Demos

Everything below runs offline from a checkout of github.com/jkomkov/bulla — no network, no LLM, no randomness.

The recourse gate — PROCEED / REFUSE with a cure

Signs two real shipped compositions, logs them, and gates them: the coherent one passes, the one with a real filesystem↔git seam is refused — and the refusal names its cure.

$ bulla/examples/gate-quickstart.sh
▸ gate coherent deed  → PROCEED (exit 0)
▸ gate seam deed      → REFUSE  (exit 1) + cure: disclose path_separator

The lying host — the loop closes

A host serves an equivocated registry root; the gate catches it because inclusion is only trusted against an independently pinned root. The same run then shows the seam breaking a real git execution and the same convention disclosure that clears the fee fixing it — the demo prints LOOP CLOSED.

$ python bulla/calibration/recourse_gate_closes_loop_git.py

The live proxy — an agent-side copilot

Fronts two backends as one MCP server and injects the bulla__* meta-tools; a canned JSON-RPC session shows the agent asking bulla__should_proceed before a risky call and getting refuse with the obstruction named. Per-call telemetry streams to events.jsonl.

$ bulla/examples/live-mcp-proxy/run_demo.sh

The diagnostics — exact or lower bound, never a guess

Bulla finds structural problems that schema validation misses entirely. And it tells you whether its repair is exact or only a lower bound.

$ cd bulla/examples/epistemic-demo
$ python run_demo.py

A. Exact regime

Two servers (analytics, storage) share path and timestamp fields. Both schemas say string. Both pass validation. But one means an absolute filesystem path; the other means a relative key. One means ISO-8601; the other means Unix epoch.

Bulla detects both hidden dimensions, recommends a minimum-cost repair, and proves it is optimal:

{
  "fee": 2,
  "geometry_dividend": 10.0,
  "sigma_star": 10.0,
  "regime": "exact",
  "recommended_repair": [
    ["analytics__query_events", "timestamp"],
    ["analytics__query_events", "path"]
  ]
}

regime=exact means: the geometry dividend is the true unavoidable cost. No cheaper repair exists. Bulla can prove it.

B. Surrogate regime

A synthetic three-server chain where token is a coloop — it must be disclosed in every repair, and its cost is unavoidable. Bulla honestly reports that its formula is an approximation:

{
  "fee": 3,
  "geometry_dividend": 6.0,
  "sigma_star": 16.0,
  "regime": "surrogate",
  "forced_cost": 10.0,
  "downgrade": "coloop_burden"
}

regime=surrogate means: the geometry dividend is useful but not a provable bound. The downgrade field says why.

Coloops are rare in the calibration corpus but arise in compositions with single-path credential or authorization dependencies.

C. Comparison

The same analytics → storage composition through two lenses:

A recommendation is only as good as its provenance. Bulla returns the epistemic status of the recommendation.

The trust ladder

D. BABEL composition challenge (canonical benchmark demo)

For multi-tool cyclic compositions, the reference surface is BABEL — frozen instances with deterministic holonomy ground truth.

$ cd benchmark/coherence-gym
$ pip install -e .
$ python -m coherence_gym demo
$ python -m coherence_gym show invoice/medium/dev_000

Score your own predictions with score-predictions. Narrative results: BABEL page.

Source

The full demo script, manifests, and frozen output are in bulla/examples/epistemic-demo.