stack pulls `torchao`, which expects torch `>=2.11`.
-=======
-- avoid doing full training inside the Space.
-
-Training belongs in Colab. The Space is for serving the environment and replay
-demo.
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 ## Success Criteria
 

outputs/trained_policy_replay.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:308f09f476570d93aedf5329db597332171a693898f72f68a3aa71b0d21e4f06
-size 380391
+oid sha256:17600954da6421338f399beb04bf88f5aac45c87007f2deaefa8ba3a81897f76
+size 793297

pyproject.toml

@@ -18,19 +18,12 @@ server = "server.app:main"
 [project.optional-dependencies]
 dev = ["pytest>=8.0.0"]
 training = [
-<<<<<<< HEAD
   "trl==0.24.0",
   "transformers==4.57.6",
   "datasets==4.3.0",
   "accelerate==1.13.0",
   "peft==0.19.1",
   "bitsandbytes==0.49.2",
-=======
-  "trl",
-  "transformers",
-  "datasets",
-  "accelerate",
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
   "unsloth",
 ]
 

requirements-train.txt

@@ -1,5 +1,4 @@
 unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git
-<<<<<<< HEAD
 trl==0.24.0
 transformers==4.57.6
 datasets==4.3.0
@@ -10,15 +9,3 @@ matplotlib==3.10.9
 seaborn==0.13.2
 pandas==3.0.2
 huggingface_hub>=0.36,<1
-=======
-trl>=0.18.2,<0.25,!=0.19.0
-transformers>=4.56,<5
-datasets>=3.0,<5
-accelerate>=1.4
-peft>=0.14
-bitsandbytes>=0.45
-matplotlib
-seaborn
-pandas
-huggingface_hub
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387

training/colab_notebook.ipynb

@@ -74,11 +74,7 @@
     "        \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\",\n",
     "    ])\n",
     "    subprocess.check_call([\"pip\", \"install\", \"-q\", \"--no-deps\",\n",
-<<<<<<< HEAD
     "        \"trl==0.24.0\", \"transformers==4.57.6\", \"datasets==4.3.0\", \"accelerate==1.13.0\", \"peft==0.19.1\", \"bitsandbytes==0.49.2\",\n",
-=======
-    "        \"trl<0.13\", \"transformers>=4.46\", \"datasets\", \"accelerate\", \"peft\", \"bitsandbytes\",\n",
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
     "    ])\n",
     "except subprocess.CalledProcessError as exc:\n",
     "    print(f\"Training extras failed to install ({exc}); continuing with heuristic-fallback path.\")\n",

training/launch_hf_job.py

@@ -6,19 +6,12 @@ import shlex
 import sys
 from textwrap import dedent
 
-<<<<<<< HEAD
 from huggingface_hub import get_token, run_job
 
 
 # Current Unsloth pulls torchao, which expects torch >= 2.11. Keep the Jobs
 # image aligned so GRPO imports fail fast only for real code issues.
 DEFAULT_IMAGE = "pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel"
-=======
-from huggingface_hub import run_job
-
-
-DEFAULT_IMAGE = "pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel"
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 DEFAULT_REPO = "https://github.com/ADITYAGABA1322/sentinel-env"
 DEFAULT_MODEL = "unsloth/Qwen2.5-0.5B-Instruct"
 
@@ -33,7 +26,6 @@ def bootstrap_repo(repo_url: str) -> list[str]:
         "command -v git || (apt-get update && apt-get install -y git)",
         f"git clone {shlex.quote(repo_url)} sentinel-env",
         "cd sentinel-env",
-<<<<<<< HEAD
         "python -m venv --system-site-packages .job-venv || (apt-get update && apt-get install -y python3-venv && python -m venv --system-site-packages .job-venv)",
         ". .job-venv/bin/activate",
         "python -m pip install --upgrade pip",
@@ -47,11 +39,6 @@ def bootstrap_repo(repo_url: str) -> list[str]:
             "from trl import GRPOConfig, GRPOTrainer; "
             "print('training imports ok')\""
         ),
-=======
-        "python -m pip install --upgrade pip",
-        "pip install -r requirements.txt",
-        "pip install -r requirements-train.txt",
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
     ]
 
 
@@ -59,16 +46,11 @@ def gpu_test_command() -> str:
     return "python -c 'import torch; print(torch.cuda.get_device_name())'"
 
