# STATUS: FROZEN

Frozen **2026-04-09** pending completion of the Recoil engine refactor.

## Do not write to this directory

The 5 LEARNINGS + 2 ANTI_PATTERNS from `tartarus-r1` (2026-03-25) are preserved
as historical data. Do not edit the existing files. Do not run `/analyze-production`
against this directory. Do not add new entries.

## Pointers

- **Decision record:** `docs/product/engine-memory-v2-consult-brief.md`
- **Auto-memory entry:** `project-engine-memory-analysis.md`
- **Brainstorming transcript:** 2026-04-09 session (recoil-engine-refactor domain)

## What triggered the freeze

Diagnosed during brainstorming 2026-04-09: the system has never actually changed
JT's behavior despite 1 successful `/analyze-production` run. Four failure modes
identified:

- **B. Narrative-only blind spot.** Designed to consume script-side exhaust
  (`script_doctor_brief`, `transition_gate_report`). The visual pipeline — where
  the next year of refactor work lives — has no equivalent input.
- **C. Promotion bottleneck.** After 1 run, 0 learnings have promoted from
  provisional → active. The "active" knowledge surface that gets loaded into
  future sessions is permanently empty.
- **E. Wrong format / wrong consumer.** Human-readable markdown in git is the
  wrong shape. The real need is probably a structured store the engine queries
  at decision time, not a static doc loaded at session start.
- **F. Framing is muddled.** Lumps together things that want different stores:
  anti-patterns, per-project state, production telemetry, policy hints for the
  feedback agent.

## What's next

1. Wait for the refactor's Phase 3 cutover (EP001 overnight unattended on the
   new substrate: sidecars, `ops.log.jsonl`, 8 critics with response_schema
   enums, feedback agent decisions, `ref_resolver` traces).
2. Observe 2-3 weeks of real use. Keep a single growing friction log at
   `scratchpads/refactor-friction.md` — one-line entries for every "I wish I
   knew," "I hit this again," "I'm doing this twice" moment.
3. Run a fresh `/consult` with the new-substrate exhaust + observed friction
   as input. The consult brief at `docs/product/engine-memory-v2-consult-brief.md`
   lists the exact inputs, the questions, and the success criteria.
4. Based on the consult output, either:
   - Design a new memory/learning layer (which may or may not absorb the
     frozen content from this directory), OR
   - Confirm the refactor itself is sufficient and retire the concept of
     a separate memory system entirely.