diff --git a/docs/progress.md b/docs/progress.md index 584081a9..1c046f7d 100644 --- a/docs/progress.md +++ b/docs/progress.md @@ -2,7 +2,7 @@ # Development progress -*Snapshot as of 2026-09-06T00:06:00Z (UTC); updated at least daily while the evidence campaign is active. The data behind this page is committed at [`docs/progress/progress.json`](progress/progress.json).* +*Snapshot as of 2026-09-06T10:26Z (UTC); updated at least daily while the evidence campaign is active. The data behind this page is committed at [`docs/progress/progress.json`](progress/progress.json).* Dynamics builds Social Security earnings histories from the PSID, and the corrected covered-earnings series is only as credible as the reading of the source documentation beneath it. Before that series ships, every piece of documentation the construction relies on is independently verified by the three arms below; the design is pre-registered and ratifies only when they complete. (Internally: the evidence campaign behind Amendment 20 (two-arm evidence charter for the covered-earnings correction design), under review in PR #405, open; draft ratified through round 6.2; next step, the A4 evidence freeze, then the C20 ratification chain toward revision 22.) @@ -320,6 +320,7 @@ From the pre-registered timeline ledger, entry 22 (registered 2026-09-03). Dates ## Recent milestones +- **2026-09-06** — Morning after the halted ceremony. Nothing ceremony-related moved overnight and nothing will until three things are ruled on: how the scratch-state battery requirement and candidate-HEAD authentication should coexist (both referees rejected the re-freeze's reconciliation as an unauthorized precommit exception), whether the two substituted sentences are restored or re-authored, and a kit round that deposits durable evidence of scratch cleanliness; the Codex home that runs the ceremony's Astra steps is at its weekly cap until 12 September, so no re-run is possible before then in any case. Two corrections to what the decision package tells Max were settled from bytes: the seven-sealed-document itemisation omits one document (a defect the errata already records as its first entry and that four instruments had re-derived independently without any of them updating the package), and the errata's own Part 2 summary says four decisions change position where its tables and enumeration show nine and eight — its detail is right, its headline is not. The structural question that follows is Max's: whether the package gets a revision folding the errata's twenty-seven defects, or ships as-is with the errata beside it. Graph side, all shadow-only and never pushed: plan revision 12 (two acceptance scopes for A-1, a per-class staged rule, a kernel totality contract, and a live-scope requirement for the closing census) was agreed by the peer gate with no findings and merged; the cohort was frozen at an immutable board coordinate with a 4,262-entry content-hashed manifest (5.55 GB across the seven declared source classes, every digest recomputed), the A4 storage record was filled from existing local settings with gated PSID documentation ingested by identity only because the host's disk is not encrypted, and the H2 stopping rule was frozen before any A-3 run (earliest H2 the evening of 12 September). A-1 accepted at LIVE scope on run 4's head 3bb53ad: an independent refute-first pass reproduced all eleven clause verdicts with nothing refuted and seven report-level qualifications recorded; run 4's third declaration rule partitioned the 29 staged-report copies 10/11/8 exactly as predicted; the plan owner reproduced its 212-test suite. Run 4's delta was accepted on reading and reproduction, not on a second adversarial pass. A-3 pass 1 landed and was accepted as one bounded pass (run 20260906-062442, 19 minutes): 2 post-H harvest events shadowed identity-only with latency recorded, zero discrepancies against bytes, H2 not yet reachable (earliest 2026-09-12). Nothing was copied into the shadow store: the plan's A4 clause requires the custody checks to be recorded per source before any copy, and the recorded table closed at H, so the pass refused every post-H copy and reported the six sources by identity. The plan owner records the post-H custody rows as a dated addendum before pass 2, which then copies the admissible sources and runs the A-2 contract over them; no plan amendment is needed for custody. Two frozen-enumerator defects (a void pattern without a negation guard; a harvest stated in words that only a creation-receipt sweep caught) are recorded for the next revision. Passes continue daily until H2. Purpose stays at 100.00%; Q5 at 63/257; the twelve decisions remain with Max. - **2026-09-05** — Late evening: the Amendment 20 closure ceremony ran to its referee step and stopped there. After the re-frozen candidate C′ (820da19) passed the scratch transition S′ (battery 220/220), the executed-transition receipt (kit gate ready 11 of 11) and a first-add on the ceremony branch, both independent closure referees withheld ratification and wrote no verdict. Both rest on the same two blocking grounds: the re-freeze's test change, which made the battery hold at the scratch state, reads as a precommit index-authentication exception the parent design does not authorize; and the repository holds no independent evidence that the scratch worktree was clean at S′, a gap in the ceremony's own evidence design. Both referees also found the candidate's replacements for the parent's status-meaning and exact-failure-control sentences not proposition-equivalent; one treats that as blocking, the other does not. Closure and the revision-22 repin were therefore not attempted, the ratification pull request stays unmerged, and the next candidate needs a design-level decision on how the scratch-state battery and candidate-HEAD authentication coexist, a