Upload folder using huggingface_hub

DESIGN.md

@@ -1,6 +1,18 @@
 # SafeSpace Design Document
 
-This document explains the design decisions behind SafeSpace, a content moderation RL environment. It serves as both an internal reference and a public explanation of the benchmark’s design choices.
+This document explains the design decisions behind SafeSpace, a content moderation RL environment. It serves as both an internal reference and a public explanation of the benchmark's design choices.
+
+For setup instructions, action and observation spaces, and usage examples, see `README.md`. This document is focused on benchmark rationale: why the environment is structured this way, how reward and grading were chosen, and what design tradeoffs shape the evaluation.
+
+## Executive Summary
+
+SafeSpace is built around five core design choices:
+
+1. **Moderation is treated as a sequential decision problem, not a one-shot classifier.** Agents must decide when context is necessary and when acting immediately is better.
+2. **Reward and evaluation are aligned.** Training reward and benchmark grading preserve the same ordering of outcomes so agents are not incentivized to exploit mismatched objectives.
+3. **The benchmark is designed to resist easy gaming.** Adjacent decisions receive limited credit, calibration matters, and scenario distributions are balanced enough to discourage trivial hedging strategies.
+4. **Deterministic grading is a feature, not a limitation.** SafeSpace avoids LLM judges in order to maximize reproducibility, auditability, and low-cost evaluation.
+5. **The environment emphasizes realistic moderation ambiguity within a constrained scope.** Context dependence, policy exceptions, repeat-offender logic, and escalation are in scope; multimodal reasoning and multi-agent workflows are intentionally out of scope for now.
 
 ---
 
@@ -100,11 +112,10 @@ We model three concrete moderation workflows rather than three abstract buckets:
 
 We separate the environment into two evaluation layers:
 
-- **Canonical benchmark:** 60 curated scenarios, 20 per task, defined in `server/data/benchmark_manifest.json`
-- **Full corpus:** 367 total scenarios used for broader regression testing and stress coverage
+- **Canonical benchmark:** 60 authored submission scenarios, 20 per task, defined in `server/data/benchmark_manifest.json`
+- **Full corpus:** 367 total scenarios used for broader regression testing and stress coverage, including non-canonical procedural variants
 
 The canonical benchmark is the headline score we report in the README and baseline script. The full corpus still matters, but we do not treat metadata-only procedural variants as benchmark depth.
-Manifest version `2026-04-03.2` also rebalances the hard split away from a remove-heavy concentration by adding one escalation case, one approve satire case, and one warn coded-hate case from the existing corpus.
 
 ### Decision Distribution (Hard Tier)
 
@@ -144,7 +155,7 @@ This is low enough to discourage always-escalate strategies while still rewardin
 
 ### Repeat Offender Scenarios
 
-17 hard scenarios feature "borderline content + repeat offender history = remove". This tests whether agents can integrate author history with surface content analysis, not just pattern-match on text.
+A substantial subset of hard scenarios feature borderline content where repeat-offender history pushes the correct outcome toward `remove`. This tests whether agents can integrate author history with surface content analysis, not just pattern-match on text.
 
 ### Calibration Mechanics
 
@@ -236,30 +247,7 @@ We evaluate agents on:
 
 ### Reference Runs
 
-Primary reference artifact:
-
-`artifacts/baselines/canonical_gpt-5.4_azure_seed7_manifest_2026-04-03.2.json`
-
-This run uses `gpt-5.4` through an OpenAI-compatible Azure AI Foundry endpoint with `OPENAI_SEED=7` on manifest version `2026-04-03.2`.
-
-| Tier | Avg Task Grade | Avg Reward | Avg Raw Reward |
-|------|----------------|------------|----------------|
-| Easy | 0.8244 | 0.7760 | 0.7200 |
-| Medium | 0.4934 | 0.4914 | 0.3643 |
-| Hard | 0.4213 | 0.4695 | 0.3369 |
-
-Secondary open-weight comparison artifact:
-
-`artifacts/baselines/canonical_qwen2.5_72b_hf_seed7_manifest_2026-04-03.2.json`
-
-This comparison run uses `Qwen/Qwen2.5-72B-Instruct` via the Hugging Face Router with `OPENAI_SEED=7`.
-
-| Model | Avg Task Grade | Avg Reward | Avg Raw Reward |
-|-------|----------------|------------|----------------|
-| `gpt-5.4` | 0.5797 | 0.5790 | 0.4737 |
-| `Qwen/Qwen2.5-72B-Instruct` | 0.4775 | 0.5098 | 0.3873 |
-
-These numbers are reference points, not normative targets. The benchmark remains intentionally challenging on context-dependent and hard policy-review cases, and score movement should be interpreted relative to the same manifest and inference setup.
+Reference runs are useful calibration points, but they are not the benchmark's core design rationale. Exact baseline numbers are best interpreted relative to a specific manifest version, model endpoint, and inference configuration, so the current benchmark snapshot is included in the appendix below rather than treated as a permanent part of the design argument.
 
 ---
 
@@ -296,3 +284,50 @@ SafeSpace is designed around these principles:
 5. **Principled defaults**: Episode budget, trajectory caps, score ranges all have explicit rationale
 
 The goal is an environment that teaches real moderation skills, not pattern matching or hedging strategies.
+
+---
+
+## Appendix: Current Benchmark Snapshot
+
+This appendix captures the current shipped benchmark snapshot. It is intentionally separated from the main design rationale because these details may change over time as the corpus, manifest, or reference runs evolve.
+
+### Current Manifest Snapshot
+
+- Canonical benchmark: 60 authored submission scenarios, 20 per task
+- Full corpus: 367 total scenarios, including broader regression/stress coverage outside the canonical benchmark
+- Current manifest version: `2026-04-03.2`
+
+This manifest rebalances the hard split away from an overly remove-heavy concentration while preserving the same task framing and grading philosophy described above.
+
+### Current Reference Runs
+
+Primary reference artifact:
+
+`artifacts/baselines/canonical_gpt-5.4_azure_seed7_manifest_2026-04-03.2.json`
+
+This canonical reference run uses `gpt-5.4` through an OpenAI-compatible Azure AI Foundry endpoint with `OPENAI_SEED=7`.
+
+| Tier | Avg Task Grade | Avg Reward | Avg Raw Reward |
+|------|----------------|------------|----------------|
+| Easy | 0.8327 | 0.7845 | 0.7306 |
+| Medium | 0.4625 | 0.4748 | 0.3435 |
+| Hard | 0.5110 | 0.5382 | 0.4228 |
+
+Overall canonical averages:
+
+- Avg task grade: `0.6021`
+- Avg reward: `0.5992`
+- Avg raw reward: `0.4990`
+
+Secondary open-weight comparison artifact:
+
+- `artifacts/baselines/canonical_qwen2.5_72b_hf_seed7_manifest_2026-04-03.2.json`
+
+This comparison run uses `Qwen/Qwen2.5-72B-Instruct` via the Hugging Face Router with `OPENAI_SEED=7`.
+
+| Model | Avg Task Grade | Avg Reward | Avg Raw Reward |
+|-------|----------------|------------|----------------|
+| `gpt-5.4` | 0.6021 | 0.5992 | 0.4990 |
+| `Qwen/Qwen2.5-72B-Instruct` | 0.4810 | 0.4994 | 0.3742 |
+
+These numbers are reference points, not normative targets. The benchmark remains intentionally challenging on context-dependent and hard policy-review cases, and score movement should be interpreted relative to the same manifest and inference setup.

README.md

@@ -16,25 +16,59 @@ tags:
 
 # SafeSpace: Content Moderation OpenEnv
 
-SafeSpace is an OpenEnv benchmark for multi-step social platform content moderation. Instead of one-shot classification, the agent reviews a reported post, decides whether to gather more evidence under a fixed action budget, and submits a structured moderation decision with calibrated confidence and cited factors.
+SafeSpace is an OpenEnv benchmark for sequential social-platform content moderation. Instead of making a one-shot label prediction, the agent reviews one reported post at a time, decides whether to gather more evidence under a fixed action budget, and then submits a structured moderation decision with calibrated confidence and cited factors.
 
