Proposed design, unimplemented. Recovery revision 3, 2 October 2026, based on origin/main at 54bb8756084c16c66d9d9870919943097d5d5b9b. This corrects seven findings in Astra round 2 and the material week-history finding in Fable round 2. Fresh Astra extra-high review passed the recovered design and static prototype before issue creation and registration; implementation acceptance remains separate. The original design, issue, prototype and reviews remain unchanged in the Factory archive; provenance records their hashes and source commit.
Nothing here records a level, activates policy, grants authority, schedules work or commissions implementation. Every cell, threshold, budget and default is proposed; [A] marks assumptions. The static policy prototype renders tables and checks static scope rules only. It has no runtime import, package entry, journal or deployment path, and proves no resolver or ledger behaviour.
E42 is registered as proposed, unimplemented scope in issue #82. E39–E41 remain fixed. Current totals are 42 epics, 15 adoption epics and 57–89 reviewed increments, including E42's 7–10. Every epic remains open. The issue specification records the implementation gates.
1. Intent and provenance
One dial per Product governs how much research, inbound work, implementation and marketing the Factory may propose and admit, within existing authority. Owner requests remain a source of work, not a ninth level or a special class. The dial records investment intent, while the separate E39 delivery mode (Prototype/MVP/Product/Scale) guides architecture: neither determines the other.
Original owner request, 2 October 2026, preserved verbatim as historical provenance:
Can you also add another dial on a project for activity level. This should determine how much research and inbound work we generate. For example, a project could be in maintenance mode, where we're not looking to create new work, we're just integrating bug and security fixes and dependency updates. It could also be in low, I guess, where it's the same, but also we are adding key new features if we find out anything about them. Could be none, so there's nothing happening on the project. I guess even in low, we are still doing research, we just have a higher bar for taking work on. It could be in maybe storming, we could call it, where it's we're just integrating lots of new things and building the product out. It could be in a engaged mode where the human overseeing it is creating work as well as the work that we find and we have to research and implement that and they're asking for changes or changing values it could be on scaling So we're looking to scale out marketing and do scale improvements on the product. I guess it could also be in sunsetting where we're winding things down and notifying people. that the product is coming down and each of these should have a different matrix around the activities that are triggered, how regular, how deep and the threshold for accepting new work, the level of investment on implementation and marketing And yeah, I guess those dials control also the level of investment in the project, even if it's early, you might not want to invest lots of tokens in it. So round that out into a, a good epic, think about any other activity types and how best to integrate this across. The modules, as I think it is quite a significant factor, create an issue for it, and also ask Astra and Fable for reviews. Then, when it's ready, Opus 5.5 builds it.
The current takeover commission in AGENTS.md supersedes the quote's historical Opus/Fable assignments: Codex coordinates and Astra extra-high independently reviews. This recovery does not commission a runtime build. Earlier review verdicts remain historical: Astra r2 FAIL; Fable r2 PASS header with one material finding. Section 12 maps all eight corrections.
2. Existing building blocks and planned seams
Source and current register were checked at the recovery base. No activity-level register, resolver, CLI, selection adapter or dial consumer exists. Existing modules are local infrastructure; their source presence is not a claim of new acceptance, live authority or customer value.
