Spec-Actualization Campaign v0.1 — mark, verify, and de-drift the whole spec tree

0. Mandate (owner's words, 2026-07-24, recorded verbatim)

• 03«актуализировать ВСЕ спецификации … Это чудовищная огромная работа. Именно поэтому я и готовлю scaffold для нее, чтобы не сбиться в ходе обхода настолько большого количества документов. Нам нужно разметить корпус фактов, которые дальше нужно будет проверять.» Work may take a month; that is accepted; quality over speed.
• Paragraph-level exhaustiveness is the point, not an option: «это in-verbatim контроль того, что мы прошли всё, каждую строчку. LLM очень любит упрощать … нужен алгоритмический надсмотрщик».
• No fractality for this campaign. «Я хочу чтобы Fable сделала максимум высокоуровневых задач (анализ и разметку спецификации и тп)» — outputs: (a) a corpus of coding tasks for Opus, (b) a corpus of spec-improvement tasks executed by whatever model the budget allows (Fable if it stretches, Opus otherwise). This deliberately overrides the standing delegation-first default for the duration of the campaign — owner decision; do not "optimize" it back.
• Stitching is non-linear: reworking B may reopen A and vice versa — plan it as a multi-pass fixpoint, not a single sweep.
• Crash-safety: any session may die (budget, power); the next session must resume from one obvious file with at most one step lost.
• Repeatability: re-runs at ~monthly cadence must cost O(delta), not O(corpus).
• Wave 1 = the host spec/ tree only; packages/ waves later; the fractality specspace excluded until the owner says otherwise.

1. Baseline (verified at authoring time, 2026-07-24)

• 04Host spec/: 91 md files, 26 699 lines. Authored packages (no vendored copies): world 154 + ai-native 140 files (~30k lines) — wave 2; fractality 700 files — out of scope.
• Free-form **Status:** lines to convert mechanically: ~55.
• specmap: index live; 34 gated orphans in vibe-spec (pre-existing).
• Progress Control: does not exist yet — no crate, no vibe progress, no campaigns/ zone, no dashboard. <status appears nowhere in the tree except PROP-043's own dogfood markers.
• Existing inline grammars that must not collide: @spec:// (~17 uses), #use/#embed/#source, <!-- REVIEW: -->.

2. Executors and the budget law

06Rules 1–4 of the repository bind every executor. Worker output is never credited; commits are human-authored surface; non-routine red lines stop for the owner no matter who is executing.

3. Campaign zone layout

campaigns/progress-2026-08/        # id fixed at Phase A close
  baseline.json                    # inter-campaign contract (PROP-043 §7.3)
  deferrals.md                     # open tails at close-out; next run drains it
  harvest/                         # doc cards (templates/harvest-card.md)
  tasks/                           # DRIFT-NNN.md / SPEC-NNN.md + INDEX.md
  run/                             # EPHEMERAL: journal.jsonl · state/*.json · RESUME.md · mirror/

08Excluded from markup scope, packaging, and registries (PROP-043 §7.4). Committed at batch boundaries — journal in the same commit as the edits it describes; fan-out via cargo xtask mirror at phase checkpoints. run/ of a closed campaign may be archived or deleted; the other four entries persist.

4. Resume protocol (crash-safety law)

1. 09Step = unit of atomicity: mark-file · verify-unit · close-obligation · execute-task. Journal writes step-start (intent, actor) before work and step-done (result ref) after; JSONL, append-only, torn tail discarded.
2. Recovery rule: step closed in journal ⇒ its edits stand; step open ⇒ git restore its files and redo the step. Steps are idempotent by construction. Maximum loss on any crash = one step.
3. RESUME.md is generated (vibe progress resume) after every step-done: where we are · unresolved steps with literal recovery commands · next steps · phase rules pointer · dashboard command. Every session of this campaign starts by reading it and ends by closing (not starting) a step.
4. Claims and staleness: journal actors (fable, opus:DRIFT-012); an in-progress task with no journal events past the threshold is returned to queued as stale by resume.
5. git = second echelon: batch commits make the worst disk-loss cost one batch, never the campaign.

4.5 Safe stop — where wave 1 could halt losing nothing (recorded retrospectively)

10Added 2026-07-31 under the owner's bring-into-line ruling: flow:campaign-plans' ##ANY-PHASE-BOUNDARY-IS-A-SAFE-STOP asks every plan to say where it can be put down, and this plan said it nowhere. Written after the campaign closed, so it is a record of where the stops actually were — not a promise made in advance.

11This campaign ran on a finer grain than the law asks for. The flow's unit is the phase boundary; §4's unit is the step — mark-file, verify-unit, close-obligation, execute-task — with step-start journalled before the work and step-done after, so the maximum loss on any crash is one step. That is strictly stronger, and it is why no session of this campaign ever needed the phase-level guarantee to recover. Both grains held; both are stated here because a stranger reading only §4 would think the campaign could be stopped anywhere, and a stranger reading only the flow would think it could be stopped only six times.

12What a stop at each boundary would have left, so «losing nothing» is a statement and not a slogan:

14The one boundary where the floor was not green, recorded rather than smoothed. The Phase C close self-check went red on cli_pkg_cycle::install_from_git_registry. Root-caused and proven in-session: a ~/.vibe/registry.toml had appeared on the machine that day, and the test isolates VIBE_REGISTRY_CACHE but not the settings chokepoint, so the global registries merged into the "hermetic" resolver and minted a second cache bucket; with VIBE_SETTINGS pointed at an empty directory the same test passes. The campaign's own gate (progress check) stayed 0 and the phase's commits were docs-only and unrelated. Ledgered as F-055 and fixed in Phase E — after which «the floor ends the phase green with no VIBE_SETTINGS override», F-055 genuinely fixed rather than worked around. ##A-PHASE-THAT-LEAVES-THE-FLOOR-RED-IS-STILL-OPEN is the rule this brushes, and the only reading under which C could close is the one the ledger states: the red was outside the phase's own diff, and it was proven so rather than assumed.

15Two things that are NOT safe stops, both learned at cost:

• 16A batch whose files are written and whose journal step is still open. §4's recovery rule is not advisory — step open ⇒ git restore its files and redo the step. Steps are idempotent by construction precisely so this is cheap.
• Delegated work committed on a filled-in task journal rather than on the completion notification. Executors write the ledger as they go, so committing on it captures an intermediate state; doing so once left the tree conform-red for twenty minutes. Now a WAL Constraint, together with its sibling: a gate never seen to go red is not known to work.

5. Phases

17Each phase: entry condition → steps → exit gate (+ prediction, per the campaign-plan discipline). Every session inside any phase obeys §4.

6. Recurrence (the monthly re-run)

48vibe progress rescan --baseline <prev>/baseline.json → new / suspect / carried-forward lists → mini-B on new+changed → re-verify suspect (+ random control of carried-forward) → mini-D on new findings → tasks → new baseline. Cost O(delta). Between runs, the maintenance discipline (PROP-043 §10) and a health-audit category ("markers vs reality") keep the delta small. This plan is the standing playbook for those runs; each run appends its own LOG entry and rewrites baseline.json.

7. Dashboard contract

49Reads run/state/*.json only (never Markdown, never computes). Zero npm dependencies. Screens: Resume (open steps red, next steps, freshness plaque), Overview (phase lane, counters), Corpus (tree colored by rollup, five views + audience filter), Stitching (ledger table + open-obligations-per-wave chart; non-falling pairs highlighted as escalations), Tasks (both queues, statuses, claim owners). Localhost, read-only, no auth.

8. Predictions (falsifiable, campaign-wide)

1. 50Wave-1 full weave fits ≤2 shards of a 1M window (digest fits trivially).
2. The exhaustive counter catches ≥1 genuinely skipped paragraph that review alone would have missed.
3. ≥60 % of **Status:** IMPLEMENTED/SHIPPED claims confirm without drift; ≤10 % of all units end unverifiable.
4. Stitching converges in ≤3 waves with ≤2 owner escalations.
5. ≥80 % of DRIFT tasks land without a returned round-trip.
6. The month budget holds: A ≈ days, B ≈ 1–1.5 weeks, C ≈ 1 week, D ≈ 3–5 days, E ≈ open-ended by queue, F+G ≈ 1 week overlapping E.

8.5 Non-goals (named retrospectively)

51Added 2026-07-31 under the owner's bring-into-line ruling: flow:campaign-plans' ##NON-GOALS-ARE-NAMED-SO-THEY-STAY-VISIBLE asks a plan to name what it deliberately does not do, and this plan named it nowhere. Every line below was a real boundary the campaign held — most of them owner rulings recorded in §9 — but none was written down as a non-goal at authoring time, and the two that ended up costing something (the judgment axis, the doc trees) are exactly the two that were never named. Reason and disposition on each, per ##EVERY-NON-GOAL-CARRIES-A-REASON-AND-A-DISPOSITION.

• 52Does NOT extend to `packages/.** *Reason:* one corpus at a time; the method had to be proven before it was scaled. *Disposition:* **wave 2**, [PACKAGES-ACTUALIZATION-CAMPAIGN-v0.1.xml`](PACKAGES-ACTUALIZATION-CAMPAIGN-v0.1.xml).
• Does NOT touch `vibevm/vibepacks/org.vibevm.fractality/`. Reason: its own specspace, own boot contract, own WAL; the mandate excluded it in as many words («the fractality specspace excluded until the owner says otherwise»). Disposition: held by the owner. *(Recorded 2026-07-31, because it turned out to matter: that specspace is a second adopter* of several flows this programme measures, and a perimeter blind to it reads adoption as absence.)
• Does NOT mark or verify the relocated legacy directories. Reason: Phase L moved terraforms / research / neworder / discipline to root legacy-spec/ as historical records, not living contracts. Disposition: rejected outright — and reinforced by the owner's ruling of 2026-07-31 that legacy-spec/** is not evidence of practice in either direction.
• Does NOT mark generated artifacts. vibevm/vibespecs/boot/STATIC.xml, vibevm/vibespecs/boot/INDEX.md and vibevm/vibespecs/WAL.xml in session form. Reason: markup written into a generated file dies at the next vibe install or wind-down. Disposition: rejected; owner rulings 2026-07-24. The authored boot snippets (00-core, 90-user) stay observed via vibevm/vibespecs/boot/[0-9]*.md.
• Does NOT edit vibevm/vibespecs/boot/90-user.xml. Reason: user-owned; 00-core's NOTOUCH-90-USER forbids it to every session. Disposition: deferred to the owner, not to a campaign — F-063's half was handed over in full rather than edited, and closed 2026-07-26 when the owner lifted the bar.
• Does NOT decide what should happen to a fact. Phase B marked what 4 917 facts are; nothing in any phase asked what should be done about them. Reason: none was given at authoring time — this is the non-goal the campaign held without ever choosing it, which is why §5-F's three views came out empty (freeze/plan 0, action="rework" 0, stage="idea" 0). Disposition: deferred to wave 2 as amendment A3(i), the judgment-marking pass.
• Does NOT write the two documentation trees. User Guide and Package Author Guide. Reason: Phase G's definition demands harvest cards and captured runs as input, and Phase C skipped the step that produces them. Disposition: deferred to wave 2 as amendment A3(ii) — the Package Author Guide in particular, since packages/ is the corpus it documents.
• Does NOT use fractality. Reason: the owner's decision, verbatim in §0, deliberately overriding the standing delegation-first default for this campaign's duration — «Я хочу чтобы Fable сделала максимум высокоуровневых задач». Disposition: held by the owner; do not "optimize" it back.