kit round that deposits durable cleanliness evidence, and a ruling on the two sentences. Every step is recorded with digests in the ceremony record. Graph side: the A-1/A-2 build was reviewed clause by clause (accepted at fixture scope; the plan's live source model not yet met), a shadow child fixed the review's concrete findings and measured the kernels' totality (123 tests; three inputs that crash rather than fail typed), and plan revision 12 is drafted with a kernel totality contract and is in its peer gate. Purpose stays at 100.00%; Q5 at 63/257; the twelve decisions remain with Max. - **2026-09-05** — Evening: the Amendment 20 A4 freeze was re-frozen after its first candidate failed the pinned battery at the scratch state (a test the freeze itself extended validated the working-tree registry as production whenever ratification was flagged ready, so it could not hold at any scratch state). The re-frozen candidate C′ (820da19, sole parent 86e6994) differs from the first only in that test and its one capture row; the §34 semantic digest is unchanged and was reproduced with the parent commit's own projection code; the candidate battery passed 220/220 and the scratch-state battery is the remaining gate. Scratch transition S′: dispatched to Astra at 18:01 EDT after the re-freeze report showed readiness 31/31 and both batteries 220/220. Read-only review of the candidate's prose found that the freeze removed seven sentences from the parent (the three permitting sentences already known, plus a meaning clause, a blanket prohibition on authority selection before activation, and a rule that a proven residue controls) and added at least twelve law-shaped sentences with no parent counterpart, one of which rules on how the campaign's own board may be read; all reproduced from bytes by a second party and carried to the two closure referees as questions, not rulings. Kit round r8b installed under the two-engine gate: referees now run a whole-document sentence diff in both directions with stated splitter parameters, source-byte probes, and a guard that the withheld-ranks sentence belongs to the missing-reason arm's occurrence census, not the purpose ledger (which stays at 100.00%). Two figures in the decision package were corrected: the MEASURED-family population is 79 occurrence lines, not 78 (one bracket opens with an en dash that a same-line scan cannot see; four parties had run the same scan), and the tag-id unit is 126 under its author's own rule (122 numbered items by five independent methods plus four verified §4/§5 items), not 127. The errata's own 'largest unswept surface', a truncation ending mid-clause somewhere in the decision package's 185 quoted spans, was swept exhaustively: every span ruled (167 complete, 9 truncated at a boundary, 8 altered, 1 not locatable) and zero mid-clause truncations, with the population re-derived blind and all nine control loci recovered. The MEASURED stratum of the decision package was then opened source by source with no wording trigger (78 of 79 loci, 35 of 35 cited files): the source-fidelity defect rate is 11 of 78 (14.1%), so the earlier 13.9% floor is now a bound, and the wording trigger it replaced had found only 3 of those 11; the union of loci carrying a defect on any axis is 15 of 79 (19.0%), stated as a union of two passes rather than one measurement. Both surfaces the errata named as unswept are closed. Graph side: the Plan A shadow plan reached revision 11 through three peer gates and merged to master; its first build slice is running on a Claude lane (Opus) since 17:08 EDT with its first commit (7515f9b, the q5graph package and its test suite) at 17:52, on a Claude lane under the agreed resource policy with no escalation. Q5 stands at 63/257; the twelve decisions remain with Max. - **2026-09-05** — The purpose census is COMPLETE: 20,815 of 20,815 prompts adopted (100.00%) at the terminal fold, ledger v1.97 (ebe3d2b9…). Blocks 7–10 of the final tranche folded through the day (v1.93–v1.97), every audit CALIBRATED; block 8 was folded in two parts after an orchestrator scope error (the brief was cut from the per-section table, leaving 64 ranks unadjudicated until an 8b completion pass) and block 9 closed the 2021 questionnaire entirely. A standing correction now requires every lane to recompute each digest it emits from the named file and run its own 64-hex self-check, after a 63-character digest was copied forward from a dispatch brief. A compile-only decision package (twelve decisions the package orders by how much work each blocks, against 63 sealed, 17 held and 1 withdrawn of the 81 annotated documents, although an errata compile finds that order does not consistently follow its own criterion) is staged for ruling; it lists five items it says the commissioning brief missed, the sharpest being that the worktree the lanes run in is two ratified design revisions behind master (19 vs 21), although the slot-catalog section the top decision turns on is byte-identical at both. Two figures already in front of the ruling were corrected: the atomisation coverage gap is 331 of 447 enumerated Social Security, pension, annuity and retirement items in no atom (74.05%, not the 72.7% the census printed and its replication reproduced without recomputing), and the corpus's page and frame digests are reproducible under the pinned Poppler pdftotext 26.04.0, which the design, the inventory code and the stage-1 replay artifact all name and the code enforces (cross-host determinism untested) (a Chronicle cross-check with pypdf 6.11.0 reproduced 0 of 101 pages, the pinned recipe 101 of 101); the final-tranche pass reports did not restate the build until the block-10 audit did. Q5 stands at 63/257, which is 77.8% of the 81 annotated questionnaires; the remaining 176 files in the fixed domain are setups, codebooks and raw data, not