Explorer v0.1 contract and packaging decision

Status: packaged Asteria renderer available; the generated enterprise-agentic smoke adapter from issue #149 is available as a separately versioned profile.

Explorer v0.1 is a deterministic projection contract and renderer boundary for benchmark packages that SynthWorld already publishes. It is not a general graph browser, an authorization engine, or an arbitrary generated-world interface. Its generated-world support is bounded to the released enterprise_agentic smoke contract; standard and longitudinal tiers are not supported.

The package ships the Asteria projection contracts introduced in 0.15.0, the independently versioned generated-smoke contracts, and a checksum-verified renderer for the published Asteria Agentic v1 package and verified generated enterprise-agentic smoke packages. The renderer produces one deterministic, self-contained HTML file. It does not accept arbitrary topology files, fixed-reference enterprise-agentic packages, unsupported generated tiers, or experiment visualizations.

Immediate scope

The first supported profile is agent-authority-v1, projected from the public artifact set of the published asteria-agentic-v1 benchmark. Its view includes:

• organisation, department, principal, logical-agent, runtime, credential, delegation, resource, and action-attempt nodes;
• containment, ownership, delegation, runtime, credential, and attempted-action relationships; and
• the public event timeline needed to replay grants, actions, revocation, evidence disposal, and audit activity.

The second supported profile is enterprise-agentic-generated-v1, projected from a verified released generated enterprise-agentic public package (smoke tier). It reuses the same node, edge, and timeline contracts and the same shared renderer; only its independently versioned source record, profile literal, and computed layout differ. Candidate C08 v2 and fixed-reference enterprise_agentic packages are not Explorer inputs. Large-world controls, standard and longitudinal generated tiers, and tier comparisons remain outside the released generated smoke contract and Explorer v0.1.

Artifact boundary

Explorer keeps its released Asteria public projection, evaluator overlay, and layout v1 contracts independently versioned. The packaged Asteria renderer uses layout manifest 2.0.0; layout 1.0.0 remains unchanged for compatibility. The generated projection and generated layout are separate 1.0.0 contracts.

The generated enterprise-agentic profile does not widen the frozen agent-authority-v1 projection, evaluator, or layout literals in place; it is an independent 1.0.0 contract with its own source record. The shared evaluator overlay contract is reused unchanged for generated worlds because its digest bindings are already world-agnostic.

The public projection is constructed field by field from AgenticPublicBundle. It cannot carry expected decisions, canonical bindings, case labels, authority truth, or failure reasons. Evaluator annotations use a separate model and a separate physical artifact. Any evaluator-rendered HTML must display:

EVALUATOR VIEW - CONTAINS REFERENCE TRUTH

Evaluator data must never be embedded in public HTML behind a client-side switch. The evaluator contract has its own schema-version identity even when its current version number matches the public projection version.

Identity, normalization, and collisions

Projection node and edge IDs are UUID5 values under the dedicated Explorer v1 namespace. UUID names use explicit node and edge domains and length-framed UTF-8 components. This prevents delimiter ambiguity and separates otherwise equal node and edge inputs. Duplicate resulting IDs, source-reference failures, and open graph references fail validation rather than receiving suffixes.

Set-like nodes, edges, properties, annotations, references, and coordinates are sorted by stable identifiers or keys. Collection-valued properties remain JSON arrays rather than delimiter-joined strings. Timeline events retain source event-index order, use UTC, and must also be strictly increasing in time. Compound node parents must be acyclic. No ordering key may depend on an evaluator verdict, case kind, failure reason, or canonical binding.

Canonical JSON uses sorted object keys, compact separators, UTF-8, no NaN values, LF line endings, and exactly one trailing newline. Digests cover those exact bytes.

Deterministic layout inputs

The layout manifest records all inputs that can alter coordinates:

• public projection digest;
• world seed and world schema version;
• visualisation profile and visualisation-profile version;
• layout engine and exact engine version;
• algorithm and direction;
• node and layer spacing;
• viewport width and height;
• coordinate precision; and
• one canonical coordinate record per rendered node.

