From 2df1d9482ba00339f9a8fb21534292adb313cfee Mon Sep 17 00:00:00 2001 From: tegwick Date: Sun, 27 Sep 2026 01:53:48 +0200 Subject: [PATCH] Preregister authority projection comparison Assistant: codex Assistant-Model: gpt-6-astra Assistant-Session: 01a0cffe-206f-7323-89d3-b20ca4cf4bf9 --- .../cycles/2026-09-27-authority-projection.md | 51 +++++++++++++++++++ 1 file changed, 51 insertions(+) create mode 100644 research/evolution/cycles/2026-09-27-authority-projection.md diff --git a/research/evolution/cycles/2026-09-27-authority-projection.md b/research/evolution/cycles/2026-09-27-authority-projection.md new file mode 100644 index 0000000..969f34d --- /dev/null +++ b/research/evolution/cycles/2026-09-27-authority-projection.md @@ -0,0 +1,51 @@ +# LEARN-0015 — Separate authority from projection and delivery + +Status: preregistered before new source collection. Date: 2026-09-27 Europe/Berlin. +Reviewer: Codex, author-model; no independent human validation. +Stream: cross-stream (C/D). Comparison group: CD-authority-projection-documentary-v1. + +## Question and prior + +Do Server/Local/Replicated Authority and Server/Client/Streaming Projection describe +six reusable patterns, or a mixture of context choices and implementation strategies? +Prior: authority is scoped to particular data/decisions, acknowledgement levels differ, +and rendering placement need not determine authority. Dedicated drafts for all six +are absent. We expect some seed labels need decomposition rather than automatic +promotion. Classification inventories are editorial priors, not empirical evidence. + +## Registered method + +Follow research/plan/AuthorityProjectionComparison.md: at most six primary-source +documents, at least two independent source families, with versions/access dates and +claim locators. Reuse existing local reviews as priors only. Candidate source targets: +Firestore offline/transaction documentation for cached versus accepted writes, +Automerge documentation for local replicas/merge, and SvelteKit/React documentation +for server/client projection and streaming. Select at most six substantive documents; +failed fetches and path discovery do not supply evidence. No broad source expansion. + +For each seed describe authority scope, acceptance boundary, cache/replica meaning, +projection inputs, execution location and delivery timing. Produce claim table, +synthetic VC-01/VC-02 contrasting decision traces, and six retain/develop/context/ +merge/reject dispositions. Seek counterexamples to server-rendered implies server +accepted and locally rendered implies locally authoritative. Unsupported combinations +remain unknown. At most two new drafts, only if distinct contracts merit them; zero +is valid. No code execution/performance/accessibility/production claim. + +## Outcomes, effort and stop + +One bounded question, one disposition: corroboration, revision or inconclusive (or +merge/rejection if a seed decision warrants it). Record uncertainty and preserved +hypotheses; no automatic M1/M2 or canonical admission. Existing Zulip follow-up remains +open and authored navigation fixture paused. No inference planned; any later live +model execution uses llm-connect. + +Stop after six substantive documents or 45 active research minutes; allow 10 minutes +of repository verification. Start a monotonic clock after preregistration commit, +before new source collection. Active elapsed session minus timed check subprocesses +includes retrieval, interpretation and recording. Exclude preregistration and final +receipt refresh/commit/snapshot/sync administration after cutoff. Tokens/cost unknown. + +Append ledger and evidence report at review boundary; update candidate profiles and +classification only if contracts change or new candidates are warranted. Snapshot +the resulting assessment even if candidate content is unchanged. Do not combine +this new documentary protocol with earlier runtime cycles to claim a trend.