Methodical LLM re-coding of the 285 entries into the derived sidecar (high-confidence, lossless, zero mutation)

Goal

Re-code the existing 285 ledger entries into high-confidence coordinates by a METHODICAL, BATCHED LLM pass — NOT regex — writing them to the DERIVED coordinate_map sidecar (§01), NEVER into the stored entries (append-only §5 absolute). The 285 entries hold irreplaceable assets (verbatim teeth, dead-end generalizations, ‘untried next lever’ pointers); they stay byte-unchanged while their coordinates land in the sidecar so the history actively drives the matrix instead of being quarantined.

Implementation Sketch

Methodical LLM execution deliverable (driven by /continue-roadmap dispatching the LLM per batch). BATCH PROTOCOL (deterministic): order entries by ascending attempt_n; bounded batch size; carry the plan-declared taxonomies + the running set of prior-batch classifications forward as context so classification stays consistent across batches. For each entry: READ in full, classify (domain, shape_id, mechanism_class, mechanism_variant, failure_mode) against the §01 taxonomies, write the result to the coordinate_map sidecar keyed by attempt_n at coordinate_confidence: high. The 285 stored entries are NEVER edited (append-only §5; the sidecar is the §13.3 derived overlay). A regex pre-pass MAY produce a SUGGESTION the LLM verifies+corrects; regex output is NEVER the final coordinate. An entry the LLM cannot confidently classify is recorded coordinate_confidence: low with an explicit reason — never guessed. SECOND PASS: after the bulk pass establishes the taxonomy’s real distribution, re-visit every low-confidence entry once with the now-established patterns; lift to high where the pattern now resolves it, leave low+reason otherwise. High-confidence coords drive propagation (D3/D4); low-confidence stay barred. Coverage gate: all 285 carry a sidecar coordinate (high OR explicitly-low-with-reason) — zero un-migrated, zero lost, zero stored-entry mutation.

Spec References

User directive 2026-06-29 (methodical LLM migration, do not lose critical info). Round-1 review (codex-F1/opencode-F5 sidecar-not-mutation, opencode-F7 batch protocol, opencode-F6 second pass). Design v7 D4 + D7 + D-CORE. learning-ledger.md §5 append-only, §13.3 derived overlay, §3 teeth-preservation.

Work Items

• Methodical batched LLM re-coding to the DERIVED coordinate_map sidecar (keyed by attempt_n): deterministic ascending-attempt_n order, bounded batch, taxonomies + prior-batch classifications carried as context; the 285 stored entries are NEVER edited.
• Regex pre-pass = SUGGESTION only (never final); unclassifiable entry = low-confidence + explicit reason, never guessed; SECOND PASS re-visits low-confidence entries once after bulk patterns establish + lifts where resolvable.
• Coverage gate: all 285 carry a sidecar coordinate (high OR explicitly-low-with-reason); high drives propagation, low barred; append-only assertion: every stored entry byte-unchanged.

Fresh intel (regenerated)

{ “schema_version”: 1, “target”: { “kind”: “plan-section”, “ref”: “ledger-matrix/s-bd0e7ad3” }, “generated_at”: “2026-06-29T18:00:13.112558+00:00”, “graph_state”: { “head_sha”: “63d372a6”, “last_code_import_at”: “2026-06-29T17:55:20.269Z”, “embedding_stale”: true, “insights_stale”: false, “cpg_stale”: false }, “surfaces”: {}, “summary”: “intel-package[plan-section:ledger-matrix/s-bd0e7ad3] agent-authored dossier”, “degraded”: false, “dossier”: { “objective_symbols”: [], “tiers”: {}, “agent_authored”: true, “constrained_symbols”: [ “validate_ledger_append_only”, “assemble_corpus”, “read_ranked”, “_FORBIDDEN_REFERENCES” ], “difficulty”: { “class”: “routine”, “research_online”: false, “research_mode”: “auto”, “signals”: [], “plan_dir”: “/home/eric/projects/ori_lang/plans/ledger-matrix” } } }

(full dossier: backcatalog-migration—s-bd0e7ad3.intel.json)