docs: simplify the recovery postmortem guidance

Author: sushildalavi

docs/RECOVERY_POSTMORTEM.md

@@ -1,66 +1,24 @@
-# Recovery Postmortem Generator
+# Recovery Postmortem
 
-Converge includes an optional recovery postmortem workflow that turns benchmark, chaos, and convergence evidence into a structured incident-style summary.
+The postmortem generator is evidence-driven. It uses benchmark artifacts, chaos artifacts, and live snapshots to generate a bounded summary instead of pretending the model has extra context.
 
-## Architecture
+## Inputs
 
-- Evidence collection gathers benchmark artifacts, chaos artifacts, workflow snapshots, convergence snapshots, retry counts, DLQ counts, worker health, and recovery timing.
-- LangGraph orchestrates the workflow:
-  - collect recovery evidence
-  - summarize the recovery timeline
-  - assess convergence
-  - identify operational risks
-  - generate the postmortem
-  - validate the schema
-  - verify evidence grounding
-- LangChain structured-output parsing keeps the response aligned with the Pydantic schema.
-- Ollama is supported as the optional local model backend.
-- A fake provider powers tests, CI, and offline validation.
+- benchmark JSON/Markdown under `benchmarks/`
+- chaos JSON/Markdown under `benchmarks/`
+- optional workflow/convergence snapshots from the API
 
-## Runtime Defaults
-
-- `AI_PROVIDER=disabled`
-- `OLLAMA_BASE_URL=http://localhost:11434`
-- `AI_MODEL=llama3.1:8b`
-- `AI_FALLBACK_MODEL=qwen2.5-coder:7b`
-- `AI_TIMEOUT_SECONDS=20`
-
-The service does not require Ollama at startup. If Ollama is unavailable, the generator falls back to deterministic evidence-based output.
-
-## How To Generate
+## Usage
 
 ```bash
 python scripts/generate_postmortem.py --artifact benchmarks/benchmark_replay_*.json --workflow-id <workflow_id>
-```
-
-You can also point the CLI at a running API:
-
-```bash
-python scripts/generate_postmortem.py --artifact benchmarks/benchmark_replay_*.json --base-url http://127.0.0.1:18000 --workflow-id <workflow_id>
-```
-
-## Evaluation
-
-```bash
 python scripts/evaluate_postmortem.py
 ```
 
-The evaluation harness checks:
-
-- schema-valid output rate
-- recovery classification accuracy
-- evidence coverage
-- average latency
-- insufficient-evidence handling
-
-## Evidence Rules
-
-- The generator must cite actual evidence collected from artifacts or live recovery state.
-- If the evidence is too thin, it returns `insufficient_evidence`.
-- The output should not claim 100K-event chaos runs or 3,000+ msg/sec throughput unless those numbers are actually measured in this repository.
+## Guardrails
 
-## CI Behavior
+- Returns `insufficient_evidence` when the source material is too thin
+- Uses local deterministic mode by default
+- Optional external providers are only used when configured
+- Does not replace the recovery engine or the operator console
 
-- CI and tests use the fake provider.
-- No paid API keys are required.
-- Ollama is optional and not part of the default startup path.