Skip to content

Reference demo

Current frozen system · HydroCore-v5 · v0.2.1 · source snapshot 4bbf6fa3ff9f

The REFERENCE INCIDENT is a deterministic, checksummed, progressive replay of HydroSwarm's frozen, WNTR-backed golden scenario. It is a product/workflow demonstration, not final HydroCore-v5 benchmark evidence. Every value traces back to a real run of scripts/run_golden.py's GoldenScenarioRunner and the frozen scenario/network fixtures.

Label: REFERENCE INCIDENT · CHECKSUMMED REPLAY. Supporting copy: "Replaying a checksummed HydroSwarm reference workflow generated from the frozen WNTR-backed scenario. Not live telemetry."

How to see it

  • From a running instance: choose Run Reference Incident on first launch, or navigate to /?experience=reference.
  • Regenerate the artifact: python scripts/build_reference_demo.py writes artifacts/reference-demo/reference-incident-v1.json and manifest.json.
  • Served by the backend: GET /api/reference-demo, resolved by hydroswarm.runtime.paths.resolve_reference_demo_path.

For the current V5 serving identity, use Final system. The reference-demo resolver is independent of model-bundle selection.

Reference Incident vs Live Example

Reference Incident is a deterministic checksummed replay of a frozen WNTR-backed workflow. Live Example runs the current frozen v0.2.1 runtime on the same bundled scenario inputs. They are not expected to reproduce identical downstream states. In the current bundled Live Example, the governed runtime abstains rather than continuing into planning; that is a current runtime observation, not M11.6 benchmark evidence.

Workflow

flowchart TD
  M0["alert"] --> M1["initial_uncertainty"]
  M1 --> M2["evidence_insufficient"]
  M2 --> M3["sample_recommended<br/>PAUSE: collect reference sample"]
  M3 --> M4["sample_received"]
  M4 --> M5["posterior_contracted"]
  M5 --> M6["plans_generated"]
  M6 --> M7["unsafe_plan_rejected"]
  M7 --> M8["safe_plan_verified"]
  M8 --> M9["human_approval_boundary<br/>PAUSE: approve plan"]
  M9 --> M10["completed<br/>final audit hash; Replay unavailable"]

Source: diagrams/reference-incident-flow.mmd.

There are two meaningful pauses: evidence collection and human approval. The replay never assumes that a sample was collected or that a response was approved without the corresponding explicit action.

The frozen milestone IDs shown above (e.g. safe_plan_verified) are artifact/stage identifiers for this deterministic replay, not a claim about real-world safety.

Stage-correctness

The generated artifact reveals fields only after the underlying frozen workflow has produced them:

  • alert: incident only;
  • initial uncertainty/evidence insufficient: broad candidate region;
  • sample recommended: recommendation before sample arrival;
  • sample received: raw evidence before posterior recomputation;
  • posterior contracted: updated source belief;
  • plans generated: proposals, not yet verified;
  • unsafe plan rejected: rejection appears after verification;
  • safe plan verified: verified modeled outcome appears;
  • human approval boundary: selected verified plan still unapproved;
  • completed: approval and final audit hash appear.

Calibration semantics

The deterministic-classical reference replay does not run the live V5 split-conformal pipeline. It therefore must not display live calibration coverage as if it were current model evidence. A LIVE incident retains current calibration/applicability fields.

Replay availability

The Reference Incident is a deterministic, checksummed, progressive replay of a frozen WNTR-backed workflow. That replay demonstrates the staged workflow itself — explicit sample and human-approval pauses, exact verification state, and a final provenance/audit hash.

Post-completion Replay of the Reference Incident is intentionally unavailable. The frozen reference artifact does not contain individual event-ledger records, so there is nothing to populate the Replay workspace with after completion. The application deliberately leaves Replay in an "unavailable" state rather than rendering an empty or fabricated ledger. This is a stated limitation of the reference artifact, not a bug: a LIVE incident's own audit/history is a separate, independently supported code path and is not affected by this limitation.

What the reference incident proves—and does not

It is strong evidence that the product can present the intended governed workflow, provenance labels, exact-plan verification states, and the human approval pause.

It does not establish V5 predictive accuracy, locked-test performance, novel-topology generalization, or field safety. Those claims belong to Scientific evidence. It also does not provide a populated post-completion Replay workspace — see Replay availability above.

Validation

The repository's reference-demo end-to-end tests check deterministic artifact construction, hash ties, stage ordering, candidate/sample behavior, verification status ordering, no premature approval, and completion only after the human approval event.