8.6 Risks and fallbacks (recorded retrospectively)

53Added 2026-07-31 under the owner's bring-into-line ruling: flow:campaign-plans' ##EVERY-RISK-CARRIES-A-DETECTION-SIGNAL-AND-A-PLAN-B asks every plan to name its risks with a detection signal and a plan B, and this plan named none. A risk register written after the campaign is a list of what happened, not of what was feared, and it is labelled as such: each row below fired, and each says what detected it — including the four whose honest answer is «nothing detected it; the next phase tripped over it». The fired-and-undetected column is the useful part of a retrospective register, because it is the part wave 2 turned into amendments.

• 54R1 — a verdict whose evidence is another spec document. Fired: F-063. Five token-precedence anchors in PROP-002 were sealed confirmed on the evidence «token loader 3-source order matches 90-user boot facts» — another spec file, carrying the identical error, in a security-relevant place. Detection that existed: none. Phase C's exit gate checked only that every marker carried a verdict, never what the verdict rested on. Fallback taken: wave 2's amendment A2 — every verdict names which source class it rests on, and one resting on the package's own artifacts alone is counted as self-referential rather than as independent confirmation.
• R2 — the campaign's own corrections introduce drift. Fired: F-065. Phase D authored a Shipped: line claiming a Baseline::store that had never been built, and the verification pass then confirmed the row on it. Detection that existed: none — nothing predicted that stitching could add false claims. Fallback taken: wave 2's prediction 6, which makes «zero new false claims, zero of them confirmed» falsifiable, with wave 1's answer on record as 1 and 1.
• R3 — a phase whose exit gate does not check its own steps. Fired: Phase C listed «harvest cards written while knowledge is hot» among its steps and gated only on «100 % of markers carry verdicts». The step was skipped, it cost nothing at the time, and Phase G arrived to consume an empty directory. Detection: the downstream phase, three days late. Fallback taken: wave 2's amendment A1 — every exit gate enumerates that phase's own steps.
• R4 — a prediction no step forces you to run. Fired: prediction 1. No step of the plan required a weave, so the claim sat untested for the whole campaign and was measured at close-out purely to fill the report row. Plan B taken: wave 2's amendment A5 — every prediction names the step that tests it, or says outright that it is scored at close-out on purpose.
• R5 — a state projection nothing refreshes drifts silently. Fired: tasks.json sat 18 tasks stale and the dashboard read five tasks for a week; findings.json survived only because it was maintained by hand all campaign. Detection: none — a projection nothing refreshes and nothing checks has no signal by construction. Fallback: the numbers of record are the ones a command prints, never the ones a projection holds.
• R6 — a gate never seen to go red is not known to work. Fired: twice, favourably — two executors ran positive controls before trusting a green result, and were right to. Fallback: now a standing WAL Constraint.
• R7 — session death, budget exhaustion, power loss mid-corpus. The one risk this plan did carry a fallback for, in §4, and it held: journal step-start / step-done, torn tail discarded, maximum loss one step. Fired: a live power cut during Phase A, which additionally exposed a missing fsync-before-rename in write_atomic — fixed with a tolerant cache load plus tests. Verdict: the crash-safety law is the one part of this plan that was written in advance and needed no amendment.
• R8 — an un-isolated test reaches real user state. Fired: F-057, filed as a stray cache directory and found to be a credential-precedence leg — vibe search resolves a GitHub token through four legs and the fourth is a file in the settings directory the test could not reach, so an un-isolated run attached the real publish token to its outbound request. Detection: accidental, on the third pass — F-055, F-056 and F-057 were one forgotten discipline caught three times by accident. Measured, not inferred: against a loopback listener recording header names and lengths only, 47 bytes of Authorization before the fix, none after; exactly one test reaches that path and it points api_base at 127.0.0.1, so the investigation closed with no rotation needed. Fallback taken: DRIFT-020 converts the discipline into a gate.