| Building block | Existing contract | Proposed activity-level seam |
|---|---|---|
| Project guidance, Portfolio, milestone objectives | Versioned values, Product identities/observed activity, separately scoped owner intent | Caller-owned Product identity; a new register for the level, with no change to those meanings |
| Lifecycle schedule | Occurrence IDs, fixed commissioned weekly ceiling, replay without dispatch, bounded call admissions | AL4 commission and state /2, per-Product sub-ceilings, calendar periods and current-decision fence; preserve /1 history |
| Discovery sources | collect(occurrenceId, brief), basic/full Source rhythm, public HTTPS reads, conditional validators and Capture novelty | AL4 versioned explicit Source selection as well as brief /2 class; neither first-party collection nor private inbox access follows from a class alone |
| Intelligence | Declared tiers and owner-approved model order, named allowances, no substitution | Admitting consumer compares the declared tier with the decision ceiling; selector unchanged |
| Product gate | Grants, per-Product aggregate, fixed Slice allowances and current checks | AL7 future revision coordinated with R15/E40; shared allowance reservations before each gate admission, never a separate per-Product copy of a shared pool |
| Opportunity gate, Planning | G1–G8 and backlog gate are planned | Neutral decision supplies class/evidence requirements, owner-request answer and pins; no consumer maps level names |
| Work items | Recovered local register/lane implementation, no live register | WI1d/AL7 owner-proposed request record and claims after a decision; activity context keys/decision links need their owning versioned contract |
| Owner rules and policy | Owner rules source and ask ledger exist; rules acceptance is not established here | Consult decision before rules/ask; sort queue by decision value; level/cap becomes a standing commission only through later reviewed policy work |
| Owner digest, dashboard | Existing bounded digest and read views | AL5 shared future digest revision with WI1c; read-only level display, no hidden write |
| Branch health, obligations | Branch-health episodes exist; OB1, security/incident/renewal typing remain prerequisites for their classes | Typed obligations only; unassessed operational Signals do not become obligations by a level change |
| Marketing | Steps 1 and 2 are integrated drafts: typed-fact material plus durable job; no live model/publication | AL8 versioned purpose record and supporting pins; present material-only brief cannot enforce purpose |
| Capacity planning | Read-only forecast, discovery ledger coverage only, native units | AL8 observes shared delivery ledger through an adapter; forecast never admits or reserves work |
| Composition catalogue | docs/composition.json owns generated cards/maps; E41 owns map maintenance | AL1a adds its future entry there and regenerates cards/maps; do not hand-edit generated module descriptions |
Existing acceptance details remain in build review. E40 also changes gate admission; AL7 must agree one future protocol with that lane, preserving prior readers and qualification pins. No independent activity /2 revision is presumed to have landed first.
3. Levels and scope
| Level | On a phone (the guide, the set reply, the digest's glossary) | Generates | Does not | Share [A] |
|---|---|---|---|---|
none | nothing runs; security, builds and renewals are watched at no model cost. | Zero-token watches of the floors; at most one open floor item per Product when a floor finding needs a decision; a one-line reply naming the levels that would run what the owner asks for | Research, triage, fixes, marketing, model calls, asks other than floor items | 0% |
maintenance | fixes, security and updates only; no research, no new features. | Security, bug and dependency fixes; broken-build restores; release-path fixes and release records; incidents; compliance and renewals; weekly triage of first-party reports; published claims corrected; the bill reported quarterly | Research, features, debt, design, cost work, promotion | 5% |
low | maintenance, plus research and the odd feature when evidence is strong. | Everything maintenance does, plus weekly light and monthly deep scans; inbound and features at the high bar; announcements of shipped changes; a monthly strategy read-out | Feature work below the high bar; proactive debt, design, cost or analytics work (as at maintenance); campaigns | 10% |
exploring (proposed) | learn cheaply before product fit; research leads, building only validates. | Daily light and weekly deep research, first in its family order; validation-only Slices (prototypes, instrumentation, experiments); conversations prepared; message experiments under owner grants | Debt, performance, localisation, launches, feature Slices before a Supported Assessment | 10% |
storming | build it out; research and inbound fill the backlog. | Daily and weekly research; features first in its family order at the standard bar; weekly debt; launch material at launches | Sustained promotion (scaling's purpose); spend without an owner grant | 35% |
engaged | storming, led by you; its questions reach you first. | Exactly what storming generates: the same cells, budget, family order and purposes (static equality rule). The owner leads: the Product's items rank first in the owner queue (§9), and from AL7 the digest notes four weeks without a recorded request from the owner | Anything storming does not | 35% |
scaling | grow it; marketing, analytics, performance and support lead. | Grow first in its family order: promotion, analytics and A/B tests, support; performance and scale; features only above the high bar; monthly cost work | Broad new feature work | 30% |
sunsetting | wind it down under your approved notice and data plan. | The approved plan: notices, export, migration help, support; security until shutdown; scale-down; staged removal; closure | Research, features, debt, promotion | 5% |
At every level above none, a request from the owner is admitted first wherever its kind runs and the class or marketing purpose is admitted (§4.2).
exploring is the proposed extra position. launching stays a release workflow, hardening is a possible later policy revision when a real Product needs it, paused means none plus a review date with no automatic restart, and retired means none after a recorded closure. ticking-over is an offered alternative spelling for low, not an extra position.
