#449 shipped an inert, duplicated, unvalidated pack off-brief; closed it but kept the on-thesis idea (ground the nouns GSM8K problems use) with the real scope it would need to be beneficial.
2.2 KiB
2.2 KiB
Backlog: everyday-atoms grounding pack (idea, not yet scoped)
Source: closed PR #449 (atomic_definitions_everyday_v1), opened off-brief by
the remote connector during the 2026-05-29 ADR-0179 work. Closed as not-mergeable
(inert, structurally duplicated, unvalidated, disconnected from the active arc),
but the idea is worth keeping.
The idea
A compact, curated pack of everyday-object atoms — the concrete nouns GSM8K word problems actually lean on (apples, baskets, boxes, strawberries, coins, …). This is on-thesis: it gives the engine "more pieces to the puzzle" so it can comprehend the entities a problem talks about, rather than only counting their numbers. It is finding/comprehending better, not storing another found answer.
Why #449 as shipped delivered no benefit
- Inert — nothing in
chat/,generate/,language_packs/, orcore/referenced it;default = false,status = "draft". No runtime effect. - Duplicated + mislocated — shipped the same 9 files twice
(
packs/everyday/andpacks/everyday/atomic_definitions_everyday_v1/), neither matching the flatpacks/<pack_id>/convention (cf.packs/en/). - Unvalidated — never run through the pack gate (
core pack validate); all gatesfalse; manifest checksums unverified (must hash the bytes on disk). - Disconnected — the math derivation/reader lane does not consume semantic packs, so even if loaded it would do nothing for GSM8K today.
What would make it beneficial (the real scope when it's time)
- One flat, conventional location:
packs/atomic_definitions_everyday_v1/. - Passes the pack gate: probes green, checksums hashing the written bytes,
validate_pack_dir/lift_from_packdelegation reviewed at the dynamic-validator trust boundary. - A consuming grounding path: an ADR that wires everyday-atom grounding into a comprehension/reader lane that actually reads it (otherwise it's a parked artifact). Likely adjacent to the ADR-0178 comprehension-guided composer, not to ADR-0179 extraction.
- Curated, not bulk — keep it compact (CLAUDE.md: no corpus bulk-ingest into runtime); grow it deterministically with pack tests.
Until 1–3 exist, this is an idea, not a pack.