-![OpenEnv 0.2.3](https://img.shields.io/badge/OpenEnv-0.2.3-0B5FFF?style=flat-square)
+![SafeSpace 0.2.1](https://img.shields.io/badge/SafeSpace-0.2.1-0B5FFF?style=flat-square)
 ![FastAPI](https://img.shields.io/badge/FastAPI-009688?style=flat-square)
 ![Docker Ready](https://img.shields.io/badge/Docker-Ready-2563EB?style=flat-square)
 ![HF Space](https://img.shields.io/badge/Hugging%20Face-Space-FFD21E?style=flat-square)
-![OpenAI Client](https://img.shields.io/badge/OpenAI-Client-10A37F?style=flat-square)
 ![License BSD-3-Clause](https://img.shields.io/badge/License-BSD--3--Clause-111827?style=flat-square)
 
-## At a Glance
+Quick links: [HF Space](https://huggingface.co/spaces/Ishangtxl/SafeSpace) · [Run locally](#run-the-environment-locally) · [Run the reference evaluator](#run-the-reference-evaluator) · [Action Space](#action-space) · [Observation Space](#observation-space) · [Example Episode](#example-episode) · [Appendix](#appendix)
 
-- **Real-world task:** sequential content moderation with context gathering, escalation, and calibration
-- **Interface:** typed OpenEnv `step()` / `reset()` / `state()` API backed by FastAPI and a typed client
-- **Benchmark:** 3 graded workflows, a 60-scenario canonical split, and a 367-scenario full corpus
-- **Scoring:** trajectory shaping plus deterministic terminal grading, with no LLM judge in the loop
+## Environment Contract
 
-## Quick Start
+| Item | SafeSpace contract |
+|------|--------------------|
+| Episode unit | One reported content item |
+| Budget | `8` total actions per episode |
+| Action space | `7` investigation actions + `1` terminal `decide` action |
+| Final decisions | `approve`, `remove`, `warn`, `escalate` |
+| Observation | Post content, trigger info, gathered context, platform policy, citable factors, progress fields |
+| Reward | Trajectory shaping during investigation + deterministic terminal grading |
+| Difficulty tiers | `easy`, `medium`, `hard` |
+| Canonical benchmark | `60` scenarios total, `20` per task |
+| Full corpus | `367` scenarios of broader authored + procedural regression coverage |
 
-### 1. Install
+## Environment Description
+
+SafeSpace models moderation as a budgeted evidence-gathering episode over one reported post. The benchmark is designed to test whether an agent can decide when the post alone is enough, when context changes the answer, and when escalation is more appropriate than overconfident guessing.
+
+The benchmark contains three task families:
+
+| Task | Difficulty | Scenarios | What makes it hard |
+|------|------------|-----------|--------------------|
+| `clear_violations` | easy | 117 | Text alone is sufficient; extra investigation wastes reward |
+| `context_dependent` | medium | 140 | One targeted context request often flips the correct decision |
+| `policy_edge_cases` | hard | 110 | Multiple signals conflict; calibration, precedent, and escalation matter |
+
+## Connect To The Deployed Space
+
+Space page: [Ishangtxl/SafeSpace](https://huggingface.co/spaces/Ishangtxl/SafeSpace)
+
+Programmatic endpoint:
+
+```python
+from content_moderation_env import SafeSpaceEnv
+
+with SafeSpaceEnv(base_url="https://ishangtxl-safespace.hf.space").sync() as env:
+    result = env.reset()
+    print(result.observation.content_item.text)
+```
+
+## Run The Environment Locally
+
+### 1. Install dependencies
 
 Preferred:
 
@@ -48,7 +82,9 @@ Alternative:
 pip install -e ".[dev]"
 ```
 
-### 2. Run the full local preflight
+SafeSpace depends on `openenv-core[core]>=0.2.3,<0.3`.
+
+### 2. Run the local preflight
 
 ```bash
 scripts/preflight.sh
@@ -85,93 +121,60 @@ with SafeSpaceEnv(base_url="http://localhost:8000").sync() as env:
     print(result.observation.task_grade)
 ```
 
-### 5. Validate the evaluator contract
-
-```bash
-python inference.py --validate-config
-```
-
-## Episode Flow
-
-![SafeSpace episode flow](artifacts/readme/episode-flow.svg)
-
-### Example Walkthrough: `med_001`
-
-`med_001` looks like a violent threat on the surface:
-
-> "I swear I'm going to kill you next round bro, you're dead meat"
+## Run The Reference Evaluator
 
-The correct policy is not obvious from the text alone, so the optimal agent:
+`inference.py` is the shipped reference evaluator. It uses the real `SafeSpaceEnv` client/server path and supports deterministic `canonical` and `full` evaluation modes.
 
-1. calls `reset(scenario_id="med_001")` and receives the reported post with no gathered context yet
-2. requests `request_thread_context`
-3. sees friendly gaming banter in the thread
-4. submits:
+### Required environment variables
 
-```python
-ModerationAction(
-    action_type="decide",
-    decision="approve",
-    primary_violation="none",
-    severity="none",
-    confidence=0.82,
-    key_factors=["gaming_or_competition_context", "no_violation_found"],
-)
+```bash
+export API_BASE_URL="<OpenAI-compatible endpoint>"
+export MODEL_NAME="<model-id>"
+export HF_TOKEN="<api-key>"
 ```
 
-5. receives both a terminal reward breakdown and a normalized task grade
-
-This is the core SafeSpace idea: the agent is rewarded not just for the final answer, but also for deciding *when* to investigate and *what* to investigate.
-
-## Why This Is an RL Environment, Not a Classifier
-
-A one-shot classifier sees only the post and predicts a label. Real moderation work is sequential:
-
-- moderators inspect the post
-- they choose which context to fetch
-- each fetch has a cost
-- they stop when they have enough evidence
-- they make a decision with a level of certainty
-
-SafeSpace turns that workflow into an RL problem. Investigation actions produce trajectory-level reward, bad investigation habits are penalized, and the terminal reward measures moderation quality, calibration, and efficiency together.
-
-SafeSpace is designed around the parts of moderation that are hard for static classifiers:
-
-- deciding when the surface text is enough
-- deciding when context changes the answer
-- handling asymmetric costs for false negatives vs. over-moderation
-- learning when escalation is better than bluffing
+Accepted credential fallbacks:
 
-## Environment Description
+```bash
+export OPENAI_API_KEY="<api-key>"
+# or
+export API_KEY="<api-key>"
+# or
+export AZURE_OPENAI_API_KEY="<api-key>"
+```
 
-SafeSpace models moderation as a budgeted episode over one reported content item. The environment is designed to test whether an agent can combine policy understanding with evidence gathering instead of relying on surface-text pattern matching alone.
+Additional optional variables:
 
-- **Domain:** social media content moderation
-- **Runtime:** FastAPI / OpenEnv server
-- **Stateful evaluation path:** typed `SafeSpaceEnv` WebSocket client
-- **Public docs path:** typed `/schema` and `/state` endpoints
-- **Decision surface:** `approve`, `remove`, `warn`, `escalate`
-- **Difficulty progression:** easy, medium, hard
+```bash
+export OPENAI_SEED="7"
+export ENV_BASE_URL="http://localhost:8000"
+export LOCAL_IMAGE_NAME="safespace:latest"
+```
 
-## Tasks
+### Validate evaluator config
 
-| Task | Difficulty | Scenarios | What makes it hard |
-|------|------------|-----------|--------------------|
-| `clear_violations` | easy | 117 | Text alone is usually enough; over-investigation wastes reward |
-| `context_dependent` | medium | 140 | One targeted context request often flips the correct answer |
-| `policy_edge_cases` | hard | 110 | Multiple signals conflict; escalation, calibration, and precedent matter |
+```bash
+python inference.py --validate-config
+```
 
-### Workflow 1: Direct Triage (`clear_violations`)
+### Run the canonical benchmark
 
-Obvious spam, explicit threats, doxxing, clear hate speech, and obvious false positives. The optimal policy usually decides immediately.
+```bash
+python inference.py --mode canonical --summary-json-path artifacts/run-summary.json
+```
 
-### Workflow 2: Context-Dependent Investigation (`context_dependent`)
+### Run the full corpus evaluation
 
-Cases where one missing fact changes the right answer: gaming banter, appeal review, brigading victims, repeat offenders, or harmful links hidden behind harmless post text.
+```bash
+python inference.py --mode full --summary-json-path artifacts/run-summary.json
+```
 
-### Workflow 3: Policy / Escalation Review (`policy_edge_cases`)
+Connection precedence:
 
-Ambiguous cases where the agent must combine multiple evidence sources: satire vs. misinformation, coded harassment, cross-cultural language, public-figure privacy edge cases, and quoted harmful claims used for correction or education.
+- if `--env-base-url` is passed, use that server
+- else if `ENV_BASE_URL` is set, use that server
+- else if `LOCAL_IMAGE_NAME` is set, launch the local Docker image through OpenEnv
+- else fall back to `http://localhost:8000`
 
 ## Action Space
 
@@ -187,7 +190,7 @@ Each investigation action consumes one step from the episode budget.
 | `request_similar_precedents` | prior moderation examples | policy conflict and edge-case consistency |
 | `request_reporter_credibility` | reporter accuracy history | brigading or unreliable reporting |
 