instruments. The open questions are recorded for ruling; the compiler ruled on none of them. diff --git a/docs/progress/progress.json b/docs/progress/progress.json index 8506f277..c7c65ec2 100644 --- a/docs/progress/progress.json +++ b/docs/progress/progress.json @@ -1,7 +1,7 @@ { "schema_version": 1, "description": "Data behind docs/progress.md (the development-progress dashboard). Update this file and run `python scripts/build_progress_page.py`; never edit docs/progress.md by hand.", - "as_of": "2026-09-06T00:06:00Z", + "as_of": "2026-09-06T10:26Z", "framework": { "amendment": "Amendment 20 (two-arm evidence charter for the covered-earnings correction design)", "pr": 405, @@ -87,6 +87,10 @@ ] }, "milestones": [ + { + "date": "2026-09-06", + "text": "Morning after the halted ceremony. Nothing ceremony-related moved overnight and nothing will until three things are ruled on: how the scratch-state battery requirement and candidate-HEAD authentication should coexist (both referees rejected the re-freeze's reconciliation as an unauthorized precommit exception), whether the two substituted sentences are restored or re-authored, and a kit round that deposits durable evidence of scratch cleanliness; the Codex home that runs the ceremony's Astra steps is at its weekly cap until 12 September, so no re-run is possible before then in any case. Two corrections to what the decision package tells Max were settled from bytes: the seven-sealed-document itemisation omits one document (a defect the errata already records as its first entry and that four instruments had re-derived independently without any of them updating the package), and the errata's own Part 2 summary says four decisions change position where its tables and enumeration show nine and eight — its detail is right, its headline is not. The structural question that follows is Max's: whether the package gets a revision folding the errata's twenty-seven defects, or ships as-is with the errata beside it. Graph side, all shadow-only and never pushed: plan revision 12 (two acceptance scopes for A-1, a per-class staged rule, a kernel totality contract, and a live-scope requirement for the closing census) was agreed by the peer gate with no findings and merged; the cohort was frozen at an immutable board coordinate with a 4,262-entry content-hashed manifest (5.55 GB across the seven declared source classes, every digest recomputed), the A4 storage record was filled from existing local settings with gated PSID documentation ingested by identity only because the host's disk is not encrypted, and the H2 stopping rule was frozen before any A-3 run (earliest H2 the evening of 12 September). A-1 accepted at LIVE scope on run 4's head 3bb53ad: an independent refute-first pass reproduced all eleven clause verdicts with nothing refuted and seven report-level qualifications recorded; run 4's third declaration rule partitioned the 29 staged-report copies 10/11/8 exactly as predicted; the plan owner reproduced its 212-test suite. Run 4's delta was accepted on reading and reproduction, not on a second adversarial pass. A-3 pass 1 landed and was accepted as one bounded pass (run 20260906-062442, 19 minutes): 2 post-H harvest events shadowed identity-only with latency recorded, zero discrepancies against bytes, H2 not yet reachable (earliest 2026-09-12). Nothing was copied into the shadow store: the plan's A4 clause requires the custody checks to be recorded per source before any copy, and the recorded table closed at H, so the pass refused every post-H copy and reported the six sources by identity. The plan owner records the post-H custody rows as a dated addendum before pass 2, which then copies the admissible sources and runs the A-2 contract over them; no plan amendment is needed for custody. Two frozen-enumerator defects (a void pattern without a negation guard; a harvest stated in words that only a creation-receipt sweep caught) are recorded for the next revision. Passes continue daily until H2. Purpose stays at 100.00%; Q5 at 63/257; the twelve decisions remain with Max." + }, { "date": "2026-09-05", "text": "Late evening: the Amendment 20 closure ceremony ran to its referee step and stopped there. After the re-frozen candidate C′ (820da19) passed the scratch transition S′ (battery 220/220), the executed-transition receipt (kit gate ready 11 of 11) and a first-add on the ceremony branch, both independent closure referees withheld ratification and wrote no verdict. Both rest on the same two blocking grounds: the re-freeze's test change, which made the battery hold at the scratch state, reads as a precommit index-authentication exception the parent design does not authorize; and the repository holds no independent evidence that the scratch worktree was clean at S′, a gap in the ceremony's own evidence design. Both referees also found the candidate's replacements for the parent's status-meaning and exact-failure-control sentences not proposition-equivalent; one treats that as blocking, the other does not. Closure and the revision-22 repin were therefore not attempted, the ratification pull request stays unmerged, and the next candidate needs a design-level decision on how the scratch-state battery and candidate-HEAD authentication coexist, a kit round that deposits durable cleanliness evidence, and a ruling on the two sentences. Every step is recorded with digests in the ceremony record. Graph side: the A-1/A-2 build was reviewed clause by clause (accepted at fixture scope; the plan's live source model not yet met), a shadow child fixed the review's concrete findings and measured the kernels' totality (123 tests; three inputs that crash rather than fail typed), and plan revision 12 is drafted with a kernel totality contract and is in its peer gate. Purpose stays at 100.00%; Q5 at 63/257; the twelve decisions remain with Max."