9. LOG (execution ledger — append per batch/wave/phase)

• 552026-07-24 · Phase A CLOSED, exit gate green in full. Commits 9446a2c b1276c3 edd487b (document package) · 8b18152 (core + adapter) · 38855c0 (campaign zone + dashboard) · ac97f26 (pilot + ratification + §3.8 amendment) · 9a69b6f (WAL). PROP-043 RATIFIED in session. Pilot: 3 genres, 46/46 paragraphs marked, one real drift caught (design/README index was incomplete — fixed). Predictions: "pilot exposes placement ambiguities" CONFIRMED (the preamble-less-H1 amendment); bonus finding outside any prediction — a live power cut exposed missing fsync-before-rename in write_atomic; fixed with tolerant cache load + tests. Floor (self-check) green; progress-core gated in conform.
• 2026-07-24 · B0 default ruled (Fable, owner-visible in session): the converted **Status:** lines are NOT deleted — the document marker is ADDED and the legacy line stays (its dates and prose are history; deduplication is Phase D material). Phase B makes no semantic edits, per its own law.
• 2026-07-24 · Phase B OPENED; B0 landed. progress.toml (wave-1 scope spec/**/*.md) committed as the campaign's first journal step (8d5ccc8); scan on the narrowed scope: 97 files (§1 counted 91 — the six newcomers are the progress-control documents authored after the baseline froze), 3 684 paragraphs, 46 pre-marked by the pilot. B0 converted 73 legacy status lines (not ~55 — the §1 estimate missed the **status: / **Status.** variants) into document markers as one reviewed diff: 73 files, +73 marker lines, 0 deletions (the B0 ruling above held — legacy lines stay). check clean; markers 68 → 141. First stitching input recorded in passing: 12 ledger findings (F-001…F-012, run/state/findings.json) — status lines contradicted by executed plans or shipped crates (SELF-SUFFICIENCY / SETTINGS-HOME / CONVERT / CULTURAL-EXTRACTION plans; the PROP-019/024/030/036/037/040 family; two missing superseded notices: TREE-TUI v0.1, PROP-026).
• Scope question for the owner (found in B0, parked): vibevm/vibespecs/boot/STATIC.xml and vibevm/vibespecs/boot/INDEX.md are generated by vibe («do not edit») yet sit inside the wave-1 scope spec/**/*.md. Markup written into them dies on the next vibe install regeneration. Options: exclude generated boot artifacts from progress.toml, or carry their markers in the generators. Until ruled, B1 batches skip both files — which keeps check --exhaustive red on them, so the Phase B exit gate needs this ruling.
• 2026-07-24 · B1 (vibevm/vibespecs/common, paragraph grammar) landed — then the grain changed. All 12 vibevm/vibespecs/common files marked paragraph-exhaustively (389 paragraphs; PROP-000 got its document marker + the missing {#invariants} anchor; open-questions sections marked spec/work), check clean, commit 91274c8. Mid-batch a real power cut killed the session; §4 recovery worked as designed — journal showed one open step (b1-prop-000), rescan verified its edits clean, the step closed without redo. In passing: PROP-000 §3 still describes the pre-2026-07-12 proprietary-EULA posture (F-014-to-be), PROP-018's MVP is implemented while its line says proposed (F-013-to-be), and the cache does not prune records that fall out of scope after progress.toml narrowing — corpus.json carries 497 entries vs 97 observed, dashboard counters inflated (DRIFT-001-to-be).
• 2026-07-24 · FACT-GRAIN DIRECTIVE (owner, in session, verbatim): «Каждый элемент такого списка по сути является самостоятельным фактом, и его тоже нужно размечать. В том числе, inline факты перечисленные внутри текста … Я хочу чтобы ты для всех текстов сделала полное деконструирование всех фактов, имеющихся в системе, совершенно всех текстов. … если какой-то абзац можно переписать в виде нумерованного или ненумерованного списка фактов и каждому из них поставить в соответствие разметку статуса — нужно сделать это. В идеале почти все тексты превращаются в какие-то списки … Это означает ЗНАЧИТЕЛЬНОЕ УСЛОЖНЕНИЕ кода Системы, инструкций Системы и рост количества фактов … Уже проиндексированное и размеченное нужно переиндексировать и переразметить». Clarifications the same session: list-item markers go at the item's start or end, @ or XML notation both; table cells are marked the same way. Ratified into PROP-043 as the fact amendment (§3.8 items 4–6, §3.9, §8) — the campaign granularity drops from paragraph to fact; deconstruction of multi-fact prose into lists is now part of the markup pass (sense-preserving, wording-preserving; semantic edits still forbidden); everything already marked (pilot + B0 doc markers + B1 common) is re-marked under the fact grammar.
• 2026-07-24 · Fact-anchor addenda (owner, same session): (1) list-item markers sit at the item's start or end, @ or XML notation both; (2) «таблицы тоже нужно в ячейках размечать аналогичным способом» — table body cells are units, marked in-cell; (3) list items get hashtag addressing — «Элементам списков стоит придумать адресацию с помощью хэштегов … формат типа 1. #RULE-001 Текст правила @status:freeze/done», refined to ##RULE-001 (double hash) «чтобы их отличать» from the #use-family directives; (4) «каждый абзац, каждый элемент списка … который имеет статус, нужно разметить с помощью якоря» — the anchored-when-marked law: a marked unit without a ##<ID> anchor is a check error. All four ratified into PROP-043 §3.8 / §8.
• 2026-07-24 · Fact-grain scanner LANDED; fact grammar re-piloted. PROP-043 amendment ratified and committed (cd2688f); the scanner (b67fa97): list items at every nesting level, lead lines, table body cells as countable units, ##<ID> fact anchors in one id space with heading anchors, duplicate detection, the anchored-when-marked error; 31 tests green, cache schema 2 (the rebuild also flushed the 400 out-of-scope records — DRIFT-001 still files the general prune defect for Opus, 910d545). Scale shift measured: wave-1 = 3 684 paragraphs → 8 219 facts; check currently carries 435 expected MissingAnchor errors on the pre-amendment markup (pilot + B1 common) — they burn down as files are re-marked. Re-pilot (6714876): PROP-029 deconstructed 9 paragraphs → 30 anchored facts, 0 unmarked, 0 issues — the demo the owner reviewed in session. Ledger: F-013 (PROP-018 MVP implemented vs proposed), F-014 (PROP-000 §3 license text predates the 2026-07-12 UPL-1.0 relicense).
• RESOLVED review point (owner, 2026-07-24) — anchor naming convention: the re-pilot mixed ##UPPER-SLUG for normative facts (ADDR-LAW, RULE-style) with ##kebab for service units (status-line, carriers-lead). Keep both registers, or fix one? Ruling: both registers stay — UPPER = normative fact, kebab = service unit; the register itself carries the normativity signal. Recorded as a decision at PROP-043 §3.8.
• RESOLVED review point (owner, 2026-07-24) — generated files in scope: vibevm/vibespecs/boot/STATIC.xml + INDEX.md are vibe-generated («do not edit») yet inside the wave-1 globs; markup written there dies on regeneration. Exclude them from progress.toml, or carry their markers in the generators? Ruling: exclude from scope. §4 of PROP-043 is include-only by design, so the exclusion is expressed by include enumeration (vibevm/vibespecs/boot/[0-9]*.md admits the authored snippets, skips the generated pair); cache rebuilt from scratch (the DRIFT-001 no-prune defect makes a stale-record purge otherwise impossible). Scope: 97 → 95 files, 8 219 → 7 872 facts. This also closes the B0-parked scope question above.
• 2026-07-24 · B1f LANDED — vibevm/vibespecs/common is fact-grain clean; the two review points RESOLVED (above). The scanner-handover step was closed retroactively (b67fa97 verified committed; RESUME had prescribed a redo of finished work). All 11 remaining vibevm/vibespecs/common files re-marked under the fact grammar: 386 paragraph-grain units → 979 anchored facts (batch commits 83bed35 / 4aed13f / d639bcf; the batch-1 message overstates its own counts — 109 units → 296 facts is the true figure, corrected here, history left unrewritten). Cluster total: 1 009 facts, 0 unmarked, 0 issues — cluster MissingAnchor 386 → 0; the wave's residue is 40 expected errors in the two pilot files (SHRINK-PLAN 28, design/README 12), owned by their B2+ batches. Grammar traps found and recorded: blockquote units cannot carry ## anchors (ledgered F-015; two units re-formed — a bold paragraph, a fenced template); a wrapped prose line whose continuation opens with + parses as a phantom list item (two fixed in PROP-019). Tasks queued for Opus: DRIFT-002 (parse.rs 809 lines > 600 budget — the standing floor is RED on this single new conform finding until it lands; the B1f batch commits were made against that known, ledgered red) and DRIFT-003 (campaign.json phase hardcoded "A", dashboard/RESUME render a stale phase). Prediction check in passing: the §5-B "unmarked counter catches what review misses" prediction keeps confirming — the counter caught both phantom-item wraps instantly.
• 2026-07-24 · DRIFT-002 executed and landed — the floor is green again. First DRIFT task through the full loop: Opus executed the parse.rs split exactly per the task file (six parse/ modules, max 261 lines; motion, not rewrite), Fable reviewed and accepted (spot-diff verbatim, differential oracle — corpus identical modulo timestamp, 31 tests green, conform 0 new, self-check all green with the real exit code). One accepted deviation: per-file //! module docs, the crate's own convention. B2 opened in parallel the same evening: templates + modules README + PROP-042/025 marked (F-016 structural index drift, F-017 code-ahead-of-spec vibe aiui scrollbar). The §5-E prediction ("≥80 % of DRIFT tasks land without a returned round-trip") starts 1/1.
• 2026-07-24 · DRIFT-003 landed — the phase lane is honest; B2 runs through batch 5. Second DRIFT task through the loop, again no returned round-trip (§5-E prediction 2/2): the phase is now an append-only journal event ({"kind":"phase","value":"B"}, last wins, absent ⇒ "A"), derived by the adapter — never compiled in, never parsed from Markdown; read_journal distinguishes a torn tail (incomplete JSON, stops) from an unknown kind (complete JSON, skipped) with both laws test-pinned; the live journal is backfilled and campaign.json/RESUME render B. B2 batches 3–5 in parallel: PROP-026 (superseded-in-topology arc split impl/done vs spec/done per fact), PROP-021/023 (bridge pair), PROP-020/022 (F-018: hooks ship while the line says proposed), PROP-041 (owner-minted per-REQ {#anchor}s reused verbatim as ## ids — one name, two notations; two heading-vs-REQ same-name collisions surfaced by the shared id space, F-019 for the stale DRAFT line). Ledger: 19 findings. B2 stands at 12/35 files.
• RESOLVED review point (owner, 2026-07-24, in session) — the WAL and the boot files: «vibevm/vibespecs/WAL.xml не должен участвовать в проверке, это генерирующиеся артефакты; также проверь про файлы внутри vibevm/vibespecs/boot». Ruling applied: vibevm/vibespecs/WAL.xml is out of scope — the checkpoint is rewritten wholesale at every session end, so markup cannot live in it (the same mortality argument as the generated boot pair). The boot check reported back: STATIC.md/INDEX.md are generated and were already excluded by the morning ruling; 00-core.xml/90-user.xml are authored, user-owned, never written by vibe (PROP-000 invariant 3) — they stay observed. Scope: 95 → 94 files; the spec/*.md include is gone (WAL was its only match), so a future top-level spec doc must be added to progress.toml explicitly.
• 2026-07-24 · Owner directive — the coder-tier engine version. The §2 coder tier ("Opus") runs on the owner-designated engine claude-opus-5 from the next session on. Verified: the id is live (one-shot CLI probe answered); the session-alias default resolves to the previous engine, so two pins are installed — the machine-local subagent-model env pin (.claude/settings.local.json, blanket) and the committed selective agent type (.claude/agents/opus5.md) — both effective from the next session (agent-type registration and settings-env injection are session-start events; verified empirically mid-session). Tasks DRIFT-002…005 of this date executed on the session-alias engine before the directive; all passed review with no returned round-trip.
• 2026-07-24 · DRIFT-005 landed — F-022 closed end to end; the DRIFT loop stands 5/5 no-return. The owner's fact-links commission is complete across all three layers: contract (PROP-014 §2.1 + PROP-035 §5/§7.3 with the reviewed heading-repeat precision), engine (DRIFT-004, core v0.8.0 mdspec fact units — all language families inherit through the shared engine at their next minting), and host compiler (DRIFT-005, vibe-spec: NodeKind::Fact IR leaves, per-fact override under :add, CompileError::DuplicateId merged-view gate, fact-addressed #embed). Code can now cite spec://…#<FACT-ID> per statement, and the §6 evidence join gains the campaign grain for Phase C. Session-end: the coder-tier engine pin (claude-opus-5) binds from the next session.
• 2026-07-25 · B2 modules sweep — 18 files in batches 8–18; B2 at 32/35. PROP-015/034/027/036/030/011/012/010/040/038/008/001/009/ 017/043/035/037/007 marked at fact grain (commits b27336ae…1e7dff01), ~1 540 units → ~1 770 anchored facts; every file 0 unmarked / 0 issues. Grammar precedents set: the Decision-paragraph idiom, ##req-*/##design-* lines, ##self-uri, checkbox anchors before [x], @impl/plan for unexecuted phase plans, superseded-arc spec/done-vs-impl/done, em-dash cells count. Ledger +7: F-023 (dangling PROP-043 launcher ref) and the stale-header family F-024…F-029 — one Phase C/D sweep fixes all. GitVerse SSH down all session (verified clean ancestor via HTTPS; plain re-fan on recovery, never --force); GitHub carries everything.
• RESOLVED scope ruling (owner, 2026-07-25, in session): «я хочу исключить из проверки vibevm/vibespecs/terraforms, vibevm/vibespecs/research, spec/neworder. Это те вещи, которые мы делали в качестве рефакторингов и исследований давным-давно». Ruling applied: the three subtrees leave the include enumeration in progress.toml — long-executed plans and studies are historical records, not living contracts. Scope: 94 → 59 files, 8 589 → 4 889 facts; the SHRINK-PLAN pilot (28 expected errors) leaves with terraforms, so the expected check residue drops to 12 (design/README, burns at its re-mark). The DRIFT-001 cache prune dropped the out-of-scope records cleanly. The campaign plan itself is now out of scope — its LOG stays the process record, unmarked.
• Next step: finish the B2 tail — PROP-005 → PROP-003 → PROP-002 (modules to 35/35), then vibevm/vibespecs/design (incl. the README re-mark burning the last 12 expected errors), vibevm/vibespecs/boot authored pair (additive markers only — user-owned files, zero re-forming), vibevm/vibespecs/manual-tests MT-01/02/03, spec/discipline/README. Journal step per file; batch commits ~1–3 files; then the Phase B exit gate (check --exhaustive clean over the 59-file scope) and the §4 boundary ritual.
• 2026-07-25 · B2 batch 20 — PROP-003 marked; third superseded-arc split. 310 units → 313 anchored facts, 0/0 (d596c631). The libsolv engine sections (§2.2/§2.3/§3.x, phase A, migration step 1) record history at spec/done — the SUPERSEDED-by-PROP-017 blockquote re-formed verbatim per F-015; the dependency vocabulary is impl/done verified against the shipped crates (features.rs incl. weak ?/ and exclusive groups, activation.rs — if_os impl/work per its recorded deviates, conditional.rs + fixpoint, manifest/i18n.rs, the four vibe-check entries, lockfile meta/package fields). Unshipped details stay spec/done: pin_preferences (recorded deviates), VIBE_LANGUAGE, dotted-key translations, --all-languages, vibe review, outdated --upstream, the LLM emission engine (vibe-llm pending); Phase F impl/plan. req-line fact ids dodge the owner-minted req-* heading anchors via -req/-design suffixes. Ledger +2: F-030 (stale design-proposal status line — F-024 family), F-031 (internal r2 leftovers: §2.8 fence r1 syntax, §4.3 → §2.5.4 misref, §2.7.5/§2.9 examples vs shipped language_chain/schema-v5 shape).
• 2026-07-25 · Owner amendment — Phase L (legacy relocation) inserted between B and C. Directive quoted verbatim in the §5 Phase L section; the four dirs vibevm/vibespecs/terraforms vibevm/vibespecs/research spec/neworder spec/discipline leave the spec tree for root legacy-spec/ once nothing references them: L1 reference inventory (living corpus + code — specmark, doctests) → L2 fact discovery → L3 fact relocation into common/design/modules/manual-tests (new specs allowed — owner grant same day) → L4 fact-grain markup at the new homes (+progress.toml scope grows) → L5 git mv to legacy-spec/. Ordering law: before Phase C, so verification covers the relocated facts. Phase C entry updated to "B and L closed". Review point (OPEN): the campaign plan itself lives in vibevm/vibespecs/terraforms/ — relocate mid-campaign or at close-out? Owner call before L5.

