# AkashicNEXUS Executable Package · Proposed Architecture

Status: Design Blueprint. No executing orchestration runtime or ChatGPT plugin is shipped in this reference package. No target date is assigned.

## Component Boundaries

Keep the component registry and prompt content portable. A component descriptor should include ID, classification, version, scope, input/output schema, source locator, source revision or retrieval date when known, prompt hash, dependencies, model/tool requirements, permissions, expected checks and rerun triggers. Unknown source revisions remain null.

AkashicSTATE records identity, versions, edition, scoped maturity, separate Meta/Mettā assessments, assessor, evidence, dependencies, visibility, HOLD, planned/actual dates and immutable history. It must not turn a recommendation into a completed action. Reassessments append a record linked to the prior record.

## Runtime To Implement

1. Validate registry and input schemas. Reject missing required inputs rather than guessing.
2. Resolve an explicit route and select relevant lenses. Record the reason for including or omitting each optional lens.
3. Run an explicitly configured model adapter with recorded model, settings, component version, source set and prompt hash. Store input/output hashes and parent run IDs.
4. Validate structured outputs and carry uncertainty, unsupported claims and missing sources forward.
5. Implement dependency invalidation for changed wording, source sets, prompts, assets or versions. Mark affected outputs stale. Preserve prior runs and resume with idempotent operations.
6. Record actual source, calculation, browser, PDF and deployment checks separately, with reproducible methods and evidence. A model-generated assertion cannot establish that a browser check ran.
7. Deliver only when required checks, access rules and HOLD permit it. Require appropriate user authorisation for external actions. Store confirmed receipts, not predicted URLs or dates.

## Deployment Targets

A static host can serve the included webpages without a build step. Deploying those files only publishes the reference guide; it does not create the runtime.

An executable service needs a server, durable storage, per-user access controls, model adapters, source retrieval and separate publishing adapters. For a future Sites implementation, use a server-backed project with durable records and an authenticated MCP endpoint. Do not copy the current project's hosting identity into another user's deployment.

A ChatGPT/Codex plugin can package reusable instructions plus MCP tools and optional UI. Suggested tool boundaries: `get_component`, `plan_run`, `get_run`, `record_check`, `compare_runs`, `get_state`, `record_state` and a controlled `rerun_affected`. These names are proposed contracts, not available endpoints. Define read/write annotations, authentication, access, idempotency, error recovery and review cases before enabling writes.

Official documentation checked 19 November 2026:
- https://developers.openai.com/plugins
- https://developers.openai.com/plugins/build/skills
- https://developers.openai.com/plugins/build/app-quickstart
- https://developers.openai.com/plugins/plugin-guidelines

Recheck documentation and account capabilities before deployment. The reference archive is not an installable plugin ZIP.

## Acceptance & Release Gates

Run all three proposed cases with controlled inputs. Capture human baseline effort, NEXUS effort, source omissions, unsupported claims introduced, planted discrepancy recall, false flags, rerun correctness and support interventions. Require independent reviewers for any independent-validation claim. Keep commercial demand and willingness-to-pay tests separate from technical pass/fail.

Silver requires observed end-to-end runs and changed-input reruns. Gold requires reliable connected execution and failure recovery. Platinum requires independent evaluation and sustained reproducibility. There are no automatic version-to-level promotions.