Unset is not a level. Revision 0 is unset, has no zone of its own and an empty budget. It answers unset; only the five floors watch at zero model cost, with at most one sourced owner item per Product. An owner request is recorded and answered with the scope it would need. Nothing else runs. A first mid-week level therefore gets ordinary budget only next Monday. Default Europe/London is proposed until the first set records the Product's fixed zone.
The dial never removes verification of admitted work, settlement/reconciliation of unknown effects, stop intent, revocations, journals, receipts, passive views or floor detection. A dated obligation remains recorded and watched at every level. Already admitted commitments settle under their original grant; a date alone grants no fresh work or funding at none or unset.
4. Matrix and evidence
4.1 Cell vocabulary and cadence
A cell is mode · cadence · tier · bar. — is off; W watches with deterministic reads/reports only; R reacts to an identified trigger; O also originates work from periodic Factory sweeps. Watch never admits a work item or model call; its sourced owner exposure item is reporting, not work admission. Inbound triage always means first-party scope; a private support channel still requires a separately commissioned adapter.
Cadences are e event, d local day, w ISO week starting Monday 00:00, m calendar month, q calendar quarter, all in the recorded Product zone. A periodic Factory-originated sweep/read-out starts at most once per period, catching up in an eligible slot of that period and coalescing when it ends. periodOf gives every consumer the same key. Individually identified typed obligations are event-driven, with duplicate refusal per obligation, grants, WIP and funding; a monthly compliance sweep never prevents a second obligation that month. Watch cadence bounds routine reporting only. Urgent findings immediately update or reopen the Product's one sourced owner item, within existing notification policy, without waiting for a reporting period. At none or unset this still starts no work.
Tier N permits deterministic checks/prepared owner steps only, no model or produced Candidate; E is an efficient ceiling; F permits frontier work within grants. The selector is unchanged and efficient tasks still need their measured qualification. A ceiling never authorises a model or substitutes a cheaper one. Watch, purpose-governed marketing and read-out-only strategy have no bar.
4.2 Classes, bars and owner requests
Class-governed admission is a kind/class pair. The kind must run, produce that class, and its bar admit it. Every class retains its own typed trigger record. The proposed classes are obligation, defect, regression, dependency, upkeep, optimisation, validation, opportunity and improvement. Defects use reproduced/first-party failure evidence; regressions cite a measured baseline; upkeep changes no user capability; optimisation cites a bill/usage baseline and revert path; validation is a bounded learning Slice; opportunity is a feature Slice; improvement is debt/design/performance/localisation beyond a regression.
| Bar | Admitted classes | Additional feature/novelty evidence |
|---|---|---|
| B0 | obligation | Typed obligation source required; never waived |
| B1 | B0, defect, regression, dependency, upkeep, optimisation | Each class's trigger/source remains required |
| B2 | B1, validation, opportunity | For validation/opportunity: score ≥1.5 × normal threshold, ≥2 customer-voice Signals, current Objective; web-only evidence admits validation only [A] |
| B3 | B1, validation; opportunity only after Supported Assessment | Normal G1–G5 evidence for validation/opportunity and Supported Assessment pin before feature implementation |
| B4 | B1, validation, opportunity, improvement | Normal gate threshold for validation/opportunity/improvement |
B2 and B3 are alternatives for a shipped Product and one before fit, not an ordered scale. Feature thresholds never apply to obligations, defects or regressions. The neutral decision carries resolved per-class requirements: applicable class, threshold values, required source/pin kinds and whether owner origin waives a requirement. The consumer never imports another policy mapping.
Kinds produce these classes (the prototype is the static source of the rendered matrix): discovery light/deep produce opportunity/validation; inbound produces obligation/defect/regression/upkeep/validation/opportunity; five floors and sunset steps produce obligation; bugs produce obligation/defect; dependency updates produce obligation/dependency; release operations obligation/defect/upkeep; cost obligation/optimisation; feature work opportunity/validation; tech debt improvement; performance regression/improvement; design validation/improvement; accessibility defect/regression/improvement; documentation and deprecation obligation/upkeep/improvement; analytics validation; support obligation/upkeep/validation. Marketing uses purposes; strategy has no work-item class.
Dependency routes. Deterministic patch/minor updates within policy are dependency-updates/dependency. A non-security major needing implementation is tech-debt/improvement, admitted at B4 in storming, engaged and scaling. A security remediation is security-fixes/obligation with typed exposure and that row's frontier ceiling. A failed bump becomes a defect or typed broken-build obligation. Dependency-updates' obligation subclass is only a bounded deterministic update required by a typed plan/obligation; it never relabels a routine major. Owner requests do not bypass these classes.