• 56RESOLVED scope ruling (owner, 2026-07-25, second in session): «spec/discipline нужно исключить из анализируемого корпуса, потому что теперь Дисциплина - это часть пакетов ai-native, а саму spec/discipline после определения и портирования ссылок - перенести в legacy-spec. Сейчас она всё ещё в основном корпусе». Ruling applied: spec/discipline/** leaves the progress.toml include enumeration — the Discipline's living home is the ai-native packages (core-ai-native + the language stacks), so the host copy is a historical record like terraforms/research/neworder. Scope: 59 → 58 files; discipline/README's 16 facts leave the corpus; the B2 tail is now boot pair + manual-tests only. The Phase L §5 note that kept discipline/README in the B scope is superseded (corrected in place); Phase L's four-directory relocation list is unchanged — discipline still relocates to legacy-spec/ after L1's reference inventory and L3's fact porting.

• 572026-07-25 · PHASE B CLOSED — the corpus is fully marked; exit gate green in full. Gate: progress check --exhaustive clean over the final scope (58 files, 4 880 facts, 4 944 markers, 0 errors, 0 warnings); floor bash tools/self-check.sh → all green, real exit code 0. Final scope after the two 2026-07-25 rulings: 58 files / 4 880 facts (94→59 terraforms/research/neworder; 59→58 discipline). The B2 tail (batches 20–26, this session): PROP-003 — 313 facts, the third superseded-arc split, vocabulary verified against the shipped crates (d596c631); PROP-002 — 359 facts, modules close 35/35 (9328becb); design/README re-mark burns the last 12 expected errors — the gate reads 0 for the first time in the campaign (cb6e55b0); loading-and-boot-model + action-system (d1a09275); workspace-and-qualified-naming + tui-visual-language — design 6/6 (91fde06c); the authored boot pair, additive-only honoured line-for-line (ae10dca2); manual-tests MT-01/02/03 — genre precedent: step items and their Expected paragraphs are separate units, pending sign-offs honestly impl/work (7d9dd964). Interleaved the same session: the Phase L owner amendment (c0147947) and the discipline scope ruling (1c48019a). Ledger at close: 35 findings (F-030…F-034 new — two stale-header rows, PROP-003's internal r2 drift, the lore "forthcoming" family). Predictions (§8) at the boundary: P2 (the counter catches what review misses) CONFIRMED repeatedly — this session it caught the PROP-003 weak-deps continuation and the workspace-naming §7 intro; P5 (DRIFT no-return) holds at 5/5; P6 for B CONFIRMED with room — the markup pass ran 2026-07-24→25, two days against the 1–1.5-week budget — while the §5-B batch-count guess (9–12) was falsified by the mid-phase fact-grain amendment (~30 batches at the finer grain; the grain change, not slippage); P1/P3/P4 stay open for C/D. Phase lane: the journal-derived phase stays B until the owner opens the next; per the 2026-07-25 amendment the next phase is L (legacy relocation), then C. Awaiting the owner's call to open Phase L.

• 58RESOLVED review point + the B→L handoff (owner, 2026-07-25, closing the session). The plan-file question is settled: the first instinct («Сам план кампании тоже перенеси в legacy-spec») was reversed in the same message — «Я передумал. План этой кампании пока переносить не нужно. Я хочу чтобы у нас остался правильный набор спецификаций и других инструкций, чтобы мы могли делать периодические проверки каждый месяц». Ruling applied: the campaign plan stays in vibevm/vibespecs/terraforms/; L5's move excludes it (the §5 Phase L review point is rewritten RESOLVED in place); its eventual relocation is a future owner call, no earlier than close-out — the §6 monthly recurrence keeps consuming it in place. And the handoff: «Перенеси все эти активности в следующую сессию и напиши как мне ее продолжить» — Phase L execution moves wholesale to the next session; this session ends at the closed-B boundary; CONTINUE.md carries the cold-resume recipe and the recommended opening phrase by which the owner opens Phase L explicitly (the phase event lands in the journal only then, per the DRIFT-003 phase lane).

• 592026-07-25 · PHASE L EXECUTED AND CLOSED — the legacy dirs are out of the spec tree. Opened on the owner's recorded phrase (journal phase event L + l1-inventory). L1 (inventory): gate-binding set = 26 sites in 13 corpus files + 1 crates doc comment (outdated.rs); spec/neworder and spec/discipline had zero corpus inbound; out-of-gate referrers classified into live docs (ROADMAP, docs/), historical reports (terraform/), campaign zone, and an explicit leave-list (packages vendored comments, neworder2 baselines, AUDIT quote, closed debt-ledger row DBT-0016, VIBEVM-SPEC — no real refs). L2 (fact discovery) verdict: every cited fact was already corpus-resident — the RP1 rejected-alternative at design/action-system.xml §4 D1, the ten design decisions, the DO1–DO18/Δ1–Δ16 sets restated in place, the settings deltas named inline, the campaign histories in PROP-038 §6 / PROP-027 / PROP-036 — so L3 ported nothing and the owner's new-spec grant went unused; every citation dissolved into archive-provenance form instead (the honest inverse of the plan's port-then-repoint expectation, recorded here as the L2→L3 finding). L4 (repoints): four batch commits 83346e78 f8f347d8 9514e8fb 1ec6a27c — 26 sites incl. both spec://org.vibevm.core/vibevm/research URI retirements, plus four word-level sites the path greps could not see (PROP-031 status-line, PROP-037 plan pointer, PROP-040 delta-mapping, PROP-041 spec.research §3.7): the literal-backtick and dotted forms needed a lookbehind/word sweep — a reusable lesson for the §6 recurrence. Scope stayed 58 files (no new files → no progress.toml growth). L5 (relocation): 70f3cbdd — 35 files git mv'd (terraforms 25, research 8, neworder 1, discipline 1) to root legacy-spec/, the campaign plan carve-out honoured; live out-of-gate pointers followed in the same commit (ROADMAP 15 occurrences, docs/ 8, terraform 2 links, findings.json 11 paths, discipline.lock recipe, progress.toml comment); historical prose, quoted URIs, the closed debt row, and the pre-broken PLAYBOOK link stayed verbatim — records are not rewritten. f311f429 regenerated the stale host specmap (absorbed B-phase drift + the move; ratchet 37 gated orphans within allowance, 0 suspects). Exit gate: the reference greps read zero into the four dirs from corpus + crates (plan carve-out aside); check --exhaustive clean (58 files, 4 880 facts, 0 errors); floor self-check all green, real exit 0. Phase C (verification) awaits the owner's opening call per the resume-boundary law.

• 602026-07-25 · PHASE C OPENED — the verify loop is live; the boot pair and manual-tests carry verdicts. Mechanics fixed for the whole phase (PROP-043 §7.1/§7.5): verdicts live in the cache's per-file campaign map — {verify_batch, verified_at, processed_hash, verdicts{anchor → {v, ev[]}}, summary} — never in markup; scan preserves the maps (verified live) and projects them into corpus.json for the dashboard. Verdict semantics by stage: impl/done ⇒ presence evidence; spec/done ⇒ absence (shipped-but-still-marked-spec is the stale-header drift); doc/done ⇒ no contradiction with the contract; dated historical records confirm unless falsified; present-state claims blocked by the GitVerse outage go unverifiable, never "probably fine". Per-file coverage is assert-gated (extractor anchors == cache marker_count == verdict keys). c0-boot (bb337e90): 64 facts — 61 confirmed / 1 drift / 2 unverifiable; the drift is real (LAYER-CODE names a nonexistent root tests/ → F-035, user-owned file so the wording fix is the owner's). c1-manual-tests: 67 facts — 61 / 6 / 0; MT-01's EXP-2/6/7/8/9 describe the pre-revision TUI keymap (shipped: F1…F6 menus, Shift+arrows tabs, Esc+confirm quit) → F-037 re-author; MT-02's footer quote omits F4 and says q-quit, its "once a picker lands" is superseded by the F4 settings menu → F-038; and the sweep caught a code-side stale clap help on --plain contradicting the shipped console-TUI default → F-036 (Phase E DRIFT candidate). MT-03 verified clean 16/16. Running tally: 131 / 4 944 markers judged — 122 confirmed / 7 drift / 2 unverifiable; findings 38. Machine-evidence base mapped: specmap carries 626 edges into vibevm/vibespecs/modules units and 111 into vibevm/vibespecs/common (section grain; facts inherit their section's edges), so the module cluster is the evidence-rich grind; design (6 files) → common → modules is the queued order.

