Scope the tail stage's frame gate to stage-attributable worsening - #571
Conversation
populace#570 round 2 established with post-fix replay receipts that the stage's absolute frame check was blocking on a condition the stage neither created nor can fix: collectibles fails at 83.9% with 67.2% of its combined mass inherited from pre-existing base rows, while the fidelity fix collapsed the only stage-caused failure (unrecaptured-1250 84.9% -> 34.2%, 1,824 -> 6,042 carriers). The stage check exists to catch stage-caused pathology; the calibrated artifact's absolute concentration is owned by the release-side gate, which measures export weights and carries its own reviewed per-run register — where collectibles was NOT among dense-P2's flagged columns. The frame check now fails a column only when it is over the absolute threshold AND the transfer strictly worsened its top-100 share vs the pre-stage frame; every column's pre/post shares and carrier counts ride the stage manifest as receipts either way. The donor-tail input check is untouched and unconditional. Review hardenings folded: the candidate merge asserts the donor/candidate column overlap decomposes into the declared partition (joint CG vector donor-owned; donor tax-unit mortgage/ALD outputs recipient-owned — derived from PUF_TAX_DETAIL_DEFAULT_TAX_UNIT_OUTPUTS, not hardcoded), so a new collision fails loud instead of silently replacing donor payload; and assignments must consume every selected donor exactly once (bijection asserted at construction). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…rance Confirm round 1 (HOLD, thin-column comparator defect): - The shared gate omits shares for columns under its min-carriers floor, so the production pre-stage collectibles (97 carriers, literal top-100 share 100%) read as pre_share 0.0 and the stage's IMPROVEMENT (100% -> 83.9%, 97 -> 1,135 carriers; review replay receipts) was reported as a worsening from zero — the rescope still raised. Pre and post shares now come from a raw per-column measurement with no thin-column skip (including the combined ST+LT column the gate also checks); over-threshold membership still comes from the production gate on the post frame. - The comparator uses a declared 1e-9 tolerance: weight splitting and summation order move shares by ULPs (the new test's own uniform weights produce 1.0000000000000002). - The replay-shaped test pins the exact 97 -> 1,135 transition passing, a stage-worsened control failing, and sub-tolerance movement reading as noise. Receipts also carry per-column distinct-value counts pre and post — a visibility signal for the share-preserving value-collapse pattern the review named (not gated; the release-side gate owns the artifact-level verdict). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Round 2 pushed for all three required items: (1) raw per-column measurement independent of the gate's thin-column floor — the comparator now sees pre-stage collectibles at its true 100%/97 carriers, so the transfer's improvement (→ 83.9%/1,135) passes; the combined ST+LT column is included in the raw receipts so a combined-column failure cannot drop silently; over-threshold membership still comes from the production gate on the post frame. (2) The exact replay-shaped test: 97 → 1,135 carriers passing, a stage-worsened control failing. (3) Declared 1e-9 worsening tolerance — my own test's uniform weights produced 1.0000000000000002, proving the point. Your High (share-preserving value collapse) is surfaced via per-column pre/post distinct-value counts in the receipts — visibility, not gating, per the required-list scope; the release-side gate owns the artifact-level verdict. Also corrected my PR-body claim per your finding: collectibles was absent from dense-P2's nine because it was thin at export (100 carriers), not because it cleared. |
… distinct counts Confirm round 2 (HOLD, commensurability): - The raw measurement now applies the production gate's population — finite rows with positive weighted |mass| — omitting only the minimum-carrier floor. Round 2's reproduction (one NaN poisoning the raw share to 0.0 while the gate reads 0.952/FAIL) is pinned in the replay test via NaN rows on both sides. - The replay regression now constructs the exact Base-P3 geometry — pre 100%/97 carriers (thin: asserted OMITTED by the production gate), post 0.839/1,135 carriers (asserted to the share and count) — and derives over-threshold membership through the production gate instead of a fabricated result, for the worsened control too. - Distinct-value counts use the same finite/positive-mass carrier mask. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Round 3 pushed for all three findings: (1) the raw measurement applies the production finite/positive-mass population (only the carrier floor omitted) — your NaN reproduction is pinned in the test with NaN rows on both sides; (2) the replay regression now constructs the exact Base-P3 geometry (pre 100%/97 asserted thin-omitted by the production gate; post 0.839/1,135 asserted to share and count) and derives membership through the production gate for both the replay and the worsened control; (3) distinct counts use the same carrier mask. Your round-2 replay table — Base-P3 passing with collectibles over-but-not-worsened — is the intended production outcome. Tail file green locally; full shard rerunning. |
…y distinct-value assertions (#572) The tolerance test's all-zero gate result made it vacuous (removing the tolerance still passed); it now runs against an over-threshold production-gate result and re-asserts stage_worsened_share is False. The replay pins distinct_values pre (1) and post (2) so the receipts' diversity signal is bound as well. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
1 participant
The final increment of the #567/#570 concentration adjudication, on the round-2 replay receipts:
The rescope: the frame check fails a column only when it is over the absolute threshold and the transfer strictly worsened its top-100 share vs the pre-stage frame. Every column's pre/post shares and carrier counts ride the stage manifest as receipts. The artifact-level absolute adjudication stays with the release-side gate (export weights + reviewed per-run register — where collectibles was not among dense-P2's flagged columns). The donor-tail input check is untouched and unconditional.
Review hardenings folded (from #570 round 2, non-blocking list): declared-ownership overlap invariant on the candidate merge (partition derived from
PUF_TAX_DETAIL_DEFAULT_TAX_UNIT_OUTPUTS, new collisions fail loud) and the donor-key bijection assertion.Tests: rescope semantics (inherited-over passes with receipts / stage-worsened-over fails / under-threshold recorded), poisoned-candidate collision raises, dropped-assignment bijection raises. Full tail file green.
Unblocks base-P3 attempt 2, bare — no register file needed.
🤖 Generated with Claude Code