Demo
A fictional LinkedIn post + the verdict the Council produced.
The examples/demo/ directory ships a real-shaped demo of what the Council does. Read this to calibrate expectations before your first review.
What’s in the demo
| File | Purpose |
|---|---|
sample_linkedin_post.md | 280-word LinkedIn post about a fictional framework (Lattice). 4 seeded issues. |
sample_verdict.json | Full structured verdict the Council produced. REVISE. |
sample_revision_brief.md | Human-readable revision brief extracted from the verdict. |
The seeded issues
- V1 voice violation — “Not a product but a protocol” (split-form negation-pivot)
- V3 voice violation — “recontextualizes how a PM team works” (hype word)
- T5 calibration — “cycle time dropped by half” (unsourced operator claim)
- Thin counter-evidence — self-serve counter named but response is rhetorical, not evidential
The Council catches all four. Skeptic + Voice + Evidence each block; Strategy passes outright. The Adjudicator merges → REVISE with a 6-step revision brief.
Reading the verdict
A REVISE verdict means:
- 1–2 (or 3+) deliberators flagged blocking concerns BUT the issues are reducible (line-level edits, not redrafts).
- The Adjudicator agreed the artifact can be saved with the revision brief.
- The producing agent should apply the brief and re-submit.
A HOLD verdict means:
- 3+ deliberators blocked OR any irreducible flag fired.
- The artifact needs a redraft, not an edit pass.
- Do not iterate — surface the blockers to the operator.
A SHIP verdict with concerns means:
- Zero blocking flags.
- Some deliberator raised non-blocking concerns worth fixing in a later pass.
- The artifact can ship; the concerns are tracked in the audit log for compounding-loop learning.
Reproduce
# From the repo root
cp council.yaml.example council.yaml
# Edit council.yaml — point context_refs at examples/ stubs, set runtime to claude_cli
python -m agent_council review examples/demo/sample_linkedin_post.md --tier=1
Real verdicts will vary across runs (LLM-as-judge non-determinism — documented in Honest limitations). The sample verdict in the demo is one representative run, not a fixture the test suite asserts against.