 
-<<<<<<< HEAD
 def train_command(args: argparse.Namespace, train: bool = True) -> str:
     lines = bootstrap_repo(args.repo_url)
     if not train:
         return shell_join(lines)
 
-=======
-def train_command(args: argparse.Namespace) -> str:
-    lines = bootstrap_repo(args.repo_url)
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
     lines.append(
         " ".join(
             [
@@ -126,15 +108,11 @@ def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
         description="Launch SENTINEL training on Hugging Face Jobs without shell quoting pain."
     )
-<<<<<<< HEAD
     parser.add_argument(
         "--mode",
         choices=["gpu-test", "import-smoke", "train-smoke", "train-full"],
         default="gpu-test",
     )
-=======
-    parser.add_argument("--mode", choices=["gpu-test", "train-smoke", "train-full"], default="gpu-test")
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
     parser.add_argument("--namespace", default=os.environ.get("HF_NAMESPACE", "XcodeAddy"))
     parser.add_argument("--flavor", default="a10g-small")
     parser.add_argument("--timeout", default="2h")
@@ -156,16 +134,11 @@ def parse_args() -> argparse.Namespace:
 
 def main() -> None:
     args = parse_args()
-<<<<<<< HEAD
     token = os.environ.get("HF_TOKEN") or get_token()
-=======
-    token = os.environ.get("HF_TOKEN")
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
     if not token:
         raise SystemExit(
             dedent(
                 """
-<<<<<<< HEAD
                 No Hugging Face token was found.
 
                 Either run:
@@ -174,28 +147,16 @@ def main() -> None:
 
                 Or log in once:
                   .venv/bin/hf auth login --add-to-git-credential
-=======
-                HF_TOKEN is not set.
-
-                Run:
-                  read -s HF_TOKEN
-                  export HF_TOKEN
-                Then paste your Hugging Face write token.
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
                 """
             ).strip()
         )
 
-<<<<<<< HEAD
     if args.mode == "gpu-test":
         command = gpu_test_command()
     elif args.mode == "import-smoke":
         command = train_command(args, train=False)
     else:
         command = train_command(args)
-=======
-    command = gpu_test_command() if args.mode == "gpu-test" else train_command(args)
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
     print("Launching HF Job:")
     print(f"  mode      = {args.mode}")
     print(f"  namespace = {args.namespace}")

ui/app/components/AgentTrustMonitor.tsx

@@ -30,11 +30,7 @@ function trustColor(t: number) {
 export default function AgentTrustMonitor({
   observation, trustDeltas, activeSpec, events, running, totalReward,
 }: Props) {
-<<<<<<< HEAD
   const agents = observation?.available_specialists ?? ["S0", "S1", "S2", "S3", "S4"];
-=======
-  const agents = observation?.available_specialists || observation?.available_workers || ["S0", "S1", "S2", "S3", "S4"];
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
   const trust = observation?.trust_snapshot ?? {};
   const lastReward = observation?.last_reward ?? 0;
 

ui/app/components/GPUClusterPanel.tsx

@@ -1,12 +1,7 @@
 "use client";
 
-<<<<<<< HEAD
 import { useState, useEffect, useRef } from "react";
 import { motion, AnimatePresence } from "framer-motion";
-=======
-import { useState, useEffect } from "react";
-import { motion } from "framer-motion";
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
 type NodeStatus = "ACTIVE" | "IDLE" | "OVERLOADED" | "FAILED";
 
@@ -18,34 +13,18 @@ interface GPUNode {
   status: NodeStatus;
 }
 
-<<<<<<< HEAD
 export default function GPUClusterPanel() {
   const [mounted, setMounted] = useState(false);
   const [nodes, setNodes] = useState<GPUNode[]>([
     { id: "GPU-1", utilization: 45, memory: 32, load: 1.2, status: "ACTIVE" },
     { id: "GPU-2", utilization: 12, memory: 8, load: 0.4, status: "IDLE" },
     { id: "GPU-3", utilization: 88, memory: 64, load: 2.8, status: "ACTIVE" },
-=======
-interface GPUClusterPanelProps {
-  sessionId?: string;
-  mode?: string;
-  gpuPool?: any[]; // Live data from observation
-}
-
-export default function GPUClusterPanel({ sessionId, mode, gpuPool }: GPUClusterPanelProps) {
-  const [mounted, setMounted] = useState(false);
-  const [nodes, setNodes] = useState<GPUNode[]>([
-    { id: "GPU-1", utilization: 0, memory: 0, load: 0, status: "IDLE" },
-    { id: "GPU-2", utilization: 0, memory: 0, load: 0, status: "IDLE" },
-    { id: "GPU-3", utilization: 0, memory: 0, load: 0, status: "IDLE" },
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
     { id: "GPU-4", utilization: 0, memory: 0, load: 0, status: "IDLE" },
   ]);
 
   const [avgLoad, setAvgLoad] = useState(0);
   const [jitter, setJitter] = useState(0.45);
 