• 612026-07-25 · c2-design — the design cluster verified; the drift is the aged-tense family. 306 units judged (300 fact anchors + 6 status-element bundles; coverage law recorded in the cache maps: verdicts key on fact anchors, table cell-markers inherit their row, <status> elements judged as _elements): 291 confirmed / 15 drift / 0 unverifiable. The drift map: loading-and-boot-model 8 — the lore's three inclusion types inline/static/dynamic (default static) against the shipped link = "static" | "dynamic" (default dynamic), plus "forthcoming" ×2 and the §6 static|static typo → F-039; action-system 4 — the F-034 forthcoming family, sharpened by aiui.rs actually shipping list_actions + invoke against the doc's "Not built now"; workspace-naming 1 — the M1.18-vs-M1.19 milestone shift → F-040; tui-visual-language 2 — the "current ASCII scaffolding" present-tense and "When §2.2 carries" against five existing anchors → F-041. Hard confirmations: the detect_tier signature matches the lore literally, PAD_X/PAD_Y/ GUTTER exact, palette hexes byte-equal in rose_pine.rs/ catppuccin.rs, exit code 7 = AMBIGUOUS_PACKAGE, the §2 module table maps 1:1 onto crates/vibe-actions/src, and the two-way design↔PROP backlinks hold 4/4 (structural-loader parked by its own spec/hold). Running tally: 437 units judged — 413 confirmed / 22 drift / 2 unverifiable; findings 41. Next: vibevm/vibespecs/common → vibevm/vibespecs/modules (the specmap-evidence-rich grind).

• 622026-07-25 · c3a-common-small — five common PROPs verified; one roster drift. PROP-006 (frozen pointer), PROP-013 (audit instance), PROP-016 (source mirrors), PROP-028 (families), PROP-029 (FQ addresses): 150 units — 149 confirmed / 1 drift / 0 unverifiable. The drift is PROP-028's family roster aged against the tree: core-ai-native ships v0.8.0 (the fact says 0.7.0), the go-ai-native family is in force (aggregator v0.1.0 + -lang + -mcp) but absent from §2.2/§2.3, and aggregators carry a LICENSE.md from the UPL relicense wave against the "vibe.toml + README and nothing else" letter → F-042 (one roster refresh fixes all three). Prime confirmations: cargo xtask mirror ran three times in-session (the tracking-ref refresh of HIST-TRACKING-REFS observed live in its output), push_args_never_force exists at xtask/src/mirror.rs:426, mirrors.toml matches the §2 block verbatim, every §open question across the five files verified genuinely open, and PROP-029's three carriers match the generated INDEX/STATIC forms character-for- character. Running tally: 587 units — 562 confirmed / 23 drift / 2 unverifiable; findings 42. Remaining: the common big seven (PROP-000, 018, 019, 024, 031, 032, 033 — 868 markers; specmap-rich on 018/019) → the modules cluster (3 300 markers).

• 632026-07-25 · c3b — PROP-000 verified: the foundation aged in six spots (the densest drift file of the phase). 162 units — 149 confirmed / 12 drift / 1 unverifiable, all twelve drifts one family row F-043: §3 still records the proprietary EULA though its own revisit trigger fired (UPL-1.0 relicense 2026-07-12, MT-05); §4 lists the retired vibe-package.toml; §6 records pre-qualified identity and four kinds while the same file's INV-VOCABULARY correctly lists five (internal r2-leftover-class inconsistency); §7 calls GitVerse "the source-of-truth" against PROP-016's no-primary model and pins KindName against the Fqdn default; §14's WAL-names-the-runs practice lapsed; §18 claims LLM-reviewed semantic conflicts whose LLM lane is pending. Hard confirms: all seven §2 crates exist, the §15 prune of PROP-001 was executed (its ARG-PRUNED cites §15 back), schemas/ + vibe-wire/generated committed, resolvo pinned, both guides exist. The unverifiable is the GitVerse-network-bound legacy-registry claim (same verdict as c0's twin).
• 2026-07-25 · c3c — PROP-018 + PROP-019 confirm wholesale on implementation evidence. 322 units — 320 confirmed / 2 drift, both drifts the same shape and both already ledgered in Phase B: a proposed-era status line over a fully shipped system (F-013 agentic modes, F-005 vvm). Beneath the headers the content carries the densest machine evidence of the phase: the agentic relay/skill/affinity/transports/explain sections hold 23 implements / 8 verifies into vibe-mcp/vibe-cli (the vibe agentic + vibe command verbs run live in this very session), and the vvm's twelve sections map 1:1 onto commands/vvm/ exactly as §3 placed them (37 implements / 38 verifies; relocate alone carries seven). Far-backlog sections verified genuinely unbuilt. Running tally: 1 071 units — 1 031 confirmed / 37 drift / 3 unverifiable; findings 43. Remaining in common: c3d = PROP-024 / 031 / 032 / 033 (384 markers), then the modules cluster (3 300).

• 642026-07-25 · c3d — the common tail verified; vibevm/vibespecs/common is CLOSED 12/12. PROP-024 / 031 / 032 / 033: 327 units — 322 confirmed / 5 drift / 0 unverifiable. All five drifts are PROP-024's: the F-006 proposed-era header (ledgered in B) plus a new family row F-044 — the §2.6 deferral fired: the TypeScript pilot shipped, so the deferred engine split executed and core-ai-native now authors the neutral engines (conform/specmap/specmark/mcp cores, vendored byte-identically per PROP-028), making CORE-STAYS-PROMPT-ONLY and OOS-TS-CHECKER false as present-state; the related line also cites the vanished vibedeps/flow-core-ai-native/0.6.0 slot. The §2.4 consumption topology confirmed verbatim (root Cargo.toml exclude = ["packages", "vibedeps"]; self-check drives the vendored engines by --manifest-path exactly as BINARY-RUN-FORM specifies). The three design proposals (031/032/033) verified honest end to end: every "schedules no implementation" claim grep-verified (no move-unit/rename-address in the stack, no code:// nodes, no [[refactoring]] manifest table, no vibe refactor CLI), and PROP-032's decided-in-place q5 record carries its owner date. Running tally: 1 398 units — 1 353 confirmed / 42 drift / 3 unverifiable; findings 44. Verified clusters: boot ✓ manual-tests ✓ design ✓ common ✓ (23 of 58 files). Remaining: the modules cluster (35 files, 3 300 markers — the specmap-richest).

• 652026-07-25 · c4a — the modules cluster opens on the campaign's own contract; the tool that runs this phase verifies itself. vibe-progress family: PROP-043 + OWNER-GUIDE + templates ×3 — **250 units (245 anchors
• 5 element bundles): 236 confirmed / 14 drift / 0 unverifiable. PROP-043's ten drifts split two ways. F-045 — the file's own status aged behind the campaign it governs: the status-line still says "implementation underway (Phase A)" and holds (provisional) sections that no longer exist (grep = 1, the sentence itself), while the phase lane reads C and §5 is fully shipped (19 implements edges; all seven subcommands + every documented flag verified live this session). F-046 — a marker-vs-implementation parity family in both directions: impl/done over unshipped fragments (EvidenceProvider wired nowhere outside the core — the adapter imports everything but evidence; 0 fact units/edges in host specmap.json — the consumed stack engine v0.7.0 predates the fact amendment, PROP-014 v0.8.0 §2.1 exists authored-side awaiting the re-mint; CMD-CHECK's "lossless folds" matches zero code; campaign.json carries no gates field; the report has no evidence column; Cache::is_current is dead outside its own tests — every run re-parses the tree), and spec/done under shipped code (§7.3: BaselineUnit matches the record field-for-field, rescan live with hash-suspect + marker-diverged; named-crates and the control sample honestly missing). OWNER-GUIDE's four drifts are all one F-020 refresh sweep (4-of-6 placements, the preamble-less amendment missed twice, FOUR-SURVIVORS omits tasks/). The templates confirm wholesale: impl-task exercised 5/5 (DRIFT-005 checked field-for-field), spec-task correctly awaits Phase D, harvest-card consistent with its empty harvest/. Method note for the batch map: coverage now reads from progress mirror's ParsedDoc (authoritative fence-aware parse) — the raw-grep extractor over-counts code-span shorthands. Running tally: 1 648 / 4 944 — 1 589 confirmed / 56 drift / 3 unverifiable; findings 46 (next free F-047).**

• 662026-07-25 · c4b — the registry core verified; the evidence-richest file confirms at 98 %. PROP-002 (360 markers, 110 specmap edges) + PROP-001 (93): 379 units — 371 confirmed / 8 drift / 0 unverifiable. The headline confirmation: RESOLVO-PRIMARY holds on live evidence — ResolvoDepSolver is the shipped production default (registry.rs unwrap_or("resolvo"), resolvo_engine = "the production DepSolver cell"), and the §2.8 fallback-seam story (naive + sat as selectable cells behind DepSolver) is exactly what shipped. That same check exposed F-047 (code-side): the two #[spec(deviates)] reasons in naive.rs/sat.rs still claim "no ResolvoSolver exists in tree" / "adopting resolvo stays an owner decision" — they aged behind the very adoption they awaited. The spec-side drift splits into two touch-up families: F-048 (PROP-002 precision: --trust-mirror promised twice and shipped nowhere — only --trust-redirect exists; vibe list --overrides promised twice and absent; the git-source ref errors ship as reason-strings, not the named MissingRef/ConflictingRefs; source_kind grew path/embedded in PROP-007/030; the cache-slot example still shows the kind-name era against the live packages/<group>.<name>/clone) and F-049 (PROP-001: no crate README behind mechanics-in-readme; NO-OFFLINE-YET aged behind the shipped --offline; the git-binary parking-lot entry resolved by the shipped VIBE_GIT_BINARY whose comment cites §6 back). Everything else confirms on dense machine evidence: the redirect subsystem wholesale (12 implements / 8 verifies, three CLI verbs, hop-limit guard, --trust-redirect at cli/registry.rs:226), the auth-silencing matrix (apply_common_env + regime-aware force_silence + 3 tests_pure verifies), merge_effective / url_is_local / registry_config_path live under their exact spec names, the enabled filter at the R-001 construction point, the mirror fall-through loop, token redaction tests under the exact names the spec cites, and the lockfile schema chain v2→v3→v4→v5 documented end-to-end across PROP-002 → PROP-007 → PROP-008 (live lock: schema 5). Running tally: 2 027 / 4 944 — 1 960 confirmed / 64 drift / 3 unverifiable; findings 49 (next free F-050).

