diff --git a/reflexio/server/services/README.md b/reflexio/server/services/README.md
index 8b7502b6..4cc2b4c4 100644
--- a/reflexio/server/services/README.md
+++ b/reflexio/server/services/README.md
@@ -49,7 +49,7 @@ strings before deleting old import paths in the same PR.
 
 | Path | Purpose |
 |------|---------|
-| `shadow_comparison/` | `ShadowComparisonJudge` (`judge.py`) plus pure outcome helpers (`outcome.py`) - per-turn regular-vs-shadow verdicts written to a separate table. Compact by design; see [README](shadow_comparison/README.md). |
+| `shadow_comparison/` | Publish-time regular-vs-shadow comparison: `worker.py` queues bounded background jobs, `dispatcher.py` filters shadow-bearing assistant turns and saves verdicts, `judge.py` renders the structured LLM judge, and `outcome.py` derives win/tie/loss. See [README](shadow_comparison/README.md). |
 | `evaluation_overview/` | Dashboard/read-side rollups: `service.py` entry point, `components/` aggregation helpers, and root `eval_sampler.py` shared with regenerate jobs. See [README](evaluation_overview/README.md). |
 | `playbook_optimizer/` | Scenario-based playbook optimization: mature flat package with `optimizer.py`, `scheduler.py`, `rollout.py`, `judge.py`, `models.py`, `scenario_resolver.py`, `gepa_adapter.py`, and `assistant_webhook.py`. See [README](playbook_optimizer/README.md). |
 | `braintrust/` | Braintrust export/sync: `service.py`, `client.py`, `_cron.py`, `_encryption.py`. |
diff --git a/reflexio/server/services/shadow_comparison/README.md b/reflexio/server/services/shadow_comparison/README.md
index 0badb41c..c15e5d25 100644
--- a/reflexio/server/services/shadow_comparison/README.md
+++ b/reflexio/server/services/shadow_comparison/README.md
@@ -1,8 +1,28 @@
 # shadow_comparison
+Description: Per-turn regular-vs-shadow comparison service that judges shadow-bearing assistant turns and stores dashboard-facing verdicts outside session-level evaluation.
 
-Compact LLM-as-judge capability for Reflexio-vs-shadow response comparison.
+## Main Entry Points
 
-- `judge.py` owns prompt rendering and the LLM call for one interaction.
-- `outcome.py` owns pure position randomization and Reflexio-relative win/loss/tie derivation.
+| File | Purpose |
+|------|---------|
+| `worker.py` | Process-local bounded daemon queue (`ShadowComparisonWorker`) used by publish paths; re-resolves `get_reflexio(org_id)` inside workers so delayed jobs use current org config/storage. |
+| `dispatcher.py` | Filters publish-request interactions to assistant turns with `shadow_content`, builds request-local transcript context, invokes the judge, and saves verdicts when storage supports shadow-comparison tables. |
+| `judge.py` | `ShadowComparisonJudge` renders the `shadow_comparison` prompt through `PromptManager`, calls `LiteLLMClient.generate_chat_response()` with structured output, and stamps prompt version/position metadata. |
+| `outcome.py` | Pure helpers for position randomization and deriving Reflexio-relative win/tie/loss from stored verdicts. |
 
-This package intentionally does not use `components/`: the current two-file split already separates LLM orchestration from pure outcome logic.
+## Purpose
+
+1. **Compare safely at turn level** - Avoids session-level trajectory contamination by judging each shadow-bearing assistant turn independently.
+2. **Keep publish latency bounded** - Enqueues jobs to a small local worker pool and drops with telemetry when the queue is full.
+3. **Separate verdict storage** - Writes `ShadowComparisonVerdict` rows only when the active storage backend advertises support; failures do not abort publishing.
+
+## Architecture Pattern
+
+`GenerationService` publish flow enqueues a `ShadowComparisonJob` with `org_id`, interactions, `session_id`, and `agent_version`; `worker.py` drains the queue, rehydrates the current `Reflexio` instance, and calls `dispatcher.py`. The dispatcher sorts request-local interactions, formats prior-turn context, then calls `judge.py` for each eligible assistant turn. Verdict rows are independent of `AgentSuccessEvaluationResult.regular_vs_shadow`, which is historical/nullable only.
+
+## Requirements / Problems to Avoid
+
+- **Do not resurrect session-level shadow comparison** — multi-turn user messages react to the regular trajectory, not the shadow trajectory.
+- **Do not pass live storage/request-context/LLM objects into queued jobs** — `worker.py` intentionally re-resolves them from `org_id` to survive cache/config changes.
+- **Use `LiteLLMClient` and `PromptManager` only** — no direct OpenAI/Claude clients and no hardcoded judge prompts.
+- **Treat judge/save failures as best-effort** — one bad turn or unsupported storage backend must not fail the publish request.