-<<<<<<< HEAD
   useEffect(() => {
     setMounted(true);
     const interval = setInterval(() => {
@@ -73,46 +52,12 @@ export default function GPUClusterPanel({ sessionId, mode, gpuPool }: GPUCluster
     }, 1500);
     return () => clearInterval(interval);
   }, []);
-=======
-  useEffect(() => { setMounted(true); }, []);
-
-  // ── LIVE SYNC FROM OBSERVATION ────────────────────────────
-  useEffect(() => {
-    if (gpuPool && Array.isArray(gpuPool)) {
-      setNodes(gpuPool.slice(0, 4).map((g: any) => {
-        const util = (g.memory_used / g.memory_total) * 100;
-        let status = g.state.toUpperCase();
-        if (status === "ALLOCATED") status = "ACTIVE";
-        
-        return {
-          id: g.id,
-          utilization: util,
-          memory: util,
-          load: (util / 100) * 4.2,
-          status: status as NodeStatus
-        };
-      }));
-    } else if (!sessionId || mode !== "cluster") {
-      // Fallback to subtle idle simulation if no live data
-      const timer = setInterval(() => {
-        setJitter(Math.random() * 0.5);
-        setNodes(prev => prev.map(n => ({
-          ...n,
-          utilization: Math.max(0, n.utilization + (Math.random() - 0.5) * 2),
-          load: n.utilization * 0.04
-        })));
-      }, 2000);
-      return () => clearInterval(timer);
-    }
-  }, [gpuPool, sessionId, mode]);
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
   useEffect(() => {
     const total = nodes.reduce((acc, n) => acc + n.utilization, 0);
     setAvgLoad(total / nodes.length);
   }, [nodes]);
 
-<<<<<<< HEAD
   if (!mounted) {
     return (
       <section className="section-block" id="gpu-cluster" style={{ opacity: 0 }}>
@@ -121,28 +66,19 @@ export default function GPUClusterPanel({ sessionId, mode, gpuPool }: GPUCluster
       </section>
     );
   }
-=======
-  if (!mounted) return null;
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
   return (
     <section className="section-block" id="gpu-cluster">
       <div className="section-label">03 // COMPUTE RESOURCES</div>
       <h2 className="section-title">GPU Compute Clusters</h2>
       <p className="section-desc">
-<<<<<<< HEAD
         Real-time telemetry from the underlying inference hardware.
         High cluster utilization may introduce latency in the trust calibration loop.
-=======
-        Real-time telemetry from the underlying inference hardware. 
-        Note how cluster utilization spikes as the RL model allocates worker jobs.
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
       </p>
 
       <div className="cluster-grid">
         {nodes.map((node) => (
           <div key={node.id} className={`card node-card ${node.status.toLowerCase()}`}>
-<<<<<<< HEAD
             <div className="card-id">{node.id} // NODE-0{node.id.split("-")[1]}</div>
 
