# SCN-001 — reviewed memory scenarios and public explanation

Accepted implementation scope, 2026-10-03. This specification summarizes the
approved product plan; it contains no private conversation or training material.

## End-to-end contract

Original wording → reviewed interpretation → explicit context → ordered memory
operations → conditional predictions → observations → retained corrections.
This is a bounded grammar and memory workflow, not arbitrary language understanding.

- Exactly two branches, baseline and alternative, each with original source,
  language version, prior revisions and a canonical memory session. The alternative
  may start as a copy, but edits and replay are independent. Every replay starts empty.
- Assumptions are descriptive, visible context. They do not secretly add operations.
- Preserve Unicode, original wording, source spans, unknown clauses and corrections.
  Unsupported clauses leave bindings unchanged and mark the overall simulation partial.
- Predictions identify branch, revision, step and key. Observation records identify
  the same target and source kind: user report, imported observation or synthetic
  model replay. Values and explicitly missing bindings are distinct from no report.
- Compare observations individually as match, mismatch or not comparable. Preserve
  conflicting reports, old predictions and discrepancies across revisions. Do not
  auto-select a branch, erase a mismatch or infer its cause.
- Preview precedes explicit adoption. Recompute through the locally configured
  engine and check the complete review digest. Edits, reset and imports invalidate
  stale requests. A failed operation cannot replace the accepted record.
- Versioned `renigma.scenario` v1 documents link sources, memory sessions, model
  identity, assumptions, predictions, observations and review decisions. Validate
  strictly; imported records cannot select model paths, code or remote services.
- Provide local preview, adopt, compare and validate APIs. Existing plan and memory
  formats stay compatible. Wrapping a legacy memory session preserves its original
  language interpretation and requires review. Keep current request, source,
  vocabulary and event limits; reject excess without truncation or partial adoption.
- UI offers source editing, copy baseline, review/adopt, observation entry and
  comparison, correction history, import/export/reset. No autosave or external action.

Synthetic acceptance: baseline `Remember color is green. Set color to blue. Recall
color.` predicts blue at its read. Alternative `Remember color is green. Recall
color.` predicts green. A user-reported green at baseline's read is a mismatch.
The alternative does not retrospectively become the baseline.

## Public website contract

Seven actual static routes: home, language, context, transformations, scenarios,
evidence and local-control. Keep the existing public Site identity and Renigma
ink/lime/cobalt palette. Preserve old anchors and the exact context-loss demo.
Each branch connects human purpose, mechanism, example, evidence/limitations and
the next related concept. Preserve the home research objective: “Change the form.
Keep the meaning.” Use concrete tested detail rather than superiority claims.

Use a shared reviewed content manifest for claims, statuses, metrics, typed graph
relationships and timeline entries. References open on explicit controls with
hover/focus previews and tap support. The timeline distinguishes execution, review
and publication; unknown dates remain unknown. Keep essential content and navigation
available without JavaScript. Respect reduced motion and provide pause controls.

Show REP-004 alongside earlier evidence: latent-only repaired 2,062 errors without
regressions; prediction-plus-latent repaired 1,186 and regressed 93, leaving 969
errors in the hard seed. These are existing-world diagnostics, not new-world
generalization. Latest findings links must no longer call REP-004 proposed.

Public scenarios are synthetic traces generated by the canonical scenario engine.
Do not duplicate the grammar in JavaScript or contact localhost automatically.
Private workbench records never enter public assets. Generative artwork is
conceptual, with provenance, and distinct from data-derived scientific figures.

## Verification and delivery

Test canonical semantics, branch isolation, source preservation, partial behavior,
stale reviews, legacy wrapping, tampering, observation provenance and limits. Browser
checks must exercise review/adoption, discrepancies, failed import preservation,
reset, source panels, timeline, all routes, keyboard/mobile and reduced motion.
Public trace and metrics checks must compare against canonical outputs and reviewed
evidence. Run pytest and math-check using the designated isolated environments.

Publish the verified source and static site to the existing public GitHub repository
and Zulu-hosted Site, retaining deployment history and PolyForm Noncommercial terms.
Fresh-setting REP-005, learned coordinates/language, NPU and other-harness replication
remain separately gated research; implementation of this workflow is not their result.