docs: reflective-evidence citation policy for PR records

docs/architecture/reflective-evidence-citation.md

@@ -0,0 +1,93 @@
+# Citing reflective-session evidence in PR records
+
+Status: proposed (draft PR) · 2026-07-28
+
+NeuralCompose generates reflective/dialectic session logs (LLM dialogue
+transcripts) during hypnagogic and dialectic sessions. Those logs are
+private, local corpus material. This document fixes the boundary between
+that material and anything that enters a pull-request record, an ADR, or a
+committed artifact — so the same log is never duplicated across branches
+and transcript text never becomes accidental source history.
+
+## Implementation status
+
+This document is the normative policy.
+
+A reference implementation currently exists on the
+`feat/local-dialectic-review` lineage, pinned at commit
+`c8731955f258d91db40bb31c7b79b25284b94d8e`:
+
+- `Scripts/quarantine_dialectic_corpus.py`
+- `Scripts/review_quarantined_dialectics.py`
+
+Those files are not yet present on `main`. Until that implementation lands,
+no PR may claim that main-branch automation enforced this policy; the
+privacy and citation rules themselves still apply manually.
+
+When the implementation lands, its tests must prove that raw transcript text
+and embeddings cannot enter the committed manifest surface.
+
+### The pinned reference implementation, described
+
+At the pinned commit, `quarantine_dialectic_corpus.py` produces SHA-256
+parse reports, metadata-only event streams, and 9-field disposition blocks,
+written to the local `InteractionLogs/local-manifests/` directory outside
+the repository; `review_quarantined_dialectics.py` performs bounded loopback
+review with whitelisted persisted keys. Raw captures never enter git; that
+invariant is audited, not assumed.
+
+## Evidence classes — keep them separate
+
+| Class | Location | Committed? |
+|---|---|---|
+| Raw reflective transcript (`dialectic-turns-*.jsonl`) | local/private, outside the repo | **Never** |
+| Redacted session manifest (parse report, event stream, review summaries) | `local-manifests/` (local); schema definitions in-repo | Optional metadata only — no text, no embeddings |
+| PR evidence summary | PR description / acceptance doc | Yes — **claim + observation + content hash + tool version only** |
+| Architectural decision (ADR) | `docs/architecture/decision-log/` | Yes — independent reasoning and review |
+| Training/embedding corpus | separately consented, quarantined lane | Only under its own explicit disposition |
+
+## The citation rule
+
+A PR record, acceptance document, or ADR may cite reflective-session
+evidence **only** in this shape:
+
+```
+Claim:       <what is being asserted>
+Observation: <redacted, aggregate, or structural finding — no dialogue text>
+Source:      dialectic-turns-<date>.jsonl sha256=<parse-report source hash>
+             parser=<PARSER_VERSION> reviewer=<REVIEWER_VERSION>
+             (and/or promptHash=<generator fingerprint>)
+```
+
+- One capture = one content hash = one local manifest. PRs cite the hash;
+  they never copy transcript text, and the same log is never attached to,
+  or reproduced in, more than one branch.
+- If a claim cannot be expressed without quoting dialogue, it does not
+  belong in a PR record.
+
+## Reflective output is not an authority
+
+A reflective model's output is a **hypothesis generator or observation** —
+never a reviewer sign-off and never normative authority. Concretely:
+
+- **Event/labeling work (e.g. PR #44):** reflective sessions may *suggest*
+  event types or failure cases; they cannot promote heuristic observations
+  into labels, science-lane evidence, or clean-session credit. Promotion
+  requires the pre-registered gates, not conversation.
+- **Client/transport work (e.g. PR #45):** only device/client acceptance
+  evidence belongs in the record. Dialogue contents and personality-analysis
+  material do not belong in a transport PR.
+- **Reconstructed ADRs (e.g. ADR-009 / PR #47):** a reconstruction records
+  the decision as it was. Later reflective conversations postdate that
+  decision and must never be folded back into it — that would convert later
+  interpretation into false history. Amendments go in new, dated ADRs.
+
+## Source-pinning for external scaffolding
+
+Externally derived scaffolding (e.g. published alignment/evaluation work)
+may inform reflective-mode *governance* — assistant-character constraints,
+honesty/uncertainty behavior, evaluation design. It never silently
+authorizes embedding-geometry claims, encoder promotion, runtime
+architecture, training-data inclusion, or scientific labels. Any such use
+must pin the exact source (paper, file, commit); an unpinned attribution is
+not a source.