core-labs/core
core-labs/core
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

docs(reader-arc): recalibration ADR-0251 + preserved knowledge + overfit inventory #81

Merged
core-labs merged 1 commit from docs/reader-arc-recalibration into main 2026-07-19 20:46:43 +00:00
Conversation 0 Commits 1 Files changed 4 +627
core-labs commented 2026-07-19 20:43:23 +00:00
Owner
Copy link

Docs-only deliverable for the reader-arc recalibration. No implementation. Design-first gate — stops here, awaiting ruling on ADR-0251 §5.

What this is

feat/reader-inc2-caseband (PR #80) pursued case 0148's first real conversion via bespoke per-case
regexes and, per its own review thread, introduced debt: a fail-open loosening of the shared _token_in
grounding guard, an empty committed "proof" artifact, and bundled provenance. Per the recalibration
ruling on that PR, the branch is abandoned (not merged) and bespoke-per-case regex work on this reader is
halted. This PR is the documentation trail for that decision — four new files, zero code changes.

Contents

  • docs/adr/ADR-0251-...md — the recalibration ADR. §§1-4 record the halt/reset/eradication decision
    as executed and verified by command (local main == origin/main == forgejo/main @ 854ce023;
    the branch's specific debt is grep-confirmed absent from main; the #77/#78 foundations are
    grep-confirmed present and untouched; smoke 176 passed; holdout_dev/v1 correct=5 wrong=0 refused=495 (n=500) matches the pre-branch baseline exactly). §5 proposes (does not build) a geometric
    surface→canonical normalization spike using existing off-serving primitives (conformal_procrustes,
    VocabManifold, the cognition pipeline's Phase C anti-unification telemetry), with an explicit
    wrong=0-gated generalization-ratio measurement contract and explicit disclaimers (no Smith-chart
    algebra per ADR-0246 §6, no Fibonacci "dimensional cascade"). Marked Proposed — awaiting ruling.
  • docs/research/reader-arc-recalibration-preserved-knowledge-2026-07-19.md — the two genuinely-
    general facts pulled from the abandoned branch before it's set aside: the has_numeric_token
    multiplier-blindness bug (root cause traced to classify_sentence()'s closed word-number set; a
    synthetic repro plus 23 real holdout_dev/v1 corpus cases matching the failure shape; confirmed not
    a wrong=0 hazard because the independent ADR-0191 completeness guard already catches it via
    lookup_multiplier; a minimal fix idea documented but deliberately not landed), and case 0148's
    correct-for-the-right-reason decomposition. Also documents where the branch's own narrative overstated
    what it had actually verified (the claimed-committed graph dump was 0 bytes).
  • docs/research/0148-first-conversion-proof.md — the real, non-empty 0148 graph dump (the abandoned
    branch committed this file empty). Generated fresh from the branch's code in a throwaway worktree
    (since removed) and cross-checked against main's current refusal behavior for the same case.
  • docs/research/reader-arc-overfit-inventory-2026-07-19.md — 31 file:line-cited bespoke/single-case
    surface patterns in the reader on main, as candidates for what a future normalizer would need to
    subsume. Inventory only — nothing is removed or modified; the existing regex reader keeps serving
    unchanged.

What this PR does NOT do

No code changes. Does not touch generate/*.py, does not re-land any part of feat/reader-inc2- caseband, does not start the ADR-0251 §5 spike, does not touch the sealed test.

Test plan

  • git status/SHA verification: local main == origin/main == forgejo/main @ 854ce023 (see
    ADR-0251 §2 for the exact commands/output)
  • grep confirms the abandoned branch's debt (_COMPARE_MASSNOUN_RE, the loosened _token_in
    signature, any empty *proof*.md) is absent from main
  • grep confirms compare_multiplicative (#77) and NEUTRAL_COUNT_VERBS (#78) remain present and
    unmodified
  • uv run core test --suite smoke -q → 176 passed, 0 failed
  • uv run python -m evals.gsm8k_math.holdout_dev.v1.runner → correct=5 wrong=0 refused=495 (n=500),
    matching the committed report.json baseline exactly
  • PR #80 commented (knowledge-capture-complete + close-not-merge request) — closing the PR itself
    needs manual action; no close/update-state tool is exposed on the Forgejo MCP connector used here