Obligation sources. Branch-health episodes exist today. OB1 legal/platform/store deadlines, CF4 renewal/rotation records, E11 security findings, E13/E38 assessed incidents, typed Marketing revalidation and AL2b sunset plan steps remain prerequisites. A Signal alone is unassessed; no source means no admission of that obligation type. The level never manufactures one.
Owner request contract. Record the owner's words in a work item with proposedBy.kind: owner, and either {governedBy: "class", activity, class} or {governedBy: "purpose", activity: "marketing-growth", purpose, supportingPins}. A changelog request is release-record, never a fabricated class. The neutral answer is first, validation-only, or reply with at most two eligible levels and required transition/plan conditions. At exploring a feature request becomes cheapest validation until Supported. From sunsetting a reply explains withdrawal; entering sunsetting needs a plan.
Owner origin waives only score, customer-voice and Factory-originated novelty requirements (G3's new-Signal test, G4 and G8 evidence-class restriction). It keeps source/obligation pins, Objective/shape, provenance where Signals are cited, Supported Assessment, measurement, duplicate refusal, backlog cap, purpose/material/supporting pins, grants, WIP, tier and funding. Admitted requests queue after floor obligations and before Factory-originated work. Refused requests are recorded with one plain reply; silence never changes the level. T1/T4 quiet counts are unknown unless owner-request records cover the whole window.
4.3 Proposed 192 cells
Status describes building blocks only, as checked on the recovery base: built does not mean the activity-level feature exists. Four rows have an underlying module. In particular inbound's first-party class and exact selection await AL4; nobody is triaging private support messages through this collector.
| Activity (family) | Status | none | maintenance | low | exploring | storming | engaged | scaling | sunsetting |
|---|---|---|---|---|---|---|---|---|---|
discovery-light sense | built | — | — | O·w·E·B2 | O·d·E·B3 | O·d·E·B4 | O·d·E·B4 | O·d·E·B4 | — |
discovery-deep sense | built | — | — | O·m·F·B2 | O·w·F·B3 | O·w·F·B4 | O·w·F·B4 | O·w·F·B4 | — |
inbound-triage sense | built | — | R·w·E·B1 | R·w·E·B2 | O·d·E·B3 | O·d·E·B4 | O·d·E·B4 | O·d·E·B4 | R·d·E·B0 |
security-fixes (floor) keep | planned | W·e·N | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 |
bug-fixes keep | planned | — | O·e·F·B1 | O·e·F·B1 | R·e·F·B1 | O·e·F·B1 | O·e·F·B1 | O·e·F·B1 | R·e·F·B0 |
dependency-updates keep | none | — | O·w·E·B1 | O·w·E·B1 | O·m·E·B1 | O·w·E·B1 | O·w·E·B1 | O·w·E·B1 | R·e·E·B0 |
build-health (floor) keep | built | W·e·N | O·e·F·B0 | O·e·F·B0 | O·e·F·B0 | O·e·F·B0 | O·e·F·B0 | O·e·F·B0 | R·e·F·B0 |
incident-response (floor) keep | planned | W·e·N | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 | R·e·F·B0 |
compliance-licences (floor) keep | planned | W·m·N | O·m·E·B0 | O·m·E·B0 | O·m·E·B0 | O·m·E·B0 | O·m·E·B0 | O·w·F·B0 | O·e·F·B0 |
renewals-access (floor) keep | planned | W·e·N | O·e·N·B0 | O·e·N·B0 | O·e·N·B0 | O·e·N·B0 | O·e·N·B0 | O·e·N·B0 | O·e·N·B0 |
release-operations keep | planned | — | R·e·F·B1 | R·e·F·B1 | R·e·F·B1 | R·e·F·B1 | R·e·F·B1 | R·e·F·B1 | R·e·F·B0 |
cost-optimisation keep | planned | W·q·N | W·q·N | W·q·N | O·q·N·B1 | R·q·E·B1 | R·q·E·B1 | O·m·F·B1 | O·e·E·B0 |
feature-work build | planned | — | — | R·e·F·B2 | O·e·F·B3 | O·e·F·B4 | O·e·F·B4 | O·e·F·B2 | — |
tech-debt build | planned | — | — | — | — | O·w·F·B4 | O·w·F·B4 | O·w·F·B4 | — |
performance-scale build | planned | — | R·e·F·B1 | R·e·F·B1 | — | R·e·F·B4 | R·e·F·B4 | O·w·F·B4 | — |
design-ux build | planned | — | — | — | O·m·F·B3 | O·m·F·B4 | O·m·F·B4 | O·m·F·B4 | — |