             <div className="node-status-badge">
@@ -150,15 +86,6 @@ export default function GPUClusterPanel({ sessionId, mode, gpuPool }: GPUCluster
                 background: node.status === "OVERLOADED" ? "var(--red)" :
                   node.status === "FAILED" ? "#555" :
                     node.status === "IDLE" ? "var(--muted)" : "var(--green)"
-=======
-            <div className="card-id">{node.id} // CORE-AX-{node.id.split("-")[1] || "0X"}</div>
-            
-            <div className="node-status-badge">
-              <div className="status-dot" style={{ 
-                background: node.status === "ACTIVE" ? "var(--green)" :
-                            node.status === "OVERLOADED" ? "var(--red)" : 
-                            node.status === "FAILED" ? "#555" : "var(--muted)" 
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
               }} />
               {node.status}
             </div>
@@ -169,16 +96,9 @@ export default function GPUClusterPanel({ sessionId, mode, gpuPool }: GPUCluster
                 <span style={{ color: "var(--cyan)" }}>{Math.round(node.utilization)}%</span>
               </div>
               <div className="metric-bar-bg">
-<<<<<<< HEAD
                 <motion.div
                   className="metric-bar-fill"
                   animate={{ width: `${node.utilization}%` }}
-=======
-                <motion.div 
-                  className="metric-bar-fill" 
-                  animate={{ width: `${node.utilization}%` }}
-                  transition={{ type: "spring", stiffness: 100, damping: 20 }}
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
                   style={{ background: node.utilization > 90 ? "var(--red)" : "var(--cyan)" } as any}
                 />
               </div>
@@ -190,16 +110,9 @@ export default function GPUClusterPanel({ sessionId, mode, gpuPool }: GPUCluster
                 <span style={{ color: "var(--green)" }}>{Math.round(node.memory)}%</span>
               </div>
               <div className="metric-bar-bg">
-<<<<<<< HEAD
                 <motion.div
                   className="metric-bar-fill"
                   animate={{ width: `${node.memory}%` }}
-=======
-                <motion.div 
-                  className="metric-bar-fill" 
-                  animate={{ width: `${node.memory}%` }}
-                  transition={{ type: "spring", stiffness: 100, damping: 20 }}
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
                   style={{ background: "var(--green)" } as any}
                 />
               </div>
@@ -223,11 +136,7 @@ export default function GPUClusterPanel({ sessionId, mode, gpuPool }: GPUCluster
         <div className="cluster-total-load">
           <span className="label">TOTAL CLUSTER LOAD</span>
           <div className="load-meter-bg">
-<<<<<<< HEAD
             <motion.div
-=======
-            <motion.div 
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
               className="load-meter-fill"
               animate={{ width: `${avgLoad}%` }}
               style={{ background: avgLoad > 80 ? "var(--red)" : "var(--cyan)", color: avgLoad > 80 ? "var(--red)" : "var(--cyan)" } as any}

ui/app/components/SpecialistNetwork.tsx

@@ -17,11 +17,7 @@ export default function SpecialistNetwork({
   trustDeltas: Record<string, number>;
   activeSpec: string | null;
 }) {
-<<<<<<< HEAD
   const ids = observation?.available_specialists ?? ["S0", "S1", "S2", "S3", "S4"];
-=======
-  const ids = observation?.available_specialists || observation?.available_workers || ["S0", "S1", "S2", "S3", "S4"];
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
   return (
     <div className="net">
       <svg className="net-svg" viewBox="0 0 100 100" preserveAspectRatio="xMidYMid meet">

ui/app/components/TrustTimeline.tsx

@@ -10,11 +10,7 @@ export default function TrustTimeline({
   observation: Observation | null;
   trustDeltas: Record<string, number>;
 }) {
-<<<<<<< HEAD
   const ids = observation?.available_specialists ?? ["S0", "S1", "S2", "S3", "S4"];
-=======
-  const ids = observation?.available_specialists ?? observation?.available_workers ?? ["S0", "S1", "S2", "S3", "S4"];
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
   return (
     <div className="tl">
       {ids.map((id) => {

ui/app/hooks/useSentinel.ts

@@ -9,7 +9,6 @@ import type {
 
 /* ── helpers ──────────────────────────────────────────── */
 
-<<<<<<< HEAD
 const API_BASE = process.env.NEXT_PUBLIC_API_URL || "";
 
 function bestSpec(obs: Observation | null): string {
@@ -17,18 +16,6 @@ function bestSpec(obs: Observation | null): string {
   return [...obs.available_specialists].sort(
     (a, b) => (obs.trust_snapshot[b] ?? 0.5) - (obs.trust_snapshot[a] ?? 0.5),
   )[0];
-=======
-const API_BASE = typeof window !== "undefined" 
-  ? (process.env.NEXT_PUBLIC_API_URL || (window.location.port === "3000" || window.location.port === "3458" ? "http://127.0.0.1:7860" : ""))
-  : "";
-
-function bestSpec(obs: Observation | null): string {
-  if (!obs) return "S0";
-  const ids = obs.available_specialists || obs.available_workers || [];
-  return [...ids].sort(
-    (a, b) => (obs.trust_snapshot[b] ?? 0.5) - (obs.trust_snapshot[a] ?? 0.5),
-  )[0] || "S0";
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 }
 
 function heuristicMove(obs: Observation | null) {
@@ -42,14 +29,9 @@ function heuristicMove(obs: Observation | null) {
 
 function randomMove(obs: Observation | null) {
   if (!obs) return { action: "delegate" as ActionType, specialist: "S0", trust: 0.5 };
-<<<<<<< HEAD
   const sp = obs.available_specialists[
     Math.floor(Math.random() * obs.available_specialists.length)
   ] || "S0";
-=======
-  const ids = obs.available_specialists || obs.available_workers || [];
-  const sp = ids[Math.floor(Math.random() * ids.length)] || "S0";
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
   return { action: "delegate" as ActionType, specialist: sp, trust: obs.trust_snapshot[sp] ?? 0.5 };
 }
 