🤖 Generated with Claude Code

**Docs-only deliverable for the reader-arc recalibration. No implementation. Design-first gate — stops here, awaiting ruling on ADR-0251 §5.** ## What this is `feat/reader-inc2-caseband` (PR #80) pursued case 0148's first real conversion via bespoke per-case regexes and, per its own review thread, introduced debt: a fail-open loosening of the shared `_token_in` grounding guard, an empty committed "proof" artifact, and bundled provenance. Per the recalibration ruling on that PR, the branch is abandoned (not merged) and bespoke-per-case regex work on this reader is halted. This PR is the documentation trail for that decision — four new files, zero code changes. ## Contents - **`docs/adr/ADR-0251-...md`** — the recalibration ADR. §§1-4 record the halt/reset/eradication decision as **executed and verified by command** (local `main` == `origin/main` == `forgejo/main` @ `854ce023`; the branch's specific debt is grep-confirmed absent from `main`; the `#77`/`#78` foundations are grep-confirmed present and untouched; smoke 176 passed; `holdout_dev/v1` `correct=5 wrong=0 refused=495 (n=500)` matches the pre-branch baseline exactly). §5 **proposes** (does not build) a geometric surface→canonical normalization spike using existing off-serving primitives (`conformal_procrustes`, `VocabManifold`, the cognition pipeline's Phase C anti-unification telemetry), with an explicit `wrong=0`-gated generalization-ratio measurement contract and explicit disclaimers (no Smith-chart algebra per ADR-0246 §6, no Fibonacci "dimensional cascade"). Marked **Proposed — awaiting ruling**. - **`docs/research/reader-arc-recalibration-preserved-knowledge-2026-07-19.md`** — the two genuinely- general facts pulled from the abandoned branch before it's set aside: the `has_numeric_token` multiplier-blindness bug (root cause traced to `classify_sentence()`'s closed word-number set; a synthetic repro plus 23 real `holdout_dev/v1` corpus cases matching the failure shape; confirmed **not** a `wrong=0` hazard because the independent ADR-0191 completeness guard already catches it via `lookup_multiplier`; a minimal fix idea documented but deliberately **not landed**), and case 0148's correct-for-the-right-reason decomposition. Also documents where the branch's own narrative overstated what it had actually verified (the claimed-committed graph dump was 0 bytes). - **`docs/research/0148-first-conversion-proof.md`** — the real, non-empty 0148 graph dump (the abandoned branch committed this file empty). Generated fresh from the branch's code in a throwaway worktree (since removed) and cross-checked against `main`'s current refusal behavior for the same case. - **`docs/research/reader-arc-overfit-inventory-2026-07-19.md`** — 31 file:line-cited bespoke/single-case surface patterns in the reader on `main`, as candidates for what a future normalizer would need to subsume. **Inventory only** — nothing is removed or modified; the existing regex reader keeps serving unchanged. ## What this PR does NOT do No code changes. Does not touch `generate/*.py`, does not re-land any part of `feat/reader-inc2- caseband`, does not start the ADR-0251 §5 spike, does not touch the sealed test. ## Test plan - [x] `git status`/SHA verification: local `main` == `origin/main` == `forgejo/main` @ `854ce023` (see ADR-0251 §2 for the exact commands/output) - [x] `grep` confirms the abandoned branch's debt (`_COMPARE_MASSNOUN_RE`, the loosened `_token_in` signature, any empty `*proof*.md`) is absent from `main` - [x] `grep` confirms `compare_multiplicative` (#77) and `NEUTRAL_COUNT_VERBS` (#78) remain present and unmodified - [x] `uv run core test --suite smoke -q` → 176 passed, 0 failed - [x] `uv run python -m evals.gsm8k_math.holdout_dev.v1.runner` → `correct=5 wrong=0 refused=495 (n=500)`, matching the committed `report.json` baseline exactly - [x] PR #80 commented (knowledge-capture-complete + close-not-merge request) — closing the PR itself needs manual action; no close/update-state tool is exposed on the Forgejo MCP connector used here 🤖 Generated with [Claude Code](https://claude.com/claude-code)
core-labs added 1 commit 2026-07-19 20:43:27 +00:00 
Records the reader-arc recalibration: halt bespoke-per-case regex work,
verified clean-base reset to main@854ce023, and eradication of
feat/reader-inc2-caseband's (PR #80) debt (loosened _token_in, empty
proof file, bundled provenance) without merging that branch.

Preserves the genuinely-general knowledge before the branch is retired:
the has_numeric_token multiplier-blindness bug (root cause + minimal fix
idea, deliberately not landed) and case 0148's correct-for-the-right-reason
decomposition, both independently re-verified against source rather than
copied from the branch's own claims (which included a stale narrative
about a graph-dump proof file that was actually committed empty).

Adds a file:line-cited overfit inventory (31 bespoke/single-case surface
patterns) as inventory only -- nothing removed, the existing regex reader
keeps serving unmodified.

ADR-0251 proposes (not builds) a geometric surface->canonical
normalization spike using existing off-serving primitives
(conformal_procrustes, VocabManifold, the cognition pipeline's Phase C
anti-unification telemetry), wrong=0-gated against holdout_dev/v1, with
Smith-chart algebra and any Fibonacci "dimensional cascade" explicitly
disclaimed. Awaiting ruling -- no implementation started.
core-labs merged commit 1de6dbc945 into main 2026-07-19 20:46:43 +00:00
core-labs deleted branch docs/reader-arc-recalibration 2026-07-19 20:46:45 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: core-labs/core#81
No description provided.
Delete branch "docs/reader-arc-recalibration"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?