• 672026-07-25 · c4c — the resolver pair verified; one c4b verdict corrected. PROP-003 (314 markers) + PROP-017 (106): 372 units — 359 confirmed / 13 drift / 0 unverifiable. PROP-017 verifies almost wholesale on the c4b code evidence (architecture item-for-item: SemverVersionSet literal to the spec, NowOrNever sync adapter, shared output builder, dominance oracle, capability closure pre-scan); its only drift is the known F-027 status-line ("implementation in progress" + an impl/work document marker over its own §6 "the port is COMPLETE"). PROP-003's eleven split three ways: F-030 (the design-proposal status line over a shipped vocabulary), four F-031 rows landing exactly as the B-phase ledgered them (the r1 __exclusive fence, the §4.3→§2.5.4 misref, and the language_chain-vs-language+language_fallback trio — lockfile.rs:130 is one merged field), and a new F-050: the solver-era tail outside the §2.2 supersede marker — §2.1 still promises "SatDepSolver becomes the default" with a [meta].solver selection key and a naive|sat CLI (shipped: naive|sat|resolvo, default resolvo, no meta.solver field), §2.11/§6 still record a sat-default flip, and vibe update --features was promised but never wired. The same evidence trail corrected c4b: PROP-002's SOLVER-IDENTITY-FIELD had confirmed on a section default, but the live lock has no solver field and PROP-017 §8 says so — verdict amended to drift, F-048 extended (f), and F-047 extended with a third stale deviates (lib.rs:288-296 still claims "SatDepSolver is not in tree" while sat.rs ships). Subskills confirmed with a live wink: this session's own MCP toolbox carries read_subskill / materialise_subskill — the M1.7 lazy-delivery surface §2.5 designed. Running tally: 2 399 / 4 944 — 2 318 confirmed / 78 drift / 3 unverifiable; findings 50 (next free F-051).

• 682026-07-25 · c4d1 — the workspace big three verified; the loading model is the session itself. PROP-007 + PROP-009 + PROP-035 (386 markers): 360 units — 351 confirmed / 9 drift / 0 unverifiable — and zero new findings: every drift row lands on ledger entries the B-phase already minted. PROP-007's four are all F-029 — the status/milestone lines, §9.3's deferral record, and the document marker still call workspace-aware vibe install "the remaining piece" while M1.18 shipped it (Workspace::discover live at plan.rs:101 and apply.rs:114). PROP-009's three: the F-026 DRAFT header pair, plus SURF-SHOW-EFFECTIVE caught shipped-but-spec/done (vibe show effective exists in its simple concatenation form; the §2.8 engine projection honestly stays v1.5). PROP-035's two are the F-028 DESIGN-provisional pair over a §17 that records the compiler shipping three times (vibe-spec 07-15, the link-type rename 07-16, normal+static AOT 07-20). The confirmations needed no reconstruction — the model under test booted this very session: STATIC.md read first, the TOML INDEX.md with its [[entry]] grammar, committed vibedeps/, when = "os:*" in the renderer, BootCategory ordering with the conflict errors deleted, vibe reinstall citing §2.10 from its own --help, and the vibe-spec pipeline (doctree fact leaves, :add/:replace with per-fact override and the merged-view DuplicateId gate, embed cycle guards, PackageFormat simple-by-default) exactly where §5–§13 put it. Running tally: 2 759 / 4 944 — 2 669 confirmed / 87 drift / 3 unverifiable; findings 50 (next free F-051).

• 692026-07-25 · c4d2 — the workspace tail closes; the honesty law meets its densest case. PROP-011/012/038/034/020/022/025 (482 markers): 474 units — 393 confirmed / 81 drift / 0 unverifiable. The headline is the pair everyone saw coming: PROP-020 (41 drift) and PROP-022 (37 drift) are whole shipped systems under "proposed" headers — hooks.rs implements the entire install-hooks contract (phases, Git-Bash-first Windows selection, DEFAULT_ALLOWED_GROUPS with org.vibevm, the non-interactive refuse, the env contract, both test seams) and the materialization system ships wholesale (the Materialization enum with doctests, hardlink/in-place through vibe-install, submodule snapshot-embedding, the destructive guard) — while the B-phase, marking wording-faithfully, left the text's own spec/done in place. Each file is ONE F-018 re-mark sweep (note extended), the same shape as PROP-000's F-043 twelve. Against that: PROP-011, PROP-012 and PROP-025 verify clean — 227/227 confirmed, the campaign's first zero-drift files, all three honestly-marked shipped contracts (the freshness oracle, the <vibevm> block engine whose canonical text lives in this repo's own CLAUDE.md, and the vibe bin family whose help text matches §4 verbatim). PROP-038 adds two F-024 rows (the §2.7 open-note its own §5 resolves, and fuzz.rs shipping the very DEF-5 its history still defers); PROP-034 takes a new F-051 for a DRAFT status line the file's own rename-note already outgrew. Running tally: 3 233 / 4 944 — 3 062 confirmed / 168 drift / 3 unverifiable; findings 51 (next free F-052).

• 702026-07-25 · c4e — the cli/actions cluster verified; Spec 2 is the reality, DRAFT is the costume. PROP-037 + PROP-036 + PROP-042 + PROP-039 (327 markers): 300 units — 290 confirmed / 10 drift / 0 unverifiable. One new family, F-052: all three contract headers still read "DRAFT — requirements" while their plans executed (PACKAGE-TREE, TREE-TUI v0.2, and the vibe-actions crate whose first consumer — the Spec-2 TUI — is the shipped vibe tree); PROP-036's §2.11 sketch keys (q quit, n order, x/t/TAB modes, F fold-all) are superseded by Spec 2's F-key scheme exactly as the c1 manual-test sweep found from the other side (F-037/F-038); and PROP-039's AIUI rows ("designed-for; not built now") aged behind their own delivery — vibe aiui state prints the §11.2 ModelView and cites the section in its --help. F-023 lands as judged (the §2.13 launcher misref into Progress-Control's PROP-043). PROP-042 verifies clean 58/58 — the fourth zero-drift file: an ACTIVE header over all seven aiui verbs live verb-for-verb, the vibe term launcher, and the icon-swap OSC. The evidence base was half-built already: c2's byte-equal palettes / detect_tier / spacing constants and c1's shipped-keymap findings carried §2.2 and §5 wholesale. Running tally: 3 533 / 4 944 — 3 352 confirmed / 178 drift / 3 unverifiable; findings 52 (next free F-053). Remaining: c4f only — vibe-mcp ×3, vibe-index PROP-005, vibe-settings ×2, and the vibe-registry rest (008/010/030/021/023/README), 957 markers.

• 712026-07-25 · c4f1 — the registry rest; the campaign's densest drift file falls. PROP-008/010/021/023/030 + the modules README (372 markers): 364 units — 233 confirmed / 131 drift / 0 unverifiable. The headline: PROP-030 takes 63 drift rows in one new family (F-053) — the embedded registry ships wholesale under a bare "proposed (2026-07-13)" status line, while the WAL's own standing constraints cite its §5 CI-off gate split and §3.3 project-local rules as living law, all five flags sit in install --help, R-001 pins its sanctioned constructor, and this very machine resolves redbook/rust-ai-native through it daily. F-018 now covers the full four-spec bridge family: PROP-021 (24 drift — recurse clone/update and snapshot embedding live in the backend) and PROP-023 (35 drift — [package].bridge parses with its doctest) join PROP-020/022; four headers, four re-mark sweeps, one family. The honest side held: PROP-008 verifies 91/92 (IMPLEMENTED M1.18+M1.19, live from the qualified refs in this repo's lock to exit code 7), PROP-010 91/93 (a genuine DRAFT the ROADMAP's M1.20 row corroborates — its only drift is a motivation pair that still claims "no --offline exists", F-054 with PROP-008's un-reconciled milestone line), and the modules README lands its ledgered F-016 six rows exactly as the B-phase wrote them. Running tally: 3 897 / 4 944 — 3 585 confirmed / 309 drift / 3 unverifiable; findings 54 (next free F-055). Remaining: c4f2 (vibe-mcp ×3 + vibe-settings ×2, 306 markers) and c4f3 (vibe-index PROP-005, 279).

• 722026-07-25 · c4f2 — mcp + settings; the cleanest batch of the phase. PROP-015/026/027 + PROP-040/041 (306 markers): 279 units — 277 confirmed / 2 drift / 0 unverifiable. Three zero-drift files in a row: PROP-015 (65/65 — vibe mcp serve/install/status live, and the judging session is itself an MCP consumer of the judged server), PROP-026 (47/47 — the superseded-in-topology arc holds exactly as the B2 markup split it: the tcg grammar normative through the family MCP servers, the standalone topology gone with wave 6), and PROP-027 (58/58 — the mcp kind and the sovereign per-family servers are this machine's installed reality). The two drift rows are the settings pair's ledgered headers: F-010 (PROP-040's DRAFT over a crates/vibe-settings whose CLI cites "§8" in its own help) and F-019 (PROP-041's DRAFT over the prefs TUI that PROP-037's F4 opens by name). Running tally: 4 176 / 4 944 — 3 862 confirmed / 311 drift / 3 unverifiable; findings 54. Remaining: c4f3 — vibe-index PROP-005 (279 markers), the last file of the phase.

• 732026-07-25 · c4f3 + PHASE C EXIT GATE — the corpus is measured. PROP-005 (279 markers, 91 edges) closes the phase all-confirmed — the eighth zero-drift file: an "implemented; folded into the workspace" header verified against the living crate, the group-native M1.19 index shape (workspace_origin in both the §2.6 schema and the shipped JTD), vibe search/outdated surfaces, and the M2.10 smoke. The exit gate is green in full: 58/58 files carry campaign maps; 4 944 / 4 944 markers carry verdicts (anchors + row-inherited cells + document elements). Phase C final tally: 4 455 units judged — 4 141 confirmed / 311 drift / 3 unverifiable = 93.0 % / 7.0 % / 0.07 % — the first measured actuality level of the spec tree. Findings stand at 54 (F-001…F-054; F-045…F-054 minted and F-018/F-024/F-047/F-048 extended by this phase's modules cluster). The §5-C prediction confirms, mirrored: drift indeed concentrates around Status lines — but in the inverse polarity: headers promised less than the tree delivers (the shipped-under-proposed families: PROP-030's 63 rows, the bridge four's 137, PROP-000's F-043 six), while honestly-updated IMPLEMENTED files confirm nearly wholesale (eight zero-drift files: PROP-011/012/025/042/015/026/027/005). The three unverifiable stay network-bound (the GitVerse legacy-registry claims). The drift ledger is sweep-shaped for Phase D: ~15 family rows cover ~80 % of the 311 rows. Phase C is complete; the phase lane stays C — Phase D opens only on the owner's call, per the §5 entry law.

