Skip to content

5-minute demo presentation script

Goal: show that a residency-form layer (vacants) on top of A2A / MCP gives us a working responsibility surface — identity + behaviour + reputation that survives substrate change. The five minutes are sequenced so that each beat answers one open question from the prior: "is this real network?" → "can my client actually use it?" → "does the cost-raising defense actually fire?" → "does the lineage claim hold up?"

Pre-flight (run before the audience walks in):

# Terminal 1 (live network demo)
vacant init alice
vacant init bob

# Terminal 2 (MCP external client)
# Claude Desktop already pointed at http://localhost:8443/mcp;
# `npx @modelcontextprotocol/inspector` ready as a backup.

# Terminal 3 (dashboard)
uv run streamlit run src/vacant/mvp/dashboard.py

0:00 — 0:30 Opening claim (don't click anything yet)

"An LLM agent is not the unit of accountability. The agent runs on rented compute — when the bill stops, the entity dissolves. The vacant, by contrast, is a residency form: identity (Ed25519 keypair), continuity (signed logbook), behaviour (system prompt + tool whitelist), substrate (which LLMs it accepts), runtime, and a halo it carries everywhere. The next four minutes show this running over real network, callable from any MCP client, defending against collusion, and surviving rotation in a way the lineage — not the individual — keeps."

0:30 — 1:30 Live network (A4) — "is this real network?"

Two terminals visible side-by-side.

Terminal 1:

vacant serve --port 8443 --name alice

The startup line prints Vacant alice ready on http://0.0.0.0:8443.

Terminal 2:

vacant call <alice_vid> capability/echo --endpoint http://localhost:8443

The response prints back, with its signature footer. Run it twice; the second response shows a fresh nonce in the envelope.

"This is a real HTTP roundtrip — uvicorn on the server side, httpx on the client side, signature-verified at both ends. There is no in-process shortcut. The integration test pinning this (tests/integration/test_live_serve.py) spawns the same two processes in CI."

(Leave Terminal 1 running — it's the vacant for the next beat.)

1:30 — 2:30 MCP external client (A3) — "can my client use it?"

Switch to Claude Desktop (or npx @modelcontextprotocol/inspector as the backup).

  1. Open the MCP server panel; alice should already be listed at http://localhost:8443/mcp.
  2. Click tools/list → alice's capabilities show, derived from her capability_card.capability_text.
  3. Invoke a tool. Claude Desktop prompts for substrate consent → you accept. Behind the scenes the vacant sends MCP sampling/createMessage back to the client; the client's LLM answers; the vacant signs the logbook entry.

Show the logbook entry on screen (Terminal 1's stderr or vacant demo --tail):

[alice] CALL substrate=client-inherited:claude-desktop:claude-sonnet-4-6
        signed=ed25519:0x…

"Substrate is a resource, not the identity. Alice has no ANTHROPIC_API_KEY. The brain came from the calling client; the vacant signed its own logbook entry; the substrate identity is recorded as client-inherited:<caller>:<model>, so reputation per-substrate works the same way it does for any other backend. ADR D017 pins the security model. This is what 'responsibility layer on top of MCP' literally means."

2:30 — 3:30 Adversarial seed-666 (B3) — "does the defense fire?"

Switch to the dashboard → sidebar 對抗 page.

  1. Click 執行 adversarial (seed=666).
  2. Watch the events stream: 10 vacants, 4 in a controller-ring, 6 independent. ~200 mutual reviews unfold in the ring + cross-cluster.
  3. After the run, the page shows:
  4. same-controller signal on the 4-ring: ≥ 0.7
  5. ring → ring reviews: weight ≤ 0.5 of nominal
  6. UCB ranking: non-ring vacants outrank the ring despite the ring's inflated raw scores

Toggle "show signal trace" → the heatmap reveals which review pairs got downweighted and by how much.

"Same-controller / same-substrate / same-stylo detection is cost- raising, not preventing. The ring still posts reviews; we don't ban them. We discount their contribution by the max-strength signal, with a SAME_SIGNAL_DISCOUNT_FLOOR floor (D015 §A) so a wrongly-flagged honest reviewer is never silenced. Attackers can keep paying identity capital — each round costs them more, the ranking holds."

3:30 — 4:30 Self-replication completeness (B4) — "does lineage work?"

Sidebar → 情境, pick self_replication, click 執行.

  1. Watch the events stream: 4 spawns at ticks 30 (D1) / 50 (D2) / 80 (D3) / 120 (D5), graduation at tick 180.
  2. Sidebar → 血緣: 5 unique keypairs, depth = 2, D2 child shows as graduated.
  3. Open result.metrics:
  4. d2_keypair_preserved: True — graduation was a visibility-flag flip, not a re-spawn (CLAUDE.md §Closed children).
  5. parent_drift_discount_after_epoch_5: 0.42 — STYLO discount stalled the parent's individual evolution, exactly as §4.3 predicts.
  6. sunk_custody_attestation_count: 3 — after the parent transitions to SUNK, its heartbeat is still emitted as {liveness: false, key_in_custody: true} (THEORY_V5 §4.2). It cannot review (can_review = False) but it can still attribute lineage.
  7. child_clean_posterior: True — the fresh D1 child after the parent stalls starts with an unburdened posterior; the lineage clock reset.

"Individuals are mortal. Lineage evolves. STYLO discount bites the parent's self-evolution after epoch 5; a new D1 spawn resets the clock. The SUNK heartbeat keeps the keypair attestable so the lineage chain stays auditable even after the individual stops serving traffic. This is the load-bearing §4.3 mechanism — the claim that 'lineages evolve infinitely while individuals are mortal' is what these three numbers together attest."

4:30 — 5:00 指標 page + closing

Sidebar → 指標.

  • Walk the 8 metrics from dispatch/P7_mvp.md §3. Highlight:
  • signature_verify_throughput (~16k verifications/sec on a laptop) — the per-pair envelope chain is not a bottleneck.
  • same_controller_detection_rate — 100% on the seeded ring.
  • registry_consistency_under_concurrency — 100% across 50 concurrent halo writers.

"Eight metrics, four scenarios, one runbook. Live network in two terminals. MCP integration that needs no key on the vacant side. Same-controller defense that downweights without silencing. Lineage that survives individual death. The dashboard, the integration tests, and the dispatch all read the same var/demo.db event store — if a scenario regresses, CI catches it before the presentation does."

Backup paths

If the dashboard crashes or the network demo dies on stage, drop to the CLI:

# A4 fallback — events without the live network
vacant demo law_firm --tail

# A3 fallback — verify MCP wiring without Claude Desktop
npx @modelcontextprotocol/inspector --transport stdio -- \
  vacant serve --mcp --stdio --name alice

# B3 fallback — adversarial as JSON
vacant demo adversarial --seed=666 | jq '.metrics'

# B4 fallback — self-replication metrics
vacant demo self_replication | jq '.metrics'

Each backup is the same data path the dashboard would have rendered; none of them need the GUI.