The renderer must not accept filesystem order, locale, host state, wall-clock time, or evaluator answers as layout inputs. The layout repeats the projection’s world seed and world schema version plus the visualisation profile and profile version. Validation compares those explicit values and also verifies the public-projection digest, which binds the complete projection source and published benchmark identity. Layout validation requires exactly one coordinate per bound projection node. Evaluator validation requires every annotation to bind a public action event and a known projection node, edge, or timeline event.

Packaged renderer

The shared browser renderer uses Cytoscape 3.34.1. For Asteria, it uses a committed layout generated by ELK.js 0.12.0. Exact npm dependencies are locked; the generated CSS, JavaScript, Asteria layout manifest, and dependency notices are checksum-bound package assets. Generated enterprise-agentic worlds use the deterministic Python grid described below, not a second browser renderer. The HTML does not load JavaScript, CSS, fonts, layout code, or data from a CDN or other network origin. Its content-security policy disables connections, and each layout is validated against the exact public projection digest before rendering.

The CLI binds to verified package directories rather than ambiguous standalone JSON:

synthworld visualize \
  --public-package ./asteria-agentic-v1/public \
  --view agent-authority \
  --output world.html

The public command reads only the public tree. To inspect reference truth, pass the separate evaluator tree explicitly:

synthworld visualize \
  --public-package ./asteria-agentic-v1/public \
  --evaluator-package ./asteria-agentic-v1/evaluator \
  --view agent-authority \
  --output evaluator-world.html

The complete loader verifies both inventories and their public/evaluator digest binding before deserializing either one. Evaluator HTML contains a prominent reference-truth watermark and evaluator annotations; public HTML contains neither. The command refuses to replace an existing output file.

The rendered page supports graph pan/zoom, node and edge inspection, event-timeline replay, revocation state, and evaluator annotations when explicitly enabled. The Asteria coordinates are pinned build artifacts; generated coordinates are computed deterministically in Python before rendering. In both profiles the browser receives preset coordinates and never recomputes layout from host state or evaluator truth.

Generated enterprise-agentic adapter

Generated worlds vary by seed and topology, so their coordinates cannot be committed build assets. The enterprise-agentic-generated-v1 adapter instead computes a deterministic kind-layered grid in Python from the projection alone and records the engine, spacing, viewport, and per-node coordinates in the independently versioned generated layout contract. The browser still receives preset coordinates and never recomputes layout.

synthworld visualize \
  --public-package generated-enterprise-agentic/public \
  --view agent-authority \
  --package-profile generated-enterprise-agentic \
  --output generated-public.html

The public command consumes only the verified public tree - no repository examples, private topology source, or evaluator access. Evaluator truth requires the separate evaluator tree, verifies both inventories, digest cross-binding, and declared generator conformance, and produces the same prominent watermark. Unsupported generated tiers or package versions fail explicitly.

Four surfaces stay distinct and must not be conflated:

• Topology import (scaffold/validate/compile-enterprise-access) authors a private fictional identity/access universe; it is never a rendering input.
• Generated benchmark artifacts (generate-enterprise-agentic --profile generated) are the checksum-bound public and evaluator trees.
• Visualization (visualize --package-profile generated-enterprise-agentic) projects and renders the verified public tree, with evaluator truth as a separately loaded, digest-bound, watermarked overlay.
• Authorization evaluation (validate/evaluate generated-enterprise-agentic) scores observed-action traces; the Explorer never imports its verdicts into public output.

Deliberate limitations

• The Asteria profile accepts only the published asteria-agentic-v1 public artifact-set digest; the generated profile accepts only verified released generated smoke packages.
• The legacy core identity world remains a deterministic smoke surface whose path topology carries no structural signal. It is not an Explorer v0.1 input, and a chain-shaped rendering of it would not be evidence of meaningful organisation structure.
• The generated enterprise-agentic smoke package is not silently coerced into the Asteria view; its separately versioned profile is selected explicitly with --package-profile generated-enterprise-agentic.
• The HTML is a local inspection aid, not a hosted service, policy engine, agent runtime, or evaluator report.
• The npm toolchain is needed only to reproduce or verify committed renderer assets; released-wheel users need only Python and a browser.