accessibility-localisation build | planned | — | R·e·F·B1 | R·e·F·B1 | — | O·m·F·B4 | O·m·F·B4 | O·m·F·B4 | — |
documentation build | planned | — | R·e·E·B1 | R·e·E·B1 | R·e·F·B3 | O·e·F·B4 | O·e·F·B4 | O·e·F·B4 | O·e·F·B0 |
marketing-growth grow | planned | — | R·e·N | R·e·N | R·e·E | O·w·E | O·w·E | O·w·E | O·e·E |
analytics-validation grow | planned | — | W·m·N | W·m·N | O·e·N·B3 | O·e·N·B4 | O·e·N·B4 | O·e·N·B4 | W·w·N |
user-comms-support grow | none | — | R·e·E·B1 | R·e·E·B1 | O·e·E·B3 | O·e·E·B4 | O·e·E·B4 | O·e·E·B4 | O·e·E·B0 |
deprecation-cleanup retire | planned | — | R·q·F·B1 | R·q·F·B1 | O·m·F·B3 | O·m·F·B4 | O·m·F·B4 | O·m·F·B4 | O·e·F·B0 |
sunset-notices-export retire | planned | — | — | — | — | — | — | — | O·e·F·B0 |
strategy-review govern | planned | W·q·N | O·q·N | O·m·N | O·w·N | O·w·N | O·w·N | O·w·N | O·w·N |
low equals maintenance except research, inbound's bar, feature work and strategy cadence; marketing additionally permits announcements. Engaged equals storming except the owner row. Tier-none on-cells admit deterministic Slices/prepared owner steps (a plan revert with read-back or renewal runbook), never a model-produced change. documentation upkeep is a reported correction; documentation for a code change still ships with that Slice. Analytics read-outs are deterministic; instrumentation is a validation Slice. Release administration reacts to defects/upkeep, with only typed sunset-plan obligations at sunsetting; release-workflow failures remain build-health episodes.
4.4 Proposed budget and permission rows
| Budget, permissions and owner row [A] | none | maintenance | low | exploring | storming | engaged | scaling | sunsetting |
|---|---|---|---|---|---|---|---|---|
| Token share: a ceiling on the Product's own use of each owner-signed weekly ceiling | 0% | 5% | 10% | 10% | 35% | 35% | 30% | 5% |
| Open Slices (proactive WIP) | 0 | 1 | 2 | 1 | 4 | 4 | 3 | 2 |
| Largest Slice allowance class | — | small | standard | small | large | large | standard | small |
| Spend the Factory may propose | never | obligations | obligations | propose | propose | propose | propose | obligations |
| Family order after floors, owner requests and Keep | — | Sense, Build, Grow, Retire | Sense, Build, Grow, Retire | Sense, Build, Grow, Retire | Build, Sense, Grow, Retire | Build, Sense, Grow, Retire | Grow, Build, Sense, Retire | Retire, Grow, Sense, Build |
| Marketing purposes, with the word the guide uses | none | accuracy: correction, release-record, capability-retirement | announce: correction, release-record, announcement, capability-retirement | experiment: correction, release-record, announcement, experiment, capability-retirement | campaign: correction, release-record, announcement, experiment, launch, capability-retirement | campaign: as storming | growth: correction, release-record, announcement, experiment, launch, promotion, capability-retirement | wind-down: correction, sunset-step |
| Owner queue rank (unset ranks 9) | 8 | 6 | 5 | 4 | 2 | 1 | 3 | 7 |
| Owner-led | no | no | no | no | no | yes | no | no |
| Quiet clause after (weeks, T1) | — | 26 | 8 | — | 8 | 8 | 8 | — |
| Owner-quiet clause after (weeks, T4) | — | — | — | — | — | 4 | — | — |
Budget rows take the weekly minimum. Family order, purposes and owner row are current permissions. Work order is floor obligations, admitted owner requests, Keep, then the listed families, oldest first within each. Govern uses no model tokens. Shares cap a Product's own use, reserve nothing and are never lent. The portfolio ceiling remains binding even if shares sum above it. Slice classes are proposed ceilings on existing GrantLimits: small 2 production/1 integration, standard 4/2, large 8/4 [A]; grants can be narrower. Spend class permits proposals only, never spending. Minutes/asks are measured after answering, not pre-admission currency.