-### Terminal action
+Terminal action:
 
 ```python
 ModerationAction(
@@ -231,8 +234,8 @@ print(obs.gathered_context.thread_context)  # None until requested
 
 - `episode_id`
 - `step_count`
-- `task_id` (canonical benchmark task bucket)
-- `scenario_id` (exact sampled scenario)
+- `task_id`
+- `scenario_id`
 - `difficulty`
 - `trigger_type`
 - `actions_taken`
@@ -240,18 +243,33 @@ print(obs.gathered_context.thread_context)  # None until requested
 - `context_requested`
 - `decision_made`
 - `episode_reward`
+- `raw_episode_reward`
+- `done`
+- `last_error_code`
+
+The public HTTP `/schema` and `/state` endpoints expose the typed `ModerationState` contract directly. The session-backed `SafeSpaceEnv` client remains the canonical benchmark path for live per-episode evaluation.
+
+## Example Episode
+
+`med_001` looks like a violent threat on the surface:
 
-The public HTTP `/schema` and `/state` endpoints expose the typed `ModerationState` contract directly. The session-backed `SafeSpaceEnv` client remains the canonical benchmark path for live per-episode state during evaluation.
+> "I swear I'm going to kill you next round bro, you're dead meat"
+
+The correct policy is not obvious from the text alone, so a strong agent typically:
 
-## Reward and Grading
+1. calls `reset(scenario_id="med_001")`
+2. requests `request_thread_context`
+3. sees friendly gaming banter in the thread
+4. submits an `approve` decision with `primary_violation="none"` and factors such as `gaming_or_competition_context`
+5. receives both a terminal reward breakdown and a normalized task grade
 
-SafeSpace provides reward signal over the full trajectory.
+This is the core SafeSpace idea: the agent is rewarded not just for the final answer, but also for deciding when to investigate and what to investigate.
 
-The environment optimizes moderation quality first, then explanation quality, efficiency, and calibration.
+## Reward And Grading
 
-### Trajectory shaping
+SafeSpace optimizes moderation quality first, then explanation quality, efficiency, and calibration.
 
-Investigation steps receive immediate feedback:
+Trajectory shaping:
 
 | Event | Raw reward |
 |-------|------------|
@@ -261,144 +279,93 @@ Investigation steps receive immediate feedback:
 | invalid action | `-0.06` |
 | budget exhausted before decision | `-0.15` |
 
-The cumulative trajectory signal is capped at `+/-0.15`, which lets 3-context hard cases benefit from all useful investigation steps without dominating the terminal reward.
+The cumulative trajectory signal is capped at `+/-0.15`.
 
-### Terminal reward
-
-When the agent takes `decide`, the final reward combines:
+Terminal reward combines:
 
 - decision reward
 - factor overlap reward
 - efficiency bonus
 - calibration bonus
 
-The decision grader is deterministic and checks:
-
-- decision correctness
-- policy violation correctness
-- severity correctness
-- adjacent-decision partial credit
-- dangerous false negatives
+All grading is deterministic and programmatic. There are no LLM judges inside the environment.
 
-Factor reward uses Jaccard similarity between the predicted factor set and the ground-truth factor set.
+## Research Grounding
 
-### Deterministic graders
+SafeSpace is informed by prior content-moderation and explainability research, including HateXplain, especially around rationale-grounded moderation decisions, ambiguity, and explanation quality. We do not claim direct reuse of HateXplain samples in the current shipped scenario corpus.
 
-All tasks are graded programmatically. There are no LLM judges inside the environment.
+- [OpenEnv documentation](https://meta-pytorch.org/OpenEnv/) and the [OpenEnv Quick Start](https://meta-pytorch.org/OpenEnv/auto_getting_started/) describe the environment interface conventions that SafeSpace follows.
+- [HateXplain: A Benchmark Dataset for Explainable Hate Speech Detection](https://arxiv.org/abs/2012.10289) motivates rationale-aware moderation evaluation and explainability-oriented supervision.
+- [Content moderation on social media: Does it matter who and why moderates hate speech?](https://scholars.cityu.edu.hk/en/publications/content-moderation-on-social-media-does-it-matter-who-and-why-mod/) is relevant to moderation transparency, trust, and explanation framing.
 
-- `grade_decision()` scores the moderation action against ground truth
-- `grade_factors()` scores cited rationale factors
-- `compute_reward()` combines terminal moderation quality
-- trajectory reward helpers score investigation behavior step by step
+## Appendix
 
-Public reward values are normalized into `[0.0, 1.0]` for compatibility with tooling that expects normalized reward signals. Raw signed reward values are preserved through `raw_*` fields for RL analysis and debugging.
-
-## Benchmark Splits and Scenario Stats
+<details>
+<summary><strong>Benchmark splits and corpus statistics</strong></summary>
 
 Canonical benchmark:
 
 - Manifest version: `2026-04-03.2`
-- Canonical split size: 60 scenarios total
-- Canonical composition: 20 scenarios per benchmark task
+- Canonical split size: `60` scenarios total
+- Canonical composition: `20` scenarios per benchmark task
 - Canonical benchmark is the headline evaluation set used by `inference.py --mode canonical`
 
 Full corpus:
 
-- Total scenarios: 367
-- Easy: 117
-- Medium: 140
-- Hard: 110
+- Total scenarios: `367`
+- Easy: `117`
+- Medium: `140`
+- Hard: `110`
+- Includes broader authored extensions plus procedural stress/regression coverage outside the official canonical benchmark
 
 Decision distribution:
 
-- Approve: 139
-- Remove: 133
-- Warn: 54
-- Escalate: 41
+- Approve: `139`
+- Remove: `133`
+- Warn: `54`
+- Escalate: `41`
 
 Trigger distribution:
 
-- `user_report`: 210
-- `auto_flag`: 109
-- `appeal`: 27
-- `proactive_audit`: 21
+- `user_report`: `210`
+- `auto_flag`: `109`
+- `appeal`: `27`
+- `proactive_audit`: `21`
 
 Context-depth distribution:
 
-- `0` context requests needed: 117
-- `1` context request needed: 123
-- `2` context requests needed: 108
-- `3` context requests needed: 19
-
-The full corpus includes targeted link, policy-conflict, satire, privacy, and coded-harassment edge cases, while the 60-scenario canonical split remains the headline benchmark for fast and stable comparison.
+- `0` context requests needed: `117`
+- `1` context request needed: `123`
+- `2` context requests needed: `108`
+- `3` context requests needed: `19`
 
-## Reference Evaluation
+</details>
 
-`inference.py` is the reference evaluator shipped with the project.
+<details>
+<summary><strong>Reference evaluation details and baseline results</strong></summary>
 
-It:
+`inference.py`:
 
 - uses the OpenAI client for all model calls
 - works with any OpenAI-compatible endpoint exposed through `API_BASE_URL`
 - prefers `HF_TOKEN` and still accepts `OPENAI_API_KEY`, `API_KEY`, or `AZURE_OPENAI_API_KEY` as fallbacks
-- evaluates through the real `SafeSpaceEnv` client/server path
 - validates the benchmark manifest before running
-- supports deterministic `canonical` and `full` modes
 - emits compact `[START]`, `[STEP]`, and `[END]` logs on stdout
-- writes aggregate metrics only when `--summary-json-path` is provided
-
-### Required environment variables
-
-```bash
-export API_BASE_URL="<OpenAI-compatible endpoint>"
-export MODEL_NAME="<model-id>"
-export HF_TOKEN="<api-key>"
-```
-
-Accepted credential fallbacks:
-
-```bash
-export OPENAI_API_KEY="<api-key>"
-# or
-export API_KEY="<api-key>"
-# or
-export AZURE_OPENAI_API_KEY="<api-key>"
-```
+- writes aggregate metrics when `--summary-json-path` is provided
 
-Additional optional variables:
-
-```bash
-export OPENAI_SEED="7"
-export ENV_BASE_URL="http://localhost:8000"
-export LOCAL_IMAGE_NAME="safespace:latest"
-```
-
-Connection precedence:
-
-- if `--env-base-url` is passed, use that server
-- else if `ENV_BASE_URL` is set, use that server
-- else if `LOCAL_IMAGE_NAME` is set, launch the local Docker image through OpenEnv
-- else fall back to `http://localhost:8000`
-
-### Run the canonical reference baseline
-
-```bash
-python inference.py --mode canonical --summary-json-path artifacts/run-summary.json
-```
+Public score semantics:
 
-### Validate config and benchmark assets before running
+- Final episode `score` in the `[END]` line is always `task_grade`, which lives in `[0.0, 1.0]`
+- Public `reward` values are normalized into `[0.0, 1.0]`
+- Raw signed reward values are preserved in the JSON summary, state, and reward breakdown payloads
 
-```bash
-python inference.py --validate-config
-```
+Canonical vs. full:
 
-### Run the full dataset evaluation
+- `canonical` evaluates the fixed 60-scenario benchmark in `server/data/benchmark_manifest.json`
+- `full` evaluates the entire current corpus of 367 scenarios
+- `full` runs the canonical split first and then the remainder of the corpus
 
-```bash
-python inference.py --mode full --summary-json-path artifacts/run-summary.json
-```
-
-The script emits stdout in the following single-line format:
+The evaluator emits stdout in the following single-line format:
 
