napkin-math(compress): cross-bucket promoter for gate-shaped items misfiled under risks

[neoneye]

Summary

Addresses the residual paperclip v53c failure mode: the LLM sometimes files a tripwire (If $X exceeds threshold, then <downside>, OR the declarative <X> exceeds threshold, <consequence>) under risks_and_shocks instead of gates_and_thresholds. The misfiled item then competes against actual risks at top-N selection and can fall out of the public output entirely.

Approach

A deterministic post-processor over the LLM emissions, not a prompt rule. The first iteration of this PR tried a risks-side prompt rule — review correctly flagged the causal model was wrong (gates emits before risks, so a risks-side prompt rule cannot move items into gates, and risks creating a worse failure mode where both buckets miss). That commit has been reverted on this branch; the gates and risks prompts are now back to their pre-PR state.

Code change

1. has_gate_shape(line) — true when the surface form matches either:

   • Canonical if/then numeric: If <... digit token ...> then <consequence>
   • Declarative comparison numeric: <subject> + comparison verb + <threshold with digit> + comma/colon + <consequence> — the actual v53c phrasing. Recognised comparison verbs: exceeds, falls below, drops below, rises above, breaches, is above/below/greater than/less than/more than, reaches, surpasses. Causal verbs (X risks Y, X causes Y, Failure of X leads to Y) are not recognised — those stay in risks.

   Numeric guard: separator comma/colon must not be followed by another digit, so commas inside numbers like $75,000 don't split the match. Qualitative if-then sentences without a numeric token between if and then are intentionally excluded so the promoter doesn't steal categorical/approval/deadline gates that the LLM already categorises correctly.

2. gate_shape_promotion(line) — returns the if/then form of line if it has any recognised gate shape, else None. For declarative inputs it produces a deterministic if/then rewrite preserving the gates bucket's output contract. Acronym casing is preserved (API job queue latency... rewrites to If API ..., NOT If aPI ...) — the case adjustment only fires on regular capitalised words (uppercase followed by lowercase). line_original is intentionally NOT rewritten; it preserves the source's native phrasing.

3. promote_gate_shaped_risks(gates, risks) — scans risks for gate-shaped items whose normalised source_quote is NOT already in the gates pool. Promoted items are MOVED to the gates candidate pool (not copied) so the risks slot is reclaimed. Items already in gates by quote are left in risks untouched (within-bucket dedupe is the existing 'do not restate' prompt rule's job).

4. Wiring — defers annotate_scored_items for gates_and_thresholds and risks_and_shocks until both have completed first+second-pass merging. After the bucket loop, the promoter runs on both merged pools, then annotate_scored_items fires on the augmented gates pool and the remaining risks pool. The promoter's count is exposed in per_bucket.gates_and_thresholds.cross_bucket_promoted_count for downstream auditing.

Empirical posture

• Unit tests: 44 pass in test_compress_report_section.py. New cases include: has_gate_shape true/false for both if/then and declarative shapes; positive regression on the literal v53c phrasing Middleware development bid exceeds $75,000, consuming budget...; negative regression on the genuine risk Supply chain disruption: 4 to 6 weeks delay and $15,000 cost increase.; end-to-end promoter test that asserts the v53c-shaped risk is moved with line_english rewritten to if/then form and source_quote / scores / status / line_original preserved; acronym preservation (API stays API); regular capitalisation lowering (Middleware → middleware).
• v56 regression sweep (paperclip 3× + 5 other probes, 290 risks candidate lines, 32 plan×section cells): 0 promotions fired. The v53c shape did NOT recur in this same-LLM same-session sweep; the v56 risks emissions are dominated by causal forms (X risks Y, X causes Y, Failure of X leads to Y) which the detector intentionally rejects. The change is a deterministic backstop for a rare LLM failure mode, analogous to PR napkin-math(bounds): Phase 4 runtime + schema readiness #747's calculation-output strip rule which also fired 0 times on its v48 regression corpus.
• The 8-run regression otherwise shows typical same-LLM variance (mostly ±1-3 items per cell, no systematic over-narrowing). One section had a 0-candidate emission failure (paperclip v56c expert_criticism gates), unrelated to this change — the gates LLM call itself is unchanged; the promoter only acts post-LLM.

What this PR explicitly does NOT claim

• Does not claim to "fix the v53c failure" by changing live LLM behaviour in the v56 session — the v53c-shaped failure didn't recur, so the promoter is uncalled.
• Does not claim to improve any bucket count metric — the v56 vs v53 count deltas are within typical same-LLM variance.
• Does claim to provide a deterministic backstop for the structural failure mode that PRs napkin-math(compress): second-pass shifts the variance failure from emission to ranking #743/napkin-math(compress): paraphrase-tolerant quote verification #744 left open: the LLM correctly emits the item with qv=True somewhere, but in the wrong bucket. When that recurs, the promoter routes it to gates with a clean if/then rewrite that preserves the gates bucket's output contract and acronym casing.