4.5 Marketing purposes
Material and purpose are separate. AL8 adds a bounded versioned purpose record to the existing durable job, with vocabulary agreement tests against the neutral contract. The current material-only brief has no such permission. Current revision purposes are sets, not ordered budget values.
| Purpose | Materials | Supporting pins it must carry | Permitted at |
|---|---|---|---|
correction | The corrected pack's own material | The published pack's digest and the validity read or fact revision that invalidated it; it may only correct or remove | every level above none |
release-record | release-notes, changelog | The release record of a shipped change | maintenance to scaling |
announcement | release-notes, changelog, site | The release record of a shipped user-visible change | low to scaling |
experiment | site | The owner's grant and the experiment record (E21) | exploring to scaling |
launch | launch, site | The launch item (L11a) and the owner-granted release | storming, engaged, scaling |
promotion | site, launch | An owner-granted campaign; paid proposals only under spend class propose | scaling |
capability-retirement | deprecation, changelog | The owner's approved Retiring disposition for the Capability | where deprecation-cleanup runs, except sunsetting |
sunset-step | deprecation, changelog, site, release-notes | The approved sunset plan revision and its step; after a withdrawal, the withdrawal record | sunsetting, and the withdrawal notice after it (§9) |
Every material requires a permitted purpose/material pair and all supporting pins. At none nothing is drafted; stale public claims become the Product's sourced owner exposure item. Drafting never publishes: publication remains a separately granted Action and Receipt. A sunset plan grants no publication itself.
4.6 Static policy rules
The inert prototype checks exactly eight levels ×24 kinds; closed vocabularies; watch has tier none/no bar; none never works; five floors watch at none and at least react with B0 elsewhere; reporting no less frequent at none than maintenance; sunset and maintenance exclusions; low's limited differences; engaged/storming equality except owner row; bounded/nested budget vocabularies; every enabled class-governed cell has a produced class; marketing purpose scope. It also exercises all advertised dependency subclasses across every level, with a mutation proving the wrong major route is caught. Future AL1a adds pinned policy digest, complete schema validation and table parity to the reference. A prototype pass establishes only these static checks.
5. Register, time and replay
The proposed sf-activity-set/1 carries Product, expected revision, level, policy digest, fixed recorded zone, resulting limits, explicit keep/set/clear operations, verbatim owner words (1–280 characters), recorder, stable request ID, and later proposal/plan pins. Omitted caps/limits are kept, not cleared. A saved request resolves operations against its expected revision and the register verifies the result; a moved head is STALE, never rebased. Only an explicit cap-clear changes one. A requested open-Slice limit above the new level is RULE; a retained larger limit stays recorded but has no effect. Attribution is not authentication: the live journal requires an owner-trusted host.
Each accepted set appends an immutable revision, including content-identical sets. Aggregate activity-level:product:<Product> has version=revision; revision 0 is unset. A digest covers all fields and effective instant. Command ID activity-level/set/<Product>/<24hex> uses SHA-256 of canonical [Product, requestId]; with 63-character Product ID it is 107 characters within the journal's 128. Changed payload under one ID conflicts. All derived verbs follow the bound.
Refusal order: grammar/size; original receipt replay or payload conflict; then for new requests only expected revision, registered Product, current policy, level/zone/plan rules, resolved-limit consistency, stored-history integrity, effective time and bounds. Replay validates the recorded result under its own protocol; it is never recomputed under today's policy/clock, registration or limits. Thus policy upgrade, withdrawn Product and a full register cannot invalidate an original receipt.
Effective instant: sample the caller-owned clock inside the journal decision after acquiring the transaction lock, as the schedule does. A pre-lock Sunday timestamp must not turn a Monday commit into a pre-week raise. Refuse a time earlier than the latest committed revision (RULE: clock-before-history); equal instants are ordered by revision. Keep the original instant on replay. Invalid clock values refuse before writing. Bounds are 500 revisions per Product, 64 revisions in force in the effective week (65th LIMIT), 16,384 canonical request bytes, history pages 1–20, caps ≤10,000,000,000, open-Slice limits 1–16, review date ≤731 days. Refuse, never truncate.