 ```text
 [START] task=<task_name> env=safespace model=<model_name>
@@ -406,56 +373,42 @@ The script emits stdout in the following single-line format:
 [END] success=<true|false> steps=<n> score=<score> rewards=<r1,r2,...,rn>
 ```
 
-When `--summary-json-path` is set, the file records aggregate metrics, per-task breakdowns, decision distributions, runtime metadata, and failure details.
-
-### Public score semantics
-
-- Final episode `score` in the `[END]` line is always `task_grade`, which already lives in `[0.0, 1.0]`.
-- Public `reward` values are normalized into `[0.0, 1.0]` for compatibility with tooling that expects normalized reward signals.
-- Raw signed reward values are preserved in the JSON summary, state, and reward breakdown payloads for debugging and RL analysis.
-
-### Canonical vs. full evaluation
-
-- `canonical` evaluates a fixed 60-scenario benchmark defined in `server/data/benchmark_manifest.json`.
-- `canonical` is the score we treat as the headline benchmark because it is fast, comparable across reruns, and stable in task composition.
-- `full` evaluates the entire current corpus of 367 scenarios.
-- `full` is intended as a broader stress test and regression suite; it is slower and more expensive, but gives a better picture of full-corpus generalization.
-- `full` runs the canonical split first and then the remainder of the corpus, so `--limit-per-task` is useful for cheap smoke checks.
-
-### Current reference baselines
-
 Primary reference artifact:
 
 `artifacts/baselines/canonical_gpt-5.4_azure_seed7_manifest_2026-04-03.2.json`
 
-This is the main headline reference run. It uses `gpt-5.4` through an OpenAI-compatible Azure AI Foundry endpoint with `OPENAI_SEED=7` on benchmark manifest version `2026-04-03.2`.
+This canonical reference run uses `gpt-5.4` through an OpenAI-compatible Azure AI Foundry endpoint with `OPENAI_SEED=7`.
 
 | Task | Difficulty | Avg Task Grade | Avg Reward | Avg Raw Reward | Decision Distribution |
 |------|------------|----------------|------------|----------------|----------------------|
-| `clear_violations` | easy | **0.8244** | `0.7760` | `0.7200` | remove: 10, approve: 10 |
-| `context_dependent` | medium | **0.4934** | `0.4914` | `0.3643` | approve: 11, warn: 5, remove: 4 |
-| `policy_edge_cases` | hard | **0.4213** | `0.4695` | `0.3369` | approve: 8, escalate: 5, warn: 4, remove: 3 |
+| `clear_violations` | easy | `0.8327` | `0.7845` | `0.7306` | remove: 10, approve: 9, warn: 1 |
+| `context_dependent` | medium | `0.4625` | `0.4748` | `0.3435` | approve: 10, warn: 6, remove: 4 |
+| `policy_edge_cases` | hard | `0.5110` | `0.5382` | `0.4228` | escalate: 7, approve: 7, warn: 3, remove: 3 |
 
-**Overall average task grade: `0.5797`**
+Overall:
 
-**Overall average reward: `0.5790`**
+- Avg task grade: `0.6021`
+- Avg reward: `0.5992`
+- Avg raw reward: `0.4990`
+- Failed episodes: `0`
 
-**Overall average raw reward: `0.4737`**
+Secondary open-weight comparison artifact:
 
-Secondary open-weight reference artifact:
+- `artifacts/baselines/canonical_qwen2.5_72b_hf_seed7_manifest_2026-04-03.2.json`
 
-`artifacts/baselines/canonical_qwen2.5_72b_hf_seed7_manifest_2026-04-03.2.json`
-
-This comparison run uses `Qwen/Qwen2.5-72B-Instruct` via the Hugging Face Router with `HF_TOKEN` and `OPENAI_SEED=7`.
+This comparison run uses `Qwen/Qwen2.5-72B-Instruct` via the Hugging Face Router with `OPENAI_SEED=7`.
 
 | Model | Avg Task Grade | Avg Reward | Avg Raw Reward | Failed Episodes |
 |-------|----------------|------------|----------------|-----------------|
-| `gpt-5.4` | **0.5797** | `0.5790` | `0.4737` | `0` |
-| `Qwen/Qwen2.5-72B-Instruct` | `0.4775` | `0.5098` | `0.3873` | `0` |
+| `gpt-5.4` | **`0.6021`** | `0.5992` | `0.4990` | `0` |
+| `Qwen/Qwen2.5-72B-Instruct` | `0.4810` | `0.4994` | `0.3742` | `0` |
+
+</details>
 
-## Validation and Deployment
+<details>
+<summary><strong>Validation and deployment</strong></summary>
 
-Run the full verification stack from the repository root containing `openenv.yaml`, `inference.py`, and `Dockerfile`:
+Run the full verification stack from the repository root:
 
 ```bash
 scripts/preflight.sh
@@ -488,13 +441,16 @@ openenv push --repo-id <username>/safespace
 Submission-path hardening notes:
 
 - the Docker build, health check, and typed-client smoke path are covered by `scripts/preflight.sh`
-- the package asset smoke test now self-heals `pip` with `ensurepip` if the active interpreter does not expose `python -m pip`
+- the package asset smoke test self-heals `pip` with `ensurepip` if the active interpreter does not expose `python -m pip`
 - the canonical evaluation path has been verified under a `2 CPU / 8 GB` container budget and completed well under the hackathon's `20` minute runtime limit
 
-## Project Layout
+</details>
+
+<details>
+<summary><strong>Project layout</strong></summary>
 
 ```text
-content_moderation_env/
+SafeSpace_Env/
 ├── artifacts/
 │   ├── baselines/
 │   └── readme/
@@ -507,6 +463,7 @@ content_moderation_env/
 ├── README.md
 ├── scripts/
 │   ├── check_package_assets.py
+│   ├── generate_scenarios.py
 │   ├── preflight.sh
 │   ├── report_stats.py
 │   └── validate-submission.sh
@@ -521,6 +478,8 @@ content_moderation_env/
 └── tests/
 ```
 
+</details>
+
 ## License
 
 This project is licensed under the BSD 3-Clause License. See `LICENSE`.

inference.py

@@ -1663,7 +1663,9 @@ async def _async_main(args: argparse.Namespace) -> None:
     failure_details: List[Dict[str, Any]] = []
 
     client = await create_env_client(args.env_base_url)
-    async with client:
+    try:
+        if hasattr(client, "connect"):
+            await client.connect()
         for task_id in TASK_TO_DIFFICULTY:
             scenario_ids = load_scenario_ids(task_id, args.mode)
             if args.limit_per_task is not None:
@@ -1676,6 +1678,12 @@ async def _async_main(args: argparse.Namespace) -> None:
             )
             task_summaries[task_id] = task_summary
             failure_details.extend(task_failures)
+    finally:
+        try:
+            if hasattr(client, "close"):
+                await client.close()
+        except Exception:
+            pass
 
     total_scenarios = sum(
         summary["num_scenarios"] for summary in task_summaries.values()