• 742026-07-25 · post-gate floor note — one red, root-caused, ledgered (F-055). The phase-close self-check run went red on cli_pkg_cycle::install_from_git_registry ("expected one registry cache bucket"). Root-caused and proven in-session: a ~/.vibe/registry.toml appeared on this machine today (15:37, vibespecs + vibespecs-gitverse), and the test isolates VIBE_REGISTRY_CACHE but not the settings chokepoint, so the global registries merge into the "hermetic" resolver and mint a second cache bucket. With VIBE_SETTINGS pointed at an empty dir the same test passes. The campaign's commits are docs-only and unrelated; the campaign's own gate (progress check) stays 0. Ledgered as F-055 (Phase E: the e2e harness pins VIBE_SETTINGS to a tempdir — the never-touch-real-user-state law, applied inbound). Findings close the phase at 55.

• 752026-07-25 · PHASE D OPENED (owner call) · wave d1 — the shipped-under-proposed families fall in one sweep. The owner opened stitching («Открываю Phase D»); the phase event D is in the journal and the lane re-derives from it. Wave 1 took the two biggest families — F-053 (PROP-030, 63 rows) and F-018 (the bridge four PROP-020/021/022/023, 137 rows) — plus F-043 (PROP-000, 12 rows): 212 of 311 drift rows closed in one wave (68 %). Every row was a reality-mismatch, so the whole wave ran through sync-from-code: the diffs were drafted, surfaced to the owner, and applied only on his approval («Правь и продолжай дальше»), which also settled the licence question in the owner's own words — «Лицензия теперь UPL, на самом деле». Mechanics: the 191 pure re-marks were scripted off the C-phase verdict map (@spec/done → @impl/done on exactly the anchors judged drift, dry-run first, progress check after) — no model touched them; the 9 prose units (5 status lines, 4 document elements) and PROP-000's 13 facts were authored by hand. The five new status lines cite the shipped surface rather than asserting it: HooksDecl + hooks.rs runner + apply wiring; the Materialization enum with its doctests and the destructive guard; --recurse-submodules + snapshot embedding + resolved_commit; [package].bridge with its doctest and the three composed mechanisms; the R-001 sanctioned constructor site with §4's knobs in install --help. What did not move is the honest half: in all five files the surviving spec/done rows are exactly related / motivation / REJ-* / OOS-* / the D1–D5 decision records — facts that never become impl. PROP-000 took twelve authored fixes (UPL-1.0 with an explicit repository scope; vibe-package.toml retired; qualified identity and the fifth kind mcp; uniqueness by (group, name); the multi-homed PROP-016 posture replacing "GitVerse = source of truth"; fqdn repo naming) plus a verdict correction in the c4b→c4c pattern: LICENSE-REVISIT was swept up in c3b's file-level confirm although its trigger had fired on 2026-07-12 — re-issued per the decision-records flow with two fresh conditions. One paired edit landed outside the corpus: vibevm/vibespecs/WAL.xml regained the outstanding-manual-runs line (MT-02, MT-03) that MT-WAL-NAMES requires, which is what made that fact true again rather than merely reworded. Ledger: 4 353 confirmed / 99 drift / 3 unverifiable = the tree measures 97.7 % true, up from 93.0 % at the Phase C gate. Findings resolved: F-007, F-014, F-018, F-043, F-053 (6 of 55 closed; 49 open). progress check 0 throughout. Next: wave d2 over the mid-size families (F-052, F-046, F-050/F-031/F-030, F-039/F-033, F-016, F-032/F-048).

• 762026-07-25 · wave d2 (a…f) — the ledger empties to nine rows, and the spec tree measures 99.7 % true. Six sub-batches, 85 rows, 30 files. d2a — the stale-header family (22 rows, 13 files): every "DRAFT — requirements" over a shipped crate rewritten to the shipped truth with the surface cited (PROP-036/037/039/040/041/034/017/035/ 009/007/008/018/019). Where a header pair contradicted itself — PROP-007/008/009 each carrying "design proposal, not implementation-locked" one line above a status recording shipped milestones — the milestone line moved. Two facts were re-based rather than softened: PROP-039's "AIUI not built now" non-goal (built: aiui.rs + vibe aiui state) and PROP-007 §9.3's deferral (resolved by PROP-009 §2.7 + M1.18 Phase 5). PROP-036 §2.11 was fixed by deletion: it restated the keymap PROP-037 owns, so the keys are gone and the section cites the contract — the single-source law is what makes that the fix. Its §2.13 launcher misref (F-023) repointed to the vibevm-term products repo, where the launchers actually went. d2b — design-doc tense (15 rows, 4 files): lore corrected to the contract per spec-genres. The vocabulary drift was handled deliberately differently: the loading-model record is a dated 2026-05-21 capture, so rewriting it into today's names would destroy the record of what was decided — it keeps its own words and gains a rename note plus a per-row shipped-name gloss. The literal "static" | "static" | "dynamic" typo is just a typo, and is fixed. d2c — PROP-003's solver tail (11 rows): corrected clause by clause rather than by widening the §2.2 supersede block, so a reader landing on any single anchor gets the truth without hunting for a marker elsewhere; [meta].solver was never wired at all, and PROP-017 §8 plus the live lockfile confirm it. d2d — index, registry, families, owner guide (29 rows, 8 files): the module index had stopped in May, so five module directories and every PROP after 012 were invisible — 26 rows added, each carrying the status Phase C verified. Four promised-absent surfaces now say so (--trust-mirror, vibe list --overrides, lockfile [meta].solver, vibe update --features): re-marked spec/done and named as unbuilt at the point of the claim. Three moved the other way (PROP-001's VIBE_GIT_BINARY parking-lot entry, PROP-038's two open notes) and PROP-024 recorded the §2.6 deferral that fired — the TypeScript pilot was the second language, and core-ai-native now authors the neutral engines. d2e — MT-01/MT-02 re-authored (6 rows): the tests walked a human through n/x/t/Tab/q, none of which exist; the steps now match the shipped F-key surface. The run status did not move — an agent may pre-run, only a person signs off, so MT-02 and MT-03 stay unsigned and vibevm/vibespecs/WAL.xml now names them (the paired fix that made MT-WAL-NAMES true). d2f — PROP-043's own status (2 rows). Method: the wave minted 31 new anchors (a supersession pointer, a rename note, 26 index rows, 3 guide placements), each sealed with its own verdict so the every-marker-carries-a-verdict invariant survives Phase D. Ledger: 4 474 confirmed / 9 drift / 3 unverifiable — 99.7 % true, from 93.0 % at the Phase C gate. Findings 42 of 55 resolved. progress check 0 throughout; the floor is red at exactly one test, the environmental F-055 recorded at the Phase C close (proven unrelated: VIBE_SETTINGS pointed at an empty dir turns it green). The nine survivors are both escalations: F-046's eight PROP-043 parity rows (the owner picks wire-or-demote per item) and F-035's single row in vibevm/vibespecs/boot/00-core.xml, a user-owned file the boot contract says may only be marked additively.

• 772026-07-25/26 · d2g, d2h, and the parity row that needed no ruling — PHASE D CLOSED at seven rows. Both escalations the d2 wave handed the owner resolved without widening anything. F-046's BASELINE-RECORD turned out not to be a wire-or-demote question at all: baseline.rs's BaselineUnit carries §7.3's fields one for one, with load/store and the rescan CLI live — shipped under a spec marker, the pattern this phase had been closing all day, so it closed the same way and the owner's list shrank by one before he had to read it (0064fd4a). F-035's single row in vibevm/vibespecs/boot/00-core.xml moved under the owner's explicit ruling, which is the only way that file may be touched: LAYER-CODE named a root tests/ directory that has never existed, and now names each crate's own. d2g/d2h took the archive (f6c17f92): six findings pointed into legacy-spec/, and the tempting answer was to close them as out of scope — nothing in the living corpus cites the archive normatively, so its stale lines mislead no build. The owner ruled otherwise and the reason holds up: a cold reader who lands in the archive directly reads "PLANNED · not started" on a campaign that finished weeks ago, or opens TREE-TUI-PLAN v0.1 with no hint that v0.2 superseded and executed it. Four plans now say EXECUTED and one SUPERSEDED, each carrying ARCHIVE — do not execute and each keeping its original line verbatim, so the record of what it once claimed survives. PROP-026 took the same treatment on the live side — its document marker had recorded the wave-6 supersession since the markup pass, but the prose still named vibe-tcg, a crate deleted whole; the status line now says so and points at the PROP-027 family servers that carry the grammar. Ledger at the Phase D exit: 4 476 confirmed / 7 drift / 3 unverifiable — 99.8 % true, from 93.0 % at the Phase C gate. Phase D closed 304 of the 311 rows it opened with. The seven survivors were all one shape — the spec promises a behaviour the code has not built — which is Phase E's material by construction rather than anything stitching could honestly fix. progress check 0 throughout.

• 782026-07-25/26 · PHASE E — the queue drained, and the last drift rows close by being built. Tasks DRIFT-006…022 opened; fourteen executed (006–014, 016–019, 021), DRIFT-015 superseded before it ran, and two (DRIFT-020, DRIFT-022) stood queued at the phase's close — 022 on an owner letter, 020 on a go. Landed: the specmap evidence join with its report column; the lossless-fold check at warning severity (EXPLICIT-BEATS blesses a divergence a document cannot distinguish from a lying fold); the gate panel in campaign.json, which the campaign then dogfooded on itself; baseline invalidation's two missing rules; blockquote fact anchors; the incremental parse path; the --plain and resolver-doc corrections; two #[spec(deviates)] reasons that turned out never to have been deviations; the cache split; the no-op-write skip; and the removal of the legacy ~/.vibevm read leg. Ledger: 4 486 confirmed / 1 drift / 3 unverifiable of 4 490 — 99.9 % true. Findings 61 of 64 resolved. The one surviving drift row (FACT-GRAIN-EVIDENCE) cannot close in this repository at all: it waits on rust-ai-native-lang v0.8.0 re-vendoring the fact-aware specmap engine, which is wave 2's Phase A2. Four new anchors were minted (PROP-042's four --view verbs), each sealed with its own verdict.