@@ -140,12 +122,7 @@ export function useSentinel() {
   const trustDeltas = useMemo(() => {
     if (!observation) return {};
     const d: Record<string, number> = {};
-<<<<<<< HEAD
     for (const id of observation.available_specialists) {
-=======
-    const ids = observation.available_specialists || observation.available_workers || [];
-    for (const id of ids) {
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
       d[id] = (observation.trust_snapshot[id] ?? 0.5) - (prevTrust[id] ?? 0.5);
     }
     return d;
@@ -173,17 +150,10 @@ export function useSentinel() {
       const s = nextSeed ?? seed;
       setRunning(true);
       abortRef.current = false;
-<<<<<<< HEAD
       const payload = { task_type: t, seed: s };
       setLastReq({ method: "POST", path: "/reset", body: payload });
       try {
         const res = await fetch(`${process.env.NEXT_PUBLIC_API_URL}/reset`, {
-=======
-      const payload = { task_type: t, seed: s, mode: "cluster" };
-      setLastReq({ method: "POST", path: "/reset", body: payload });
-      try {
-        const res = await fetch(`${API_BASE}/reset`, {
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
           method: "POST",
           headers: { "Content-Type": "application/json" },
           body: JSON.stringify(payload),
@@ -225,31 +195,17 @@ export function useSentinel() {
 
       setActiveSpec(specialist);
 
-<<<<<<< HEAD
       const payload = {
         session_id: sid,
         task_type: obs.task_type,
         action_type: action,
-=======
-      const isCluster = active?.info?.environment_mode === "cluster" || sessionId === "cluster";
-      const mappedAction = (isCluster && action === "delegate") ? "allocate" : action;
-
-      const payload = {
-        session_id: sid,
-        task_type: obs.task_type,
-        action_type: mappedAction,
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
         specialist_id: specialist,
         subtask_response: action === "solve_independently" ? "SELF_SOLVED" : null,
         reasoning: `ui-${action}${specialist ? `-${specialist}` : ""}`,
       };
       setLastReq({ method: "POST", path: `/step?session_id=${sid}`, body: payload });
       try {
-<<<<<<< HEAD
         const res = await fetch(`${process.env.NEXT_PUBLIC_API_URL}/step?session_id=${encodeURIComponent(sid)}`, {
-=======
-        const res = await fetch(`${API_BASE}/step?session_id=${encodeURIComponent(sid)}`, {
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
           method: "POST",
           headers: { "Content-Type": "application/json" },
           body: JSON.stringify(payload),

ui/app/lib/types.ts

@@ -14,10 +14,6 @@ export type Observation = {
   subtasks_total: number;
   subtasks_remaining: number;
   available_specialists: string[];
-<<<<<<< HEAD
-=======
-  available_workers?: string[];
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
   trust_snapshot: Record<string, number>;
   stakes_level: number;
   step_count: number;
@@ -25,10 +21,6 @@ export type Observation = {
   last_action_summary: string | null;
   last_reward: number;
   episode_status: string;
-<<<<<<< HEAD
-=======
-  gpu_pool?: any[];
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 };
 
 export type Reward = {
@@ -49,10 +41,6 @@ export type StepResult = {
     score: number;
     adversarial_detections?: number;
     adversarial_poisonings?: number;
-<<<<<<< HEAD
-=======
-    environment_mode?: string;
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
   };
 };
 

ui/app/page.tsx

@@ -256,15 +256,7 @@ export default function Page() {
       <div className="divider" />
 
       {/* GPU CLUSTER */}
-<<<<<<< HEAD
       <GPUClusterPanel />
-=======
-      <GPUClusterPanel 
-        sessionId={s.info?.session_id} 
-        mode={s.info?.environment_mode} 
-        gpuPool={s.observation?.gpu_pool}
-      />
->>>>>>> a89a58750afb4cf3e8d49f13fe66d7c227911387
 
       <div className="divider" />