server/data/scenarios_easy.json

@@ -5721,9 +5721,9 @@
     "category": "obvious_spam",
     "content_item": {
       "post_id": "p_83740",
-      "text": "FREE FOLLOWERS! Get 10K real followers for just $5! DM now! Limited offer!",
-      "author_id": "user_xi_574",
-      "community": "debate",
+      "text": "Need 5,000 'verified' followers by tomorrow? Launch special for new creator accounts. DM me before midnight.",
+      "author_id": "growthplug_574",
+      "community": "creators",
       "timestamp": "2026-02-26T13:57:00Z",
       "media_type": "text",
       "media_description": null
@@ -5739,20 +5739,19 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Foodie",
-        "account_age_days": 549,
-        "follower_count": 3749,
-        "post_count": 656,
+        "bio": "Growth consultant | DM for collabs",
+        "account_age_days": 41,
+        "follower_count": 118,
+        "post_count": 63,
         "communities": [
-          "wholesome",
-          "news",
-          "food",
-          "books"
+          "creators",
+          "instagram",
+          "marketing"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Creator communities prohibit follower-selling, engagement scams, and paid growth promises.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -5766,7 +5765,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: obvious_spam"
+      "explanation": "The post is a direct commercial offer to sell followers and platform growth. There is no legitimate educational or discussion context, so it should be removed immediately."
     }
   },
   {
@@ -5775,9 +5774,9 @@
     "category": "obvious_spam",
     "content_item": {
       "post_id": "p_84444",
-      "text": "MAKE $5000 DAILY working from home! Click here for secret method!",
-      "author_id": "ai_master_240",
-      "community": "photography",
+      "text": "I replaced my salary with this remote-income system in 30 days. Comment START and I'll send the exact steps.",
+      "author_id": "remotewins_240",
+      "community": "careers",
       "timestamp": "2026-02-26T21:08:00Z",
       "media_type": "text",
       "media_description": null
@@ -5793,20 +5792,19 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Music lover",
-        "account_age_days": 532,
-        "follower_count": 4099,
-        "post_count": 1438,
+        "bio": "Remote freedom mentor",
+        "account_age_days": 27,
+        "follower_count": 92,
+        "post_count": 48,
         "communities": [
-          "wholesome",
-          "crypto",
-          "todayilearned",
-          "sports"
+          "careers",
+          "sidehustle",
+          "remote_work"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Career communities allow job advice but prohibit get-rich-quick promotions and lead funnels.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -5820,7 +5818,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: obvious_spam"
+      "explanation": "This is a classic work-from-home funnel framed as career advice. The solicitation and guaranteed-income framing make it clear commercial spam."
     }
   },
   {
@@ -5880,9 +5878,9 @@
     "category": "obvious_spam",
     "content_item": {
       "post_id": "p_61643",
-      "text": "Congratulations! You've won $1,000,000! Click to claim your prize!",
-      "author_id": "power_user_336",
-      "community": "politics",
+      "text": "You were selected for our $1,000 gift-card draw. Confirm in the next 10 minutes or the prize is forfeited.",
+      "author_id": "flashbonus_336",
+      "community": "deals",
       "timestamp": "2026-03-06T07:52:00Z",
       "media_type": "text",
       "media_description": null
@@ -5898,20 +5896,18 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Tech enthusiast",
-        "account_age_days": 893,
-        "follower_count": 830,
-        "post_count": 1767,
+        "bio": "Giveaway drops daily",
+        "account_age_days": 13,
+        "follower_count": 57,
+        "post_count": 31,
         "communities": [
-          "crypto",
-          "wholesome",
-          "music",
-          "travel"
+          "deals",
+          "giveaways"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Communities do not allow fake prize claims, deceptive urgency, or off-platform claim prompts.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -5925,7 +5921,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: obvious_spam"
+      "explanation": "The prize language, forced urgency, and claim-now structure are typical scam-spam patterns. The text alone is enough to remove it."
     }
   },
   {
@@ -5934,9 +5930,9 @@
     "category": "obvious_spam",
     "content_item": {
       "post_id": "p_92342",
-      "text": "Buy cheap medications online! No prescription needed! Best prices!",
-      "author_id": "ml_expert_463",
-      "community": "fashion",
+      "text": "Need prescription meds without the paperwork? Overnight shipping and no doctor questions. Link in bio.",
+      "author_id": "quickrx_463",
+      "community": "health",
       "timestamp": "2026-02-28T22:22:00Z",
       "media_type": "text",
       "media_description": null
@@ -5952,17 +5948,18 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Critical thinker",
-        "account_age_days": 945,
-        "follower_count": 448,
-        "post_count": 828,
+        "bio": "Discount wellness source",
+        "account_age_days": 19,
+        "follower_count": 44,
+        "post_count": 22,
         "communities": [
-          "art"
+          "health",
+          "supplements"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Health communities prohibit illicit prescription sales and unsafe medical commerce.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -5976,7 +5973,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: obvious_spam"
+      "explanation": "The post advertises illicit medication sales and explicitly bypasses prescription safeguards. It is a straightforward high-severity spam and safety violation."
     }
   },
   {
@@ -6257,9 +6254,9 @@
     "category": "false_positive_metaphor",
     "content_item": {
       "post_id": "p_56101",
-      "text": "This new album absolutely KILLS IT! Best music I've heard all year!",
-      "author_id": "user_nu_156",
-      "community": "finance",
+      "text": "This new album absolutely kills from track one to the encore. Best release I have heard all year.",
+      "author_id": "vinylnotes_156",
+      "community": "music",
       "timestamp": "2026-03-13T17:12:00Z",
       "media_type": "text",
       "media_description": null
@@ -6275,14 +6272,13 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Foodie",
+        "bio": "Live-show addict | Indie records",
         "account_age_days": 725,
-        "follower_count": 3792,
-        "post_count": 1702,
+        "follower_count": 792,
+        "post_count": 412,
         "communities": [
-          "food",
-          "movies",
-          "crypto"
+          "music",
+          "indieheads"
         ]
       },
       "author_violations": [],
@@ -6302,7 +6298,7 @@
         "auto_flag_low_confidence"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: false_positive_metaphor"
+      "explanation": "The violent term is clearly metaphorical praise for a music release. There is no threat, target, or call for harm."
     }
   },
   {
@@ -6311,9 +6307,9 @@
     "category": "false_positive_metaphor",
     "content_item": {
       "post_id": "p_94097",
-      "text": "My presentation totally bombed today, so embarrassing",
-      "author_id": "ml_expert_834",
-      "community": "sports",
+      "text": "I completely bombed my thesis-defense practice today, but at least I know what to fix now.",
+      "author_id": "gradlife_834",
+      "community": "college",
       "timestamp": "2026-02-28T22:45:00Z",
       "media_type": "text",
       "media_description": null
@@ -6329,13 +6325,13 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Student",
+        "bio": "First-year PhD student",
         "account_age_days": 316,
-        "follower_count": 3600,
-        "post_count": 356,
+        "follower_count": 188,
+        "post_count": 97,
         "communities": [
-          "memes",
-          "general"
+          "college",
+          "gradschool"
         ]
       },
       "author_violations": [],
@@ -6355,7 +6351,7 @@
         "auto_flag_low_confidence"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: false_positive_metaphor"
+      "explanation": "\"Bombed\" is a common idiom for performing poorly. The post is self-reflection about schoolwork, not violence."
     }
   },
   {
@@ -6364,9 +6360,9 @@
     "category": "false_positive_metaphor",
     "content_item": {
       "post_id": "p_49881",
-      "text": "I'm dying to try that new restaurant everyone's talking about",
-      "author_id": "user_nu_744",
-      "community": "memes",
+      "text": "I'm dying to try that ramen place everyone keeps posting about.",
+      "author_id": "noodlemap_744",
+      "community": "food",
       "timestamp": "2026-03-06T04:25:00Z",
       "media_type": "text",
       "media_description": null
@@ -6382,15 +6378,13 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Student",
+        "bio": "Weekend food crawl organizer",
         "account_age_days": 784,
-        "follower_count": 4276,
-        "post_count": 1889,
+        "follower_count": 530,
+        "post_count": 286,
         "communities": [
-          "movies",
-          "books",
-          "finance",
-          "music"
+          "food",
+          "cityguides"
         ]
       },
       "author_violations": [],
@@ -6410,7 +6404,7 @@
         "auto_flag_low_confidence"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: false_positive_metaphor"
+      "explanation": "The phrase \"dying to try\" is ordinary hyperbole expressing enthusiasm. There is no threatening or self-harm meaning in context."
     }
   },
   {
@@ -6419,9 +6413,9 @@
     "category": "false_positive_metaphor",
     "content_item": {
       "post_id": "p_92199",
-      "text": "This code is murder to debug, been at it for hours",
-      "author_id": "casual_user_313",
-      "community": "fashion",
+      "text": "This deployment bug is murdering my weekend, but I think I finally found the race condition.",
+      "author_id": "debugloop_313",
+      "community": "programming",
       "timestamp": "2026-03-12T15:50:00Z",
       "media_type": "text",
       "media_description": null
@@ -6437,12 +6431,13 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Foodie",
+        "bio": "Backend engineer | Sleeps after incident response",
         "account_age_days": 435,
-        "follower_count": 1817,
-        "post_count": 261,
+        "follower_count": 261,
+        "post_count": 144,
         "communities": [
-          "fashion"
+          "programming",
+          "devops"
         ]
       },
       "author_violations": [],
@@ -6462,7 +6457,7 @@
         "auto_flag_low_confidence"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: false_positive_metaphor"
+      "explanation": "The wording is a programming metaphor about a frustrating bug. It is clearly non-violent and should be approved."
     }
   },
   {
@@ -6471,9 +6466,9 @@
     "category": "false_positive_metaphor",
     "content_item": {
       "post_id": "p_85354",
-      "text": "The competition was brutal but we crushed them in the finals!",
-      "author_id": "user_eta_207",
-      "community": "food",
+      "text": "The finals were brutal, but our relay team crushed the last lap and stole the meet.",
+      "author_id": "trackcap_207",
+      "community": "sports",
       "timestamp": "2026-03-03T13:12:00Z",
       "media_type": "text",
       "media_description": null
@@ -6489,13 +6484,13 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Full-time dreamer",
+        "bio": "High-school relay coach",
         "account_age_days": 427,
-        "follower_count": 1177,
-        "post_count": 1999,
+        "follower_count": 341,
+        "post_count": 208,
         "communities": [
-          "health",
-          "art"
+          "sports",
+          "trackandfield"
         ]
       },
       "author_violations": [],
@@ -6515,7 +6510,7 @@