Permissions apply at the recorded instant to the next admission. Ordinary token budget, open Slices, Slice class and spend class take the week's minimum. Lowering never interrupts admitted work or blind-retries an unknown effect. Frozen acceptance/allowance stays pinned. New production reads the current fence; verification and integration of already sealed Candidates retain their grant checks. Raising never revives parked work by itself. A Product keeps its pinned policy until the owner's explicit next set; no silent repin.
6. Pure decision and current fence
Proposed API, not an exported contract:
resolveActivity({ productId, activity, asOf, current, revisions, ceilings? }): ActivityDecision;
periodOf(cadence, asOf, recordedZone): string;
digestClauses({ productId, asOf, revisions, facts }): DigestClauses;
// current: trusted caller/port's current ActivityPin; pure code cannot discover an omitted newer head.
// revisions: contiguous through current, starting at revision 1 or a revision at/before Monday;
// explicit revision 0 when unset. The resolver derives implied unset time and refuses missing coverage.
// ceilings: owner-signed weekly values and scope, echoed in pins.week, never permission by themselves.
// Decision: answer (may-run/watch-only/off/unset), mode, cadence, period, tierCeiling,
// bar {code, classes, afterSupported, evidence: resolved per-class requirements} | null,
// ownerRequest:
// {governedBy: "class", answers: per produced class, evidenceWaivers} |
// {governedBy: "purpose", answers: per purpose, requiredSupportingPins, evidenceWaivers: []} |
// {governedBy: "none"},
// answer: {admit: "first"} | {admit: "validation-only"} |
// {admit: "reply", levels, transitionRequirements},
// budget: {tokens: per allowance known | unknown-with-reason-and-atMost, openSlices, sliceClass, spend},
// hardOwnerCaps: per allowance {state: "absent", revisions} | {state: "bounded", tokens, revisions},
// floor: {wip, sliceClass, reservePerProduct, trigger: "typed-obligation", cadence: "event", dedupe: "obligation-pin"} | null,
// purposes, familyRank, owner {queueRank, ownerLed}, pins {register, budgetRevisions, week, policy},
// displayLevel (display only), digest.
There is no caller-selected zone: snapshots must agree on the Product's fixed zone. The neutral schema, parsers and frozen digest-bound decisions expose every value a consumer branches on. Policy/core are the only mapping. Consumers import only the neutral contract and a caller-owned currentActivity port, never map display words to limits. Future graph/import checks and a distinctive-level-name/displayLevel scan prevent re-derivation; allowance/material vocabularies use agreement tests, not runtime imports that couple modules.
Admission fence: read a fresh Product/activity decision immediately before admission; parse and check scope, class or purpose, evidence, grants, budget and a consumer-owned revision high-water mark. Pass dynamic context outside the command payload so receipt replay stays stable. Pin the admitted decision. Immediately before the effect, reread the current register revision; if changed, mark not-sent with no usage. Even a content-identical new revision invalidates the old pin deliberately. A change after this final check applies at the next admission; started work settles. Unavailable/corrupt ports refuse. Crash between admission and effect leaves unknown charged at its bound, reconciled without dispatch; original admission replay always says dispatch false. Lowering requires no new grant; enlargement still requires the existing signed upper bound or a new owner commission.
7. Weekly budgets and shared reserve
7.1 Complete history window
The resolver derives Monday 00:00 through asOf from recorded times and Product zone. Snapshots must be contiguous and start at revision 1 or a revision recorded at/before Monday; a later-revision suffix starting mid-week refuses INVALID: week-not-covered. If revision 1 first occurs mid-week, include implied revision 0's empty ordinary budget before it. Any future snapshot refuses INVALID: revision-after-asof; backwards times refuse; equal times use revision order. The run must end at the trusted caller/port's supplied current pin. This head input is a trust requirement, not proof that a pure function can discover missing history independently. The resolver ignores nothing supplied and never guesses a gap.
7.2 Ordinary budget and separate hard cap
For each allowance, each revision in force in the validated window contributes: zero for none/unset; otherwise the minimum of floor(signed weekly ceiling ×level share) and a present owner cap, if both exist; the share-derived ceiling alone; the owner cap alone if that is the only signed ceiling; or unknown if neither exists. The weekly ordinary budget is their minimum when all are known. Any unknown keeps the answer unknown with a reason and smallest known atMost for display only; zero is not substituted and atMost never admits. Open Slices, class and spend use the same weekly minimum. Supplied ceiling values/scope live in pins.week; their enforcing owner fixes them when the week opens and never raises them during it. Discovery still needs its S1 commission even if a Product cap is known.