79What is worth keeping from this phase is not the count. Four executors volunteered evidence against their own work. Two successive measurements each corrected the previous one's headline — DRIFT-010's ×1.44–2.06 speed-up was a debug profile, and in release the parse is 10.3 ms against 7.5 ms of payload serde; DRIFT-017 then corrected that, because the writes were ~14 % of a warm run rather than the bulk, since comparing requires serialising and only write+fsync are saved. The next performance lever is the serialisation, not the IO. Two others refused to absorb adjacent fixes they had found and filed them instead. The one that mattered most went the other way entirely: F-057 was filed as a stray cache directory and turned out to be a credential-precedence leg — vibe search resolves a GitHub token through four legs and the fourth is a file in the settings dir the test could not reach, so an un-isolated run attached the real publish token to its outbound request. Measured, not inferred, against a loopback listener recording header names and lengths only: 47 bytes of Authorization before the fix, none after. The investigation closed with no rotation needed — exactly one test reaches that path and it points api_base at 127.0.0.1. F-055, F-056 and F-057 were one forgotten discipline caught three times by accident, which is what DRIFT-020 exists to convert into a gate. Two lessons cost real time and are now Constraints in the WAL: a gate never seen to go red is not known to work (two executors ran positive controls before trusting a green result), and delegated work is committed on the completion notification, never on a filled-in task journal — executors write §9 as they go, and committing on it captured an intermediate state and left the tree conform-red for twenty minutes. The floor ends the phase green with no VIBE_SETTINGS override — F-055 genuinely fixed rather than worked around, verified again at this session's open.

10. Deferrals

100(empty — drained into campaigns/<id>/deferrals.md at close-out)

11. REPORT (filled at close-out against §8)

101Filled 2026-07-26. Scored against §8 as written, not as one would like it to read. Where a prediction is scored against a number the campaign itself moved, the number used is the one from before the phase that moved it — otherwise the campaign grades its own homework.

1021. Wave-1 full weave fits ≤2 shards of a 1M window (digest fits trivially). — CONFIRMED, with room to spare — but only because close-out went and measured it. progress weave --max-tokens 1000000 over the 58-file wave-1 corpus emits 1 shard, 1 138 441 bytes (≈ 300 k tokens at prose-markdown density — roughly a third of the window, so the sharder never had to split). weave --digest is 200 454 bytes, ≈ 6 % of a 1M window: "trivially" is right.

103The prediction is confirmed and the process around it is not. No LOG entry records a weave run at any point in wave 1 — Phase B worked from per-file mirrors instead, so nothing in the plan's own steps ever exercised the command, and the number above was produced at close-out purely to fill this row. A prediction that no step of the plan forces you to run is one you will reach the end of the campaign without having tested. Wave 2 should either wire the prediction to a step or drop it.

1042. The exhaustive counter catches ≥1 genuinely skipped paragraph that review alone would have missed. — CONFIRMED, more than once. The sharpest case is a grammar trap no reader would have seen: a wrapped prose line whose continuation opens with + parses as a phantom list item, and the counter caught both occurrences in PROP-019 instantly. The pilot produced a second, independent catch — spec/design/README's index was incomplete, found by the count rather than by reading it. The B-phase LOG scored this prediction in passing at the time, which is the only reason it is cleanly citable now.

1053. ≥60 % of IMPLEMENTED/SHIPPED claims confirm without drift; ≤10 % of all units end unverifiable. — CONFIRMED, both clauses, with wide margin. Scored at the Phase C exit gate, before any stitching: 93.0 % confirmed, 311 drift rows of the 4 455 facts then counted (7.0 %), 3 unverifiable = 0.07 % against a 10 % ceiling. (The corpus reached 4 490 only later, as Phase D and Phase E minted anchors for facts the tree had gained.) The post-stitch figure of 99.9 % is not the answer to this prediction — stitching corrected the drift, it did not verify it away, and quoting it here would be circular.

1064. Stitching converges in ≤3 waves with ≤2 owner escalations. — CONFIRMED; comfortable on waves, exactly at the bound on escalations. Two waves, not three: d1 and d2, the latter run in eight sub-batches (a–h). Exactly two escalations, both genuine and both resolved — F-046's PROP-043 parity rows (wire-or-demote, per item) and F-035's single row in vibevm/vibespecs/boot/00-core.xml, a user-owned file. One row that had been sitting in the escalation pile turned out not to need a ruling at all (BASELINE-RECORD), which is worth reading as a warning rather than a win: it was returned to the owner's desk because nobody had checked the code, and when someone did, the answer was already there. It was also — see prediction 6's note and F-065 — checked wrongly.

1075. ≥80 % of DRIFT tasks land without a returned round-trip. — CONFIRMED at 100 %. Of the twenty-two DRIFT tasks authored in wave 1, none was ever set to returned. DRIFT-015 is the one non-landing, and it was superseded before it ran rather than returned — the conform finding it targeted had been resolved another way. The margin is large enough to be worth distrusting slightly: the template is genuinely carrying enough context, but four executors also volunteered adjacent findings rather than returning, so some of what a returned status would have measured shows up in the findings ledger instead (F-057, F-061, F-063, F-064 all arrived this way).

1086. The month budget holds. — FALSIFIED, in the favourable direction, and only partly scoreable. Predicted: A ≈ days, B ≈ 1–1.5 weeks, C ≈ 1 week, D ≈ 3–5 days. Actual: A on 2026-07-24, B on 07-24/25, C on 07-25, D on 07-25/26, E on 07-25/26 — a month of plan executed in three days, roughly an order of magnitude fast. The honest reading is not "we beat the estimate": the estimate was built on human-paced reading and the work was done by delegated batch execution, so it was measuring the wrong quantity from the start. The "F+G ≈ 1 week overlapping E" clause is permanently unscoreable for wave 1: both phases were deferred to wave 2 by owner ruling on 2026-07-26, and not for want of time — close-out measured their inputs and found them absent (F's three views are empty; G's harvest directory was never written). A phase that cannot start because an earlier phase skipped a step is not evidence about a schedule.

12. Quick-start for the executing session

110Added 2026-07-29 by owner ruling: flow:campaign-plans' ##COLD-A-LITERAL-QUICK-START-BLOCK requires it and this plan had none. Wave 1 is closed out, so this block is a verification quick-start — how a cold session re-measures what the campaign claims, rather than how it resumes it.

cargo run -q -p vibe-cli --bin vibe -- progress report --json --campaign campaigns/progress-2026-08 | head -40
python -c "import json;d=json.load(open('campaigns/progress-2026-08/run/state/findings.json',encoding='utf-8'));print(len(d['findings']),'findings,',sum(1 for f in d['findings'] if f['status']!='resolved'),'unresolved')"
bash tools/self-check.sh; echo "EXIT=$?"       # the gate panel — 0

112The campaign's own claims live in §11 REPORT, scored against §8's predictions; §9's LOG read from the end is what actually happened, batch by batch. Wave 2 — the packages, the sibling corpus — is PACKAGES-ACTUALIZATION-CAMPAIGN-v0.1.xml, and its own quick-start is that plan's §10.

13. Whole-campaign acceptance (recorded retrospectively)

113Added 2026-07-31 under the owner's bring-into-line ruling: flow:campaign-plans' ##ACCEPTANCE-IS-A-RUNNABLE-SCRIPT-ASSERTING-THE-END-STATE asks every campaign for a runnable script asserting its end state, run on a green floor at close and cited by the report, and this plan had none — §11's REPORT scored the predictions without one. The script below was written after the fact and its numbers were re-measured at HEAD fffcb494 on 2026-07-31, not copied from the close-out. Where the two differ, both are shown: a closed campaign whose end state has since improved should say so rather than freeze a figure.

1 — the gate panel. Not re-run in the pass that authored this block;

it is a multi-minute build gate and running it is the executing

session's job, not the drafter's.

114bash tools/self-check.sh; echo "EXIT=$?" # 0

2 — every observed paragraph carries a marker

115./target/debug/vibe.exe progress check --exhaustive \ --campaign campaigns/progress-2026-08 # clean, 0 warnings

→ progress check: clean (260 files, 0 warning(s)) EXIT=0

NOTE: the check covers the whole progress.toml scope, not just this

campaign's 58 files — --campaign selects the state zone, not the

perimeter. Wave 2 shares this gate by design and it must stay at 0.

3 — the host corpus's measured actuality: the campaign's headline

116python campaigns/packages-2026-09/tasks/summary.py | grep '^ host'

→ host (58 files) confi 4496 drift 0 unver 3 total 4499 99.9 %

4 — nothing evaporated: every leftover is a commit or a named deferral

117grep -c '^- ' campaigns/progress-2026-08/deferrals.md # 13, each owned

5 — the recurrence artifact exists, so the next run costs O(delta)

118test -s campaigns/progress-2026-08/baseline.json; echo "EXIT=$?" # 0

119What the acceptance shows, and the one thing it cannot.

• 120The end state is better than the close recorded. At close-out (2026-07-26) the ledger read 4 486 confirmed / 1 drift / 3 unverifiable of 4 490 — 99.9 %. Today it reads 4 496 / 0 / 3 of 4 499 — 99.9 %. The single surviving drift row, FACT-GRAIN-EVIDENCE, closed exactly where deferrals.md said it would and nowhere else: wave 2's Phase A step 2, commit 07a38e1a. The denominator moved by nine as later work minted anchors in host files. A deferral that closes where its own reasoning said it would is the strongest evidence the deferrals ledger is doing its job.
• Three units stay unverifiable and always will here. They are network-bound GitVerse legacy-registry claims; unverifiable was chosen over "probably fine" and the choice is the point.
• What this script cannot assert is the mandate. The owner asked to actualize all the specifications; the script proves the corpus is marked, judged and drift-free, and proves nothing about whether the resulting documents are useful. §11's {#report-gaps} is where that gap is stated: the campaign marked what 4 917 facts are and never what should happen to them, so every forward-looking view came out empty. An acceptance script that passes on an empty view is telling you the view was not part of the contract.
• Step 3 is served by a tool that lives in the next campaign's zone. Wave 1 shipped no summary tool; the number above is printed by campaigns/packages-2026-09/tasks/summary.py, which reads the shared verdict cache. That is a finding, not a convenience: a closed campaign whose headline can only be reproduced by its successor's tooling is one re-run away from being unmeasurable.