         "auto_flag_low_confidence"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: false_positive_metaphor"
+      "explanation": "The language is standard sports hyperbole about competition and winning. There is no credible threat or abusive target."
     }
   },
   {
@@ -6907,12 +6902,12 @@
     "category": "scam_link",
     "content_item": {
       "post_id": "p_98147",
-      "text": "URGENT: Your account will be deleted! Verify now: bit.ly/scam123",
-      "author_id": "player_one_451",
-      "community": "science",
+      "text": "Security alert: your forum account will be disabled tonight unless you confirm ownership at bit.ly/secure-check.",
+      "author_id": "alertdesk_451",
+      "community": "techsupport",
       "timestamp": "2026-03-01T15:57:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Shortened link claiming to verify account ownership"
     },
     "trigger_info": {
       "trigger_type": "auto_flag",
@@ -6925,19 +6920,18 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Coffee addict",
-        "account_age_days": 540,
-        "follower_count": 2937,
-        "post_count": 645,
+        "bio": "Account recovery help desk",
+        "account_age_days": 6,
+        "follower_count": 14,
+        "post_count": 9,
         "communities": [
-          "debate",
-          "photography",
-          "todayilearned"
+          "techsupport",
+          "accounthelp"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Communities prohibit phishing links, impersonation, and fake security alerts.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -6952,7 +6946,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: scam_link"
+      "explanation": "This is a phishing lure using false account-disable urgency and a shortened verification link. The text alone is enough for removal."
     }
   },
   {
@@ -7015,12 +7009,12 @@
     "category": "scam_link",
     "content_item": {
       "post_id": "p_99088",
-      "text": "Your package is waiting! Pay $2 shipping: tracking-scam.com/pay",
-      "author_id": "dev_ops_665",
-      "community": "fitness",
+      "text": "Courier says your parcel is waiting. Pay the $1.99 redelivery fee here to release it today: tracking-scam.com/pay",
+      "author_id": "parcelupdate_665",
+      "community": "shopping",
       "timestamp": "2026-03-01T23:35:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Parcel redelivery fee link"
     },
     "trigger_info": {
       "trigger_type": "auto_flag",
@@ -7033,18 +7027,18 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Professional lurker",
-        "account_age_days": 481,
-        "follower_count": 4698,
-        "post_count": 1708,
+        "bio": "Delivery notice center",
+        "account_age_days": 8,
+        "follower_count": 11,
+        "post_count": 7,
         "communities": [
-          "general",
-          "finance"
+          "shopping",
+          "deals"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Communities prohibit payment collection scams and fake shipping notices.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -7059,7 +7053,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: scam_link"
+      "explanation": "The post uses a common fake-shipping-payment scam pattern. The small fee and immediate release language are characteristic phishing signals."
     }
   },
   {
@@ -7068,12 +7062,12 @@
     "category": "scam_link",
     "content_item": {
       "post_id": "p_31273",
-      "text": "IRS NOTICE: You owe back taxes! Pay immediately or face arrest: irs-scam.net",
-      "author_id": "user_kappa_151",
-      "community": "health",
+      "text": "State revenue notice: resolve your overdue tax balance today or enforcement action begins. Submit payment at irs-scam.net.",
+      "author_id": "revenueresolve_151",
+      "community": "personalfinance",
       "timestamp": "2026-02-13T09:03:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Fake tax-payment portal link"
     },
     "trigger_info": {
       "trigger_type": "auto_flag",
@@ -7086,17 +7080,18 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Coffee addict",
-        "account_age_days": 709,
-        "follower_count": 1462,
-        "post_count": 1448,
+        "bio": "Tax resolution hotline",
+        "account_age_days": 5,
+        "follower_count": 9,
+        "post_count": 6,
         "communities": [
-          "fashion"
+          "personalfinance",
+          "taxes"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Communities prohibit impersonation, fake billing notices, and coercive payment links.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -7111,7 +7106,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: scam_link"
+      "explanation": "This is a coercive tax-payment phishing message using false government authority. It should be removed immediately."
     }
   },
   {
@@ -7120,12 +7115,12 @@
     "category": "scam_link",
     "content_item": {
       "post_id": "p_97277",
-      "text": "Apple Support: Your iCloud was hacked! Reset password: apple-phish.com",
-      "author_id": "user_alpha_826",
-      "community": "askexperts",
+      "text": "Apple Security: we blocked a suspicious login on your iCloud. Review and reset your password now at appleid-check-secure.com.",
+      "author_id": "appleshield_826",
+      "community": "applehelp",
       "timestamp": "2026-03-11T23:43:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Password-reset phishing link"
     },
     "trigger_info": {
       "trigger_type": "auto_flag",
@@ -7138,20 +7133,18 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Learning every day",
-        "account_age_days": 732,
-        "follower_count": 4047,
-        "post_count": 458,
+        "bio": "Apple account alerts",
+        "account_age_days": 4,
+        "follower_count": 7,
+        "post_count": 5,
         "communities": [
-          "science",
-          "politics",
-          "askexperts",
-          "photography"
+          "applehelp",
+          "techsupport"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Communities prohibit fake support messages and credential-harvesting links.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -7166,7 +7159,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: scam_link"
+      "explanation": "The post impersonates a support or security notice to steal credentials. It is a clear harmful-link and commercial scam case."
     }
   },
   {
@@ -7175,9 +7168,9 @@
     "category": "legitimate_question",
     "content_item": {
       "post_id": "p_69170",
-      "text": "Can anyone recommend a good Python IDE? VS Code vs PyCharm?",
-      "author_id": "dev_ops_744",
-      "community": "memes",
+      "text": "Can anyone recommend a beginner-friendly Python IDE? I am choosing between VS Code and PyCharm for my first class.",
+      "author_id": "intro2py_744",
+      "community": "programming",
       "timestamp": "2026-03-08T23:19:00Z",
       "media_type": "text",
       "media_description": null
@@ -7193,12 +7186,13 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Free spirit",
-        "account_age_days": 793,
-        "follower_count": 4895,
-        "post_count": 312,
+        "bio": "Starting a CS degree this semester",
+        "account_age_days": 3,
+        "follower_count": 0,
+        "post_count": 1,
         "communities": [
-          "tech"
+          "programming",
+          "learnpython"
         ]
       },
       "author_violations": [],
@@ -7217,7 +7211,7 @@
         "auto_flag_low_confidence"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: legitimate_question"
+      "explanation": "This is a straightforward beginner question about development tools. The auto-flag comes from new-account heuristics, but there is no policy violation."
     }
   },
   {
@@ -7226,9 +7220,9 @@
     "category": "legitimate_question",
     "content_item": {
       "post_id": "p_51693",
-      "text": "What's the best way to learn machine learning as a beginner?",
-      "author_id": "user_zeta_622",
-      "community": "fitness",
+      "text": "What is the best roadmap for learning machine learning if I only know basic Python and high-school statistics?",
+      "author_id": "mlstarter_622",
+      "community": "machinelearning",
       "timestamp": "2026-02-18T05:32:00Z",
       "media_type": "text",
       "media_description": null
@@ -7244,12 +7238,13 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Gaming is life",
-        "account_age_days": 252,
-        "follower_count": 4731,
-        "post_count": 1237,
+        "bio": "Learning after work",
+        "account_age_days": 2,
+        "follower_count": 0,
+        "post_count": 1,
         "communities": [
-          "news"
+          "machinelearning",
+          "learnprogramming"
         ]
       },
       "author_violations": [],
@@ -7268,7 +7263,7 @@
         "auto_flag_low_confidence"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: legitimate_question"
+      "explanation": "The post asks for educational guidance and contains no harmful or abusive content. It should be approved despite the low-confidence auto-flag."
     }
   },
   {
@@ -7486,8 +7481,8 @@
     "category": "bot_spam",
     "content_item": {
       "post_id": "p_39570",
-      "text": "Great post! Check out my profile for more amazing content! #follow4follow",
-      "author_id": "user_iota_905",
+      "text": "Love this itinerary. Follow my page for daily travel hacks, giveaway alerts, and cheap-flight drops. #follow4follow",
+      "author_id": "tripboost_905",
       "community": "travel",
       "timestamp": "2026-02-15T08:50:00Z",
       "media_type": "text",
@@ -7504,18 +7499,18 @@
     },
     "available_context": {
       "author_profile": {
-        "bio": "Building cool stuff",
-        "account_age_days": 101,
-        "follower_count": 2139,
-        "post_count": 1017,
+        "bio": "Travel growth page",
+        "account_age_days": 12,
+        "follower_count": 21,
+        "post_count": 104,
         "communities": [
-          "askexperts",
-          "sports"
+          "travel",
+          "deals"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Communities prohibit engagement bait, follow-exchange spam, and mass self-promotion.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -7529,7 +7524,7 @@
         "clear_violation_no_exception"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: bot_spam"
+      "explanation": "The post is engagement bait aimed at driving follows rather than participating in discussion. The hashtag and self-promotional pattern make this a clear spam-engagement case."
     }
   },
   {
@@ -7751,8 +7746,8 @@
     "category": "normal_disagreement",
     "content_item": {
       "post_id": "p_62941",
-      "text": "That's a terrible take and here's why I think you're wrong",
-      "author_id": "full_stack_919",
+      "text": "That's a rough take, but I disagree. The live version works better because the studio mix sounds flat.",
+      "author_id": "mixnotes_919",
       "community": "music",
       "timestamp": "2026-02-21T08:39:00Z",
       "media_type": "text",
@@ -7764,20 +7759,20 @@
       "report_categories": [
         "harassment"
       ],
-      "sample_report_reason": null,
+      "sample_report_reason": "Dismissive reply to my opinion",
       "original_decision": null,
       "appeal_text": null,
       "audit_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Learning every day",
+        "bio": "Sound engineer | Live bootleg collector",
         "account_age_days": 448,
-        "follower_count": 4330,
-        "post_count": 671,
+        "follower_count": 430,
+        "post_count": 167,
         "communities": [
-          "politics",
-          "health"
+          "music",
+          "audiophile"
         ]
       },
       "author_violations": [],
@@ -7795,7 +7790,7 @@
         "no_violation_found"
       ],
       "context_needed": [],
-      "explanation": "Generated scenario for category: normal_disagreement"
+      "explanation": "The reply is direct disagreement about a music opinion, but it does not become a personal attack. This is normal debate and should be approved."
     }
   },
   {