The decision independently returns each effective hard owner cap: absent if no revision in the validated week carries one, or bounded by the minimum of all present caps with their revision pins. Absence contributes no restriction, not zero or unknown. A newly set cap lowers immediately; a raised/cleared cap cannot increase the hard limit until next Monday. This calculation does not depend on level shares. After none→maintenance the ordinary budget may be zero with either an absent hard cap or an existing bounded one. All relevant ordinary and reserve-funded Product usage counts against the effective hard cap; unknown effects count at their maximum admitted bound.
No lending, carry-over or top-up is permitted. A raise changes current permissions but only next week's ordinary budget. The reply names the next Monday's actual date. Exhausted proactive work waits; unknown budget refuses budget-unknown. Owner minutes are measured, never converted to tokens.
7.3 Event-driven floor path
At a level above none, a typed obligation whose kind/class runs may use one open floor Slice per Product, class small, with at most half the shared reserve per Product per week [A]. It does not consume proactive WIP. Distinct obligations wait in deadline order for the slot; completing one permits the next in the same month without another sweep. Duplicate obligation pins cannot create a second dispatch. No level slows detection. At none/unset, an urgent finding updates/reopens the one sourced owner item immediately, but starts no work.
Funding first uses known unspent ordinary Product budget, then an owner-signed shared floor reserve (proposed 15% of the delivery ceiling), withheld from the ordinary pool. It adds no capacity. A cap, missing reserve, exhausted reserve or exhausted Product reserve share produces one owner item naming the missing resource. The reserve never bypasses a hard cap.
7.4 One authoritative shared allowance ledger
The enforcing ledger is one shared transaction owner, separate from per-Product gate aggregates and capacity forecasts. It serialises all ordinary and reserve delivery reservations under portfolio/account/commission identity, allowance/native unit and ISO accounting week in the commission's recorded zone. Product zones must match; zone mismatch refuses, never converts. It atomically checks total ceiling, ordinary pool after reserve, Product ordinary budget, effective hard owner cap and per-Product reserve share. It never trusts caller-supplied remaining balances.
A narrow caller-owned contract provides idempotent reserve(request), settle(receipt, observedUsage) and release(receipt, provedNotSent). The bounded stable request identity/canonical digest binds ledger scope/week, Product, exact gate/admission ID, grant pin, activity-decision pin, obligation source pin for floors, maximum usage by allowance and funding split. Changed payload under one ID conflicts. A digest-bound receipt cannot be reused for another Product, gate, grant, week or admission. The gate requires that matching receipt before admitting; the current fence and authority still apply. One effect consumes one reservation; multi-allowance debits are checked in the same ledger decision.
Two stores are never one transaction. A crash after reservation, gate admission or dispatch leaves the full bound charged until reconciliation proves the outcome. No timeout reclaims it; an unknown effect cannot be sent again. Release requires durable refusal/not-sent proof or an atomic cancellation fence that makes this admission permanently unable to dispatch. A missing gate receipt or elapsed time is insufficient. A settled known usage releases only its proved unused portion; it cannot enlarge the original allowance. Settlement/release each use stable IDs so crashes/retries cannot refund twice. Gate replay returns original receipts with dispatch false. At a week boundary, an unused reservation cannot fund a new week's admission; unknown work stays attributed to its original week until reconciled, with no rollover credit.
A one-time cap exception, if the owner chooses one, is a new separately signed fixed Budget/grant, explicitly naming Product, obligation, allowances and authority to exceed the standing cap. Record/debit it separately in the same ledger; it cannot reset the standing week, replace an unknown reservation or reuse an admission. An ordinary grant without explicit cap-exception authority is insufficient. No such exception is granted here.
7.5 Forecast boundary and units
| Bound | Allowance/unit | Enforcement | Forecast |
|---|---|---|---|
| Discovery | discovery-window / discovery-ledger-token | Schedule per-Product ledger at admission, under S1 | Existing discovery adapter; activity budget beside headroom |
| Delivery ordinary | slice, product, improvement-window / proposed estimated delivery tokens | Shared allowance ledger before gate admission | Future delