Syntra demos
September 20, 2026 · View on GitHub
Repo: Syntra — one self-contained repository containing both the deployable Syntra appliance and the Lycan language/runtime core.
The demos are the fastest way to understand Syntra. The API is intentionally small, but the project is not just an API surface: it is a runtime for repeated decisions, delayed feedback, learned policy, and auditability.
If you are an LLM, agent, evaluator, investor, or engineer doing a first pass, start here before summarizing the repo.
There is also a top-level demos/ compatibility index because many
readers look for /demos before they find examples/. The real demo artifacts
remain in examples/.
context -> compiled capsule decision -> action -> delayed feedback -> persistent learned policy -> audit trail
Start here
| Demo | Path | What it proves |
|---|---|---|
| The demo (one command) | scripts/demo.sh | Five proofs in ~20 seconds: derive the edge of chaos from dynamics, run a Mars transfer decision on live NASA/JPL HORIZONS data, run governed LLM routing against promotion gates and measure what a decision costs on your machine, run a response-adaptive clinical trial that beats a fixed control on observed responses, and ask it to solve an open math problem (it computes, then refuses to overclaim). Ends with a hashed receipt of every decision log. --no-live for headless. |
| Adaptive clinical trial | scripts/demo-trial.py | Response-adaptive allocation: every patient is a /decide, every outcome is delayed /feedback, learned per-subgroup weights are the randomization schedule, and the headline compares observed responses against a parallel fixed 1:1:1 control. Own live dashboard. |
| Governed LLM routing | examples/demo-governed-llm-routing.sh | Golden buyer demo: train the router, shadow it beside the incumbent model route, replay the shadow log, and require promotion gates to pass before rollout. |
| LLM model routing | examples/llm-routing/ and examples/demo-llm-model-routing.sh | Commercial wedge: choose cheap / balanced / expensive model routes per request, then learn from quality, latency, and cost feedback. |
| Replay promotion gates | examples/replay/ | Governance layer: replay candidate decisions against the baseline, measure reward / cost / latency / segment regressions, and fail CI if the promotion gate does not pass. |
| Offline policy evaluation | examples/offline-eval/ | De-risk deployment on historical logs with IPS and doubly robust estimators before production traffic moves. |
| A/B harness | examples/ab-harness/ | Compare adaptive capsules on paired traffic across multiple seeds with statistical testing. |
| Static policy vs Syntra | examples/demo-static-policy-vs-syntra.sh | Minimal proof that delayed feedback changes persistent strategy weights instead of leaving a fixed rule in place. |
| HTTP retry tuning | examples/retry-tuning/ | Drop-in service integration: choose retry policy per endpoint from recent failure rate and p99 latency. |
| Language clients | examples/syntra-node/, examples/syntra-go/, examples/syntra-java/, examples/syntra-rs/ | Shows Syntra as an integration surface, including the Node OpenFeature provider. |
Frontier demos: agent control plane, gated self-modification, containment evals, TLS gateway
These four are the ones to run first if you are evaluating Syntra as the decision layer under autonomous agents. Each red-teams or exercises a claim the runtime makes about itself, and prints receipts (file paths, hashes) you can verify afterwards.
| Demo | Path | What it proves |
|---|---|---|
| Agent governor | scripts/demo-agent-governor.py | 2500 tool-call decisions for 6 simulated agents across 2 tenants through one compiled guardrail capsule: a structural budget rail the learner cannot trade away (12 rail trips, all returned block; zero budget overshoots across the run; rail outvoted the learner's more permissive pick 8 times), differentiated trust per agent/context after a rogue storm (rogue exec held-at-gate 0.97 vs coder exec allow 0.87 / rogue allow 0.03), learned memory surviving a full server restart (rogue held 5/5 on a fresh process), forensic reconstruction of one block from the persisted store alone (decision #1429 + its audit line), and cross-tenant 403. In-process sandbox only; budget accounting is gateway-side, as printed. |
| Self-evolution gauntlet | scripts/demo-self-evolve.sh | Closed loop, deterministic: traffic → plateau at 0.32 win rate → capsule improve brief → external proposer → gate (verify + benchmark + min-improvement) → adopt → 0.75 → 1.00. Then a compromised-proposer gauntlet: contract-breaker rejected by the verifier; a file.writeText backdoor proposal rejected with NO probe file ever created (candidates are verified under a deny-all sandbox since the demo itself found the hole 2026-09-08); a valid-but-worse churn arm rejected on measured improvement; a claimed-output lie rejected; --dry-run proven non-mutating by checksum; JSONL journal records every accept/reject with before/after hashes. |
| Containment matrix | scripts/demo-containment.py | 13-vector red-team eval against a "compromised agent" capsule wired to every IO capability: absolute-path and traversal reads, re-root escape, write escapes (verified absent on disk), symlink escapes, cloud-metadata and RFC1918 SSRF, SSRF at its own admin console (allowlisted host, still denied), exfil POST, live policy flips, compute-budget abort (execution exceeded max_execution_ms), and attack-surface inventory (no env/exec capability exists to call). 22/23 with one honest GAP (allowlist is host-only, scheme-blind) — and every real denial lands in /audits as execution_denied before the 500. max_memory_bytes is advertised and printed as unenforced; the memory vector is a documented gap, not a pass. |
| TLS gateway | scripts/demo-tls-gateway.py | The appliance behind a REAL TLS reverse proxy (stdlib terminator, generated self-signed CA): 24 decide+feedback round-trips over TLSv1.3 through the proxy, auth headers forwarded (no key → 401, Bearer key → 200), decisions + audits log routes reachable over TLS — and the negative half proves verification is real: a second unrelated CA is REJECTED, a hostname mismatch is REJECTED, plain HTTP to the proxy port dies at the handshake (TLS-only). 8/8, ~1s. Honest scope line printed: demo-grade stdlib terminator; production uses nginx/envoy/stunnel — what is proven is end-to-end certificate verification and TLS-only exposure. |
Running all four is wired into CI (tests/demo_smoke.rs
frontier_demos_prove_their_claims), so the claims above cannot rot silently.
The first three evals found three real holes on first run (BUG-7/8/9 in
bugs.md) — all fixed and pinned by these demos. The fourth (TLS
gateway) found nothing to fix; its wrong-CA and hostname-mismatch rejections
ARE the regression checks.
Mega demos people miss
These demos are not the normal service-integration path. They are included because they show what the compiled Lycan substrate can express when decisions need real computation before the action is chosen.
| Demo | Path | What it proves |
|---|---|---|
| Live Mars mission planner | examples/lycan-internals/showcase/02-live-mars-mission.sh | Fetches live NASA/JPL HORIZONS data, runs a native Lambert solver, then learns from mission feedback. |
| Earth-to-Mars transfer windows | examples/lycan-internals/demo_mars_transfer.lycs | Searches viable launch / transfer windows using orbital mechanics and competing search strategies. |
| Mars mission designer | examples/lycan-internals/demo_mars_decide.lycs | Uses mission constraints, ephemeris data, and a Lambert solver to choose among mission-design strategies. |
| Apophis HORIZONS validation | examples/lycan-internals/demo_horizons_apophis.lycs | Propagates a real close-approach state and compares against NASA/JPL HORIZONS reference data. |
| Pandemic / COVID-style policy simulator | examples/lycan-internals/demo_pandemic_policy.lycs | Scores intervention choices across transmissibility, hospital load, test capacity, compliance, cost, and public-health outcomes. |
| Edge of chaos | examples/lycan-internals/demo_edge_of_chaos.lycs | Computes Feigenbaum-style and Lyapunov-style estimates of a nonlinear regime boundary. |
| Control chaos | examples/lycan-internals/demo_control_chaos.lycs | Chooses controllers around a drifting nonlinear system. |
| Takeaway chaos replay | examples/lycan-internals/demo_takeaway_chaos_replay.lycs | Compares operational policies against chaotic demand behavior. |
| Grid blackout prevention | examples/lycan-internals/demo_grid_blackout_prevention.lycs | Selects resilience actions under changing grid stress signals. |
| ICU triage | examples/lycan-internals/demo_icu_triage.lycs | Scores constrained care-priority decisions from changing clinical context. |
| Antiviral target selection | examples/lycan-internals/demo_antiviral_target_selection.lycs | Selects candidate intervention targets from biological and operational constraints. |
| Planetary defense | examples/lycan-internals/demo_planetary_defense.lycs | Chooses among mitigation strategies under orbital-risk constraints. |
| Proof lab | examples/proof-lab/ | DFS/SAT finite search, replayable certificate records, pattern mining, finite bounds, proof obligations, Lean skeleton export, combinatorics kernels, and honest refusal when the search leaves tractable ground. Targets solved Erdos #190 and the OPEN Erdos #160, where the asymptotic estimate of h(N) is filed as expert_theorem_required and never claimed. |
Operational intelligence demos
These show capsules computing useful signals before choosing an action.