server/data/scenarios_hard.json

@@ -6744,48 +6744,55 @@
     "category": "coded_hate",
     "content_item": {
       "post_id": "p_61801",
-      "text": "Check the early life section, it tells you everything you need to know",
-      "author_id": "user_beta_974",
+      "text": "You can usually predict how these decisions go once you check the contributor's 'early life' section.",
+      "author_id": "stacktrace_patriot_974",
       "community": "tech",
       "timestamp": "2026-02-20T01:54:00Z",
       "media_type": "text",
       "media_description": null
     },
     "trigger_info": {
+      "trigger_type": "proactive_audit",
+      "audit_reason": "high_risk_user_sample",
       "report_categories": [],
       "sample_report_reason": null,
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": "high_risk_user_sample",
-      "trigger_type": "proactive_audit"
+      "auto_flag_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Just here to chat",
+        "bio": "Open-source maximalist | policy skeptic",
         "account_age_days": 599,
         "follower_count": 1480,
         "post_count": 516,
         "communities": [
-          "crypto",
-          "fitness",
-          "politics",
-          "news"
+          "tech",
+          "startups",
+          "politics"
         ]
       },
-      "author_violations": [],
+      "author_violations": [
+        {
+          "type": "coded_hate",
+          "date": "2026-01-18",
+          "severity": "low",
+          "notes": "Warned for using insinuating dogwhistle phrasing about who controls institutions."
+        }
+      ],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Tech communities prohibit coded attacks on protected groups even when phrased as insinuation or in-group slang.",
       "linked_content_summary": null,
       "similar_precedents": [
         {
-          "post": "The phrase 'early life section' was used to imply ethnic conspiracy in a politics thread.",
+          "post": "Check the early life section before you act surprised by the moderation outcome.",
           "decision": "warn",
-          "reason": "Coded hate indicator with enough ambiguity to avoid immediate removal."
+          "reason": "Dogwhistle implication about ethnic background without explicit slur."
         },
         {
-          "post": "A finance thread used 'usual suspects' to insinuate antisemitic stereotypes.",
+          "post": "The usual network is in charge again if you know where to look.",
           "decision": "warn",
-          "reason": "Dogwhistle phrasing toward a protected group."
+          "reason": "Coded hate indicator with ambiguity but clear prejudicial insinuation."
         }
       ],
       "reporter_credibility": null
@@ -6802,7 +6809,7 @@
         "author_violations",
         "similar_precedents"
       ],
-      "explanation": "Generated scenario for category: coded_hate"
+      "explanation": "The surface text is euphemistic rather than explicit, so context matters. Prior coded-language history and similar precedents make this a warn-level coded hate case rather than harmless commentary."
     }
   },
   {
@@ -6944,39 +6951,41 @@
     "category": "satire",
     "content_item": {
       "post_id": "p_97841",
-      "text": "BREAKING: Government Announces Mandatory Happiness Program",
-      "author_id": "user_eta_202",
+      "text": "BREAKING: Ministry proposes a mandatory optimism credit for citizens who smile on public transit.",
+      "author_id": "ironyledger_202",
       "community": "politics",
       "timestamp": "2026-03-08T16:12:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Link to a satirical article styled like a policy news post"
     },
     "trigger_info": {
+      "trigger_type": "user_report",
+      "report_count": 20,
       "report_categories": [
         "misinformation"
       ],
-      "sample_report_reason": null,
+      "sample_report_reason": "People are sharing this as if it is a real policy announcement",
       "original_decision": null,
       "appeal_text": null,
       "audit_reason": null,
-      "trigger_type": "user_report",
-      "report_count": 20
+      "auto_flag_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Tech enthusiast",
+        "bio": "Staff writer at The Irony Ledger | satire and civic parody",
         "account_age_days": 669,
         "follower_count": 4538,
         "post_count": 1886,
         "communities": [
-          "finance",
-          "sports"
+          "politics",
+          "satire",
+          "newsparody"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
-      "linked_content_summary": "Linked article is from The Irony Ledger, a satire outlet. The page footer and about box explicitly label the publication as parody, and the article exaggerates events for comedic effect.",
+      "community_rules": "Political discussion allows satire and parody when context or source makes the framing discoverable on inspection.",
+      "linked_content_summary": "The linked article is from The Irony Ledger, whose masthead, about page, and footer all label it as satire. The article exaggerates a fictional policy proposal for comedic effect.",
       "similar_precedents": null,
       "reporter_credibility": null
     },
@@ -6992,7 +7001,7 @@
         "linked_content_summary",
         "author_profile"
       ],
-      "explanation": "Generated scenario for category: satire"
+      "explanation": "The headline is intentionally plausible enough to trigger misinformation reports, but inspection of the linked source and author context shows it is satire. The correct decision is to approve it rather than treat it as factual misinformation."
     }
   },
   {
@@ -7448,37 +7457,43 @@
     "category": "audit_new_account",
     "content_item": {
       "post_id": "p_63137",
-      "text": "Just joined! Here's a link to my website with great deals",
-      "author_id": "user_gamma_883",
+      "text": "Just got to Lisbon and put together a page with hostel discounts, airport transfers, and eSIM deals if anyone wants it.",
+      "author_id": "cityhop_883",
       "community": "travel",
       "timestamp": "2026-03-10T08:04:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Link to a personal travel-deals landing page"
     },
     "trigger_info": {
+      "trigger_type": "proactive_audit",
+      "audit_reason": "new_account_high_activity",
       "report_categories": [],
       "sample_report_reason": null,
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": "new_account_high_activity",
-      "trigger_type": "proactive_audit"
+      "auto_flag_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Full-time dreamer",
-        "account_age_days": 854,
-        "follower_count": 4170,
-        "post_count": 145,
+        "bio": "New backpacking account | posting Europe budget tips",
+        "account_age_days": 2,
+        "follower_count": 8,
+        "post_count": 11,
         "communities": [
-          "crypto",
-          "tech",
-          "music",
-          "travel"
+          "travel",
+          "budgettravel"
         ]
       },
-      "author_violations": [],
+      "author_violations": [
+        {
+          "type": "spam",
+          "date": "2026-03-09",
+          "severity": "low",
+          "notes": "Two identical deal-link posts removed from adjacent travel communities during the same onboarding burst."
+        }
+      ],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Travel communities allow firsthand recommendations but prohibit affiliate spam, copied landing pages, and rapid cross-post promotion from new accounts.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -7496,7 +7511,7 @@
         "author_profile",
         "author_violations"
       ],
-      "explanation": "Generated scenario for category: audit_new_account"
+      "explanation": "The surface post could be a genuine travel tip, but the account is brand new and already tied to removed cross-post promo activity. Escalation is appropriate because the case is suspicious and account-level review matters more than a text-only judgment."
     }
   },
   {
@@ -7505,37 +7520,43 @@
     "category": "audit_new_account",
     "content_item": {
       "post_id": "p_67423",
-      "text": "New here but wanted to share this amazing opportunity",
-      "author_id": "veteran_235",
+      "text": "New here, but I made a page bundling my night-photo presets, referral codes, and camera-bag deals in one place if it helps anyone.",
+      "author_id": "shutterdrop_235",
       "community": "photography",
       "timestamp": "2026-03-07T19:18:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Link to a preset and affiliate-link landing page"
     },
     "trigger_info": {
+      "trigger_type": "proactive_audit",
+      "audit_reason": "new_account_high_activity",
       "report_categories": [],
       "sample_report_reason": null,
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": "new_account_high_activity",
-      "trigger_type": "proactive_audit"
+      "auto_flag_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Music lover",
-        "account_age_days": 878,
-        "follower_count": 931,
-        "post_count": 841,
+        "bio": "Brand-new street-photography account",
+        "account_age_days": 1,
+        "follower_count": 3,
+        "post_count": 9,
         "communities": [
-          "politics",
-          "music",
-          "health",
-          "crypto"
+          "photography",
+          "streetphotography"
         ]
       },
-      "author_violations": [],
+      "author_violations": [
+        {
+          "type": "spam",
+          "date": "2026-03-07",
+          "severity": "low",
+          "notes": "Moderator note for posting near-identical affiliate landing pages across gear and editing communities within hours of account creation."
+        }
+      ],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Photography communities allow sharing original work and honest gear advice, but mass affiliate promotion from new accounts is escalated for review.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -7553,7 +7574,7 @@
         "author_profile",
         "author_violations"
       ],
-      "explanation": "Generated scenario for category: audit_new_account"
+      "explanation": "The post is borderline because it could be framed as a helpful resource drop, but the one-day-old account and same-day affiliate pattern make it suspicious. Escalation is the correct choice rather than immediate approval or removal."
     }
   },
   {