Test plan

• pytest worker_plan/.../tests/test_compress_report_section.py (44 pass)
• v56 paperclip 3× + 5 regression probes — no systematic over-narrowing, 0 false-positive promotions
• Reverted the wrong risks-side prompt rule from this PR's first commit
• Detector covers both if/then numeric AND declarative comparison forms (the actual v53c phrasing)
• Acronym casing preserved by the if/then rewrite
• CI green on this branch

🤖 Generated with Claude Code

[neoneye]

Addressed review (1755269c). The risks-side prompt rule has been reverted on this branch; the new approach is the deterministic post-processor (option 2):

1. has_gate_shape(line) — true for If <... digit token ...> then <...> surface form. Language-neutral.
2. promote_gate_shaped_risks(gates, risks) — scans risks for gate-shaped items whose normalised quote is not in gates, moves them to the gates candidate pool, removes them from risks. Items already in gates by quote are left in risks untouched.
3. Wiring: defers annotate_scored_items for both buckets until both have emitted; runs promoter between LLM emission and top-N filter.

Empirical finding worth surfacing: 0 promotions fired across all 8 v56 runs (paperclip 3× + 5 regression probes, 32 plan×section cells). The v53c-style miscategorisation didn't recur in this LLM session, so the promoter had nothing to act on. The change is a deterministic backstop, same posture as PR #747's calculation-output strip (which also fired 0 times on its v48 regression corpus).

37 unit tests pass (28 prior + 9 new). PR title and body rewritten to reflect the new approach and the honest empirical posture.

[neoneye]

Addressed (b495818b). The detector now actually covers the v53c shape:

Extended has_gate_shape to recognise two forms:

1. Canonical If <... digit ...> then <...> (unchanged from prior commit)
2. Declarative <subject> + comparison verb + <threshold with digit> + comma/colon + <consequence> — the actual v53c phrasing

Comparison verb list (structural cues, not domain vocabulary): exceeds, falls below, drops below, rises above, breaches, is above/below/greater than/less than/more than, reaches, surpasses. Causal verbs (X risks Y, X causes Y, Failure of X leads to Y) are NOT recognised — those stay in risks.

Numeric guard: separator comma/colon must not be followed by another digit so commas inside numbers like $75,000 don't split the match.

Deterministic if/then rewrite preserves the gates bucket's output contract: a declarative line is rewritten as If <subject> <verb> <threshold>, then <consequence> with case adjustments. line_original stays unchanged.

The exact tests you requested are in:

• Positive: test_has_gate_shape_accepts_declarative_v53c_form uses the literal v53c line "Middleware development bid exceeds $75,000, consuming budget planned for the physical handoff accumulation system."
• Negative: test_has_gate_shape_rejects_genuine_risk_with_colon_delay uses "Supply chain disruption: 4 to 6 weeks delay and $15,000 cost increase."
• End-to-end: test_promote_gate_shaped_risks_rewrites_declarative_v53c_form asserts the v53c-shaped risk is promoted with line_english rewritten to if/then form, source_quote/scores/status preserved, line_original untouched, and the item removed from risks.

42 unit tests pass total.

Honest empirical posture (audited the extended detector against v56): across 290 risks candidate lines in the 8 v56 runs, 0 match the extended pattern. The actual v56 risks emissions are dominated by causal forms, not declarative comparison. The v53c shape didn't recur in v56. The promoter is still a deterministic backstop — covered by the regression test on the literal historical line, dormant on the live sweep. Same posture as PR #747's calculation-output strip rule.

PR title remains "addresses paperclip v53c miscategorisation" — the change now actually does, by regression-test construction on the literal phrasing.

[neoneye]

Addressed both cleanup items (542cc196):

1. Acronym casing preserved. _lowercase_first_preserving_acronyms only lowercases the first character when it's uppercase AND followed by a lowercase letter (a regular capitalised word). API / OPC UA / 5G all have a non-lowercase second character, so they're left unchanged. Two new regression tests: one asserts API job queue latency exceeds 100ms, ... rewrites to If API job queue latency exceeds 100ms, then ... with aPI explicitly absent; the other asserts Middleware still lowers to middleware so the normal capitalisation path still works.
2. PR body updated to reflect 44 tests (was 37) and to describe both the canonical if/then detector AND the declarative comparison detector with the comparison-verb list, numeric guard, rewrite behaviour, and acronym preservation.

44 unit tests pass.