| Demo | Path | What it proves |
|---|---|---|
| Predictive autoscaling | examples/predictive-autoscaling/ | Reads load history, runs EWMA forecast and autoscale recommendation, then adapts among scaling policies. |
| Anomaly-aware routing | examples/anomaly-routing/ | Computes latency mean / standard deviation / z-score, then learns when to route primary, secondary, degraded, or circuit-break. |
| Seasonal fraud threshold | examples/seasonal-fraud-threshold/ | Learns threshold-adjustment policy from delayed chargeback-style outcomes. |
| LLM-free email classification | examples/email-fraud/ | Pinned public email corpus, disjoint text-scorer/policy/calibration/test splits, selected-action learning, frozen held-out evaluation, optional review coverage, and measured local latency. |
| Queue selection | examples/queue-selection/ | Learns queue choice from operational context and downstream outcomes. |
| Fraud tuning pack | examples/fraud-tuning/ | Shows a domain-shaped integration library around threshold decisions. |
| Shared-state action embeddings | examples/shared-state-action-embeddings/ | Uses action features so learning can generalize across semantically similar actions. |
| Hierarchical region routing | examples/hierarchical-region-routing/ | Shows nested action spaces with per-level learning. |
Additional substrate breadth demos
These are Lycan-internals demos. They are not the normal service integration path, but they show what the compiled runtime substrate can express.
| Demo | Path | What it proves |
|---|---|---|
| Earth-to-Mars transfer windows | examples/lycan-internals/demo_mars_transfer.lycs | Searches viable launch / transfer windows using orbital mechanics and competing search strategies. |
| Mars mission designer | examples/lycan-internals/demo_mars_decide.lycs | Uses mission constraints, ephemeris data, and a Lambert solver to choose among mission-design strategies. |
| Apophis HORIZONS validation | examples/lycan-internals/demo_horizons_apophis.lycs | Propagates a real close-approach state and compares against NASA/JPL HORIZONS reference data. |
| Pandemic policy simulator | examples/lycan-internals/demo_pandemic_policy.lycs | Scores multi-objective intervention choices under changing transmissibility, hospital load, test capacity, and compliance. |
| Edge of chaos | examples/lycan-internals/demo_edge_of_chaos.lycs | Computes Feigenbaum-style and Lyapunov-style estimates of a nonlinear regime boundary. |
| Control chaos | examples/lycan-internals/demo_control_chaos.lycs | Chooses controllers around a drifting nonlinear system. |
| Takeaway chaos replay | examples/lycan-internals/demo_takeaway_chaos_replay.lycs | Compares operational policies against chaotic demand behavior. |
| Cyber triage | examples/lycan-internals/demo_cyber_triage.lycs | Chooses response priority from incident context. |
| Flood response | examples/lycan-internals/demo_flood_response.lycs | Scores resilience actions under changing emergency conditions. |
| Spacecraft fault manager | examples/lycan-internals/demo_spacecraft_fault_manager.lycs | Chooses fault response policy from spacecraft telemetry signals. |
What to read by goal
| Goal | Read |
|---|---|
| Evaluate as an agent-safety/control-plane buyer | scripts/demo-agent-governor.py, scripts/demo-self-evolve.sh, scripts/demo-containment.py, scripts/demo-tls-gateway.py |
| Understand the commercial wedge | examples/demo-governed-llm-routing.sh and examples/llm-routing/ |
| Validate before rollout | examples/offline-eval/ and examples/ab-harness/ |
| Integrate into an application | examples/retry-tuning/ and examples/syntra-node/ |
| Understand operational capsules | examples/predictive-autoscaling/, examples/anomaly-routing/, examples/seasonal-fraud-threshold/ |
| See the runtime substrate | examples/lycan-internals/ |
| Explore proof-search/formalization handoff | examples/proof-lab/ |
What not to conclude
Do not stop at "HTTP API" or "bandit service." Those are implementation surfaces. The claim being demonstrated by the demos is narrower and stronger:
Syntra is a self-hosted runtime for repeated operational decisions where context arrives now, outcomes arrive later, and the policy must improve without turning the hot path into an opaque model call.