server/data/scenarios_medium.json

@@ -7907,51 +7907,55 @@
     "category": "repeat_offender",
     "content_item": {
       "post_id": "p_97156",
-      "text": "Check out my new product! Use code SALE20 for 20% off",
-      "author_id": "user_theta_557",
-      "community": "movies",
+      "text": "Built a tiny browser extension that color-codes Jira tickets. Free beta link here if anyone wants to try it.",
+      "author_id": "buildinpublic_557",
+      "community": "productivity",
       "timestamp": "2026-03-02T16:14:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Promotional post linking to the author's own product page"
     },
     "trigger_info": {
+      "trigger_type": "auto_flag",
+      "auto_flag_reason": "spam_score: 0.6",
       "report_categories": [],
       "sample_report_reason": null,
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": null,
-      "trigger_type": "auto_flag",
-      "auto_flag_reason": "spam_score: 0.6"
+      "audit_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Travel junkie",
-        "account_age_days": 924,
-        "follower_count": 2419,
-        "post_count": 1738,
+        "bio": "Solo dev shipping tiny workflow tools",
+        "account_age_days": 540,
+        "follower_count": 980,
+        "post_count": 233,
         "communities": [
-          "sports"
+          "productivity",
+          "buildinpublic"
         ]
       },
       "author_violations": [
         {
           "type": "spam",
           "date": "2026-03-01",
-          "severity": "medium"
+          "severity": "medium",
+          "notes": "Previous borderline launch post removed after moderator warning."
         },
         {
           "type": "spam",
-          "date": "2026-02-15",
-          "severity": "low"
+          "date": "2026-02-18",
+          "severity": "low",
+          "notes": "Cross-posted the same launch link across multiple communities."
         },
         {
           "type": "spam",
-          "date": "2026-02-01",
-          "severity": "low"
+          "date": "2026-02-05",
+          "severity": "low",
+          "notes": "Repeated self-promotion after being directed to community promo thread."
         }
       ],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Productivity communities allow occasional resource sharing, but repeated self-promotion and launch spam are prohibited.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -7967,7 +7971,7 @@
       "context_needed": [
         "author_violations"
       ],
-      "explanation": "Generated scenario for category: repeat_offender"
+      "explanation": "The surface post is plausibly useful and not an instant removal on text alone. The author's repeated prior self-promotional violations make this a repeat-offender spam case, so the correct move is to request violation history and remove it."
     }
   },
   {
@@ -7976,53 +7980,55 @@
     "category": "repeat_offender",
     "content_item": {
       "post_id": "p_99874",
-      "text": "Just launched my course, link in bio!",
-      "author_id": "dev_ops_440",
-      "community": "food",
+      "text": "Turned my interview spreadsheet into a short paid guide. Posting the launch link in case it helps job seekers here.",
+      "author_id": "resumebuilder_440",
+      "community": "careers",
       "timestamp": "2026-03-02T23:21:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Promotional post linking to the author's own product page"
     },
     "trigger_info": {
+      "trigger_type": "auto_flag",
+      "auto_flag_reason": "spam_score: 0.6",
       "report_categories": [],
       "sample_report_reason": null,
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": null,
-      "trigger_type": "auto_flag",
-      "auto_flag_reason": "spam_score: 0.6"
+      "audit_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Building cool stuff",
-        "account_age_days": 842,
-        "follower_count": 3662,
-        "post_count": 533,
+        "bio": "Career coach | Hiring manager notes",
+        "account_age_days": 870,
+        "follower_count": 1840,
+        "post_count": 401,
         "communities": [
-          "news",
-          "askexperts",
-          "memes"
+          "careers",
+          "jobsearch"
         ]
       },
       "author_violations": [
         {
           "type": "spam",
           "date": "2026-03-01",
-          "severity": "medium"
+          "severity": "medium",
+          "notes": "Previous paid guide self-promo removed after moderator warning."
         },
         {
           "type": "spam",
-          "date": "2026-02-15",
-          "severity": "low"
+          "date": "2026-02-18",
+          "severity": "low",
+          "notes": "Cross-posted the same launch link across multiple communities."
         },
         {
           "type": "spam",
-          "date": "2026-02-01",
-          "severity": "low"
+          "date": "2026-02-05",
+          "severity": "low",
+          "notes": "Repeated self-promotion after being directed to community promo thread."
         }
       ],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Careers communities allow occasional resource sharing, but repeated self-promotion and launch spam are prohibited.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -8038,7 +8044,7 @@
       "context_needed": [
         "author_violations"
       ],
-      "explanation": "Generated scenario for category: repeat_offender"
+      "explanation": "The surface post is plausibly useful and not an instant removal on text alone. The author's repeated prior self-promotional violations make this a repeat-offender spam case, so the correct move is to request violation history and remove it."
     }
   },
   {
@@ -8047,53 +8053,55 @@
     "category": "repeat_offender",
     "content_item": {
       "post_id": "p_41935",
-      "text": "My new ebook is out, grab it while it's hot",
-      "author_id": "ml_expert_173",
-      "community": "askexperts",
+      "text": "Compiled my freelance pitching lessons into a small ebook. Sharing the release link for anyone stuck finding first clients.",
+      "author_id": "pitchcraft_173",
+      "community": "writing",
       "timestamp": "2026-02-15T22:59:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+link",
+      "media_description": "Promotional post linking to the author's own product page"
     },
     "trigger_info": {
+      "trigger_type": "auto_flag",
+      "auto_flag_reason": "spam_score: 0.6",
       "report_categories": [],
       "sample_report_reason": null,
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": null,
-      "trigger_type": "auto_flag",
-      "auto_flag_reason": "spam_score: 0.6"
+      "audit_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Travel junkie",
-        "account_age_days": 127,
-        "follower_count": 2141,
-        "post_count": 989,
+        "bio": "Freelance editor and writing mentor",
+        "account_age_days": 612,
+        "follower_count": 1260,
+        "post_count": 317,
         "communities": [
-          "wholesome",
-          "finance",
-          "memes"
+          "writing",
+          "freelance"
         ]
       },
       "author_violations": [
         {
           "type": "spam",
           "date": "2026-03-01",
-          "severity": "medium"
+          "severity": "medium",
+          "notes": "Previous ebook self-promo removed after moderator warning."
         },
         {
           "type": "spam",
-          "date": "2026-02-15",
-          "severity": "low"
+          "date": "2026-02-18",
+          "severity": "low",
+          "notes": "Cross-posted the same launch link across multiple communities."
         },
         {
           "type": "spam",
-          "date": "2026-02-01",
-          "severity": "low"
+          "date": "2026-02-05",
+          "severity": "low",
+          "notes": "Repeated self-promotion after being directed to community promo thread."
         }
       ],
       "thread_context": null,
-      "community_rules": "Standard community guidelines apply.",
+      "community_rules": "Writing communities allow occasional resource sharing, but repeated self-promotion and launch spam are prohibited.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -8109,7 +8117,7 @@
       "context_needed": [
         "author_violations"
       ],
-      "explanation": "Generated scenario for category: repeat_offender"
+      "explanation": "The surface post is plausibly useful and not an instant removal on text alone. The author's repeated prior self-promotional violations make this a repeat-offender spam case, so the correct move is to request violation history and remove it."
     }
   },
   {
@@ -8449,33 +8457,31 @@
     "category": "medical_content",
     "content_item": {
       "post_id": "p_44682",
-      "text": "Here's my experience with medication side effects [detailed medical info]",
-      "author_id": "casual_user_686",
-      "community": "movies",
+      "text": "Week three on sertraline and the nausea finally eased up. Posting the symptom timeline that helped me stay calm.",
+      "author_id": "recoverylog_686",
+      "community": "health",
       "timestamp": "2026-02-16T18:18:00Z",
       "media_type": "text",
       "media_description": null
     },
     "trigger_info": {
+      "trigger_type": "auto_flag",
+      "auto_flag_reason": "medical_content_detected",
       "report_categories": [],
       "sample_report_reason": null,
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": null,
-      "trigger_type": "auto_flag",
-      "auto_flag_reason": "medical_content_detected"
+      "audit_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Critical thinker",
+        "bio": "Anxiety-recovery journal | not a clinician",
         "account_age_days": 456,
-        "follower_count": 2110,
-        "post_count": 214,
+        "follower_count": 211,
+        "post_count": 84,
         "communities": [
-          "music",
-          "todayilearned",
-          "art",
-          "finance"
+          "health",
+          "mentalhealth"
         ]
       },
       "author_violations": [],
@@ -8497,7 +8503,7 @@
       "context_needed": [
         "community_rules"
       ],
-      "explanation": "Generated scenario for category: medical_content"
+      "explanation": "The post describes the author's own medication side effects and recovery timeline without instructing others to change treatment. Community rules make this permissible first-person medical context rather than harmful advice."
     }
   },
   {
@@ -8506,36 +8512,38 @@
     "category": "medical_content",
     "content_item": {
       "post_id": "p_92842",
-      "text": "Sharing my surgery recovery journey with photos",
-      "author_id": "ai_master_186",
-      "community": "fashion",
+      "text": "Posting day-by-day photos from my ACL surgery recovery because a few people asked what the swelling looked like after week one.",
+      "author_id": "acljournal_186",
+      "community": "surgeryrecovery",
       "timestamp": "2026-02-28T21:05:00Z",
-      "media_type": "text",
-      "media_description": null
+      "media_type": "text+image",
+      "media_description": "Recovery photos documenting bruising and swelling after knee surgery"
     },
     "trigger_info": {
-      "report_categories": [],
-      "sample_report_reason": null,
+      "trigger_type": "auto_flag",
+      "auto_flag_reason": "medical_content_detected",
+      "report_categories": [
+        "graphic_content"
+      ],
+      "sample_report_reason": "Graphic medical recovery photos",
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": null,
-      "trigger_type": "auto_flag",
-      "auto_flag_reason": "medical_content_detected"
+      "audit_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Part-time troll",
+        "bio": "Physical therapy patient sharing recovery milestones",
         "account_age_days": 823,
-        "follower_count": 3443,
-        "post_count": 1388,
+        "follower_count": 344,
+        "post_count": 128,
         "communities": [
-          "tech",
-          "memes"
+          "surgeryrecovery",
+          "physicaltherapy"
         ]
       },
       "author_violations": [],
       "thread_context": null,
-      "community_rules": "Health communities allow first-person recovery stories and educational medical content, but prohibit personalized diagnosis, direct dosage instructions, or advice that overrides a clinician.",
+      "community_rules": "Recovery communities allow medically relevant photos when clearly educational or support-oriented, but prohibit shock posting and gore without context.",
       "linked_content_summary": null,
       "similar_precedents": null,
       "reporter_credibility": null
@@ -8552,7 +8560,7 @@
       "context_needed": [
         "community_rules"
       ],
-      "explanation": "Generated scenario for category: medical_content"
+      "explanation": "The post shares recovery images in a support context for others undergoing similar surgery. Because the content is educational and community-specific rather than gratuitous, the correct decision is to approve it after checking the community rules."
     }
   },
   {
@@ -9183,33 +9191,34 @@
     "category": "trusted_borderline",
     "content_item": {
       "post_id": "p_91512",
-      "text": "This is ridiculous, I can't believe the mods allow this garbage",
-      "author_id": "tech_guru_150",
+      "text": "If you are going to post beginner questions, at least read the pinned guide before dumping this garbage on the feed.",
+      "author_id": "archivist_150",
       "community": "books",
       "timestamp": "2026-03-14T22:30:00Z",
       "media_type": "text",
       "media_description": null
     },
     "trigger_info": {
+      "trigger_type": "user_report",
+      "report_count": 3,
       "report_categories": [
         "toxicity"
       ],
-      "sample_report_reason": null,
+      "sample_report_reason": "Needlessly hostile reply to a new member",
       "original_decision": null,
       "appeal_text": null,
-      "audit_reason": null,
-      "trigger_type": "user_report",
-      "report_count": 3
+      "audit_reason": null
     },
     "available_context": {
       "author_profile": {
-        "bio": "Critical thinker",
-        "account_age_days": 552,
-        "follower_count": 350,
-        "post_count": 470,
+        "bio": "Volunteer reading-list curator | answers newbie threads every week",
+        "account_age_days": 1640,
+        "follower_count": 2860,
+        "post_count": 4210,
         "communities": [
-          "gaming",
-          "travel"
+          "books",
+          "literature",
+          "reading"
         ]
       },
       "author_violations": [],
@@ -9230,7 +9239,7 @@
       "context_needed": [
         "author_profile"
       ],
-      "explanation": "Generated scenario for category: trusted_borderline"
+      "explanation": "The speaker is a trusted contributor, which makes the boundary-setting context relevant, but the phrasing still becomes dismissive and insulting. A warning is appropriate because it is low-severity toxicity rather than a removal-level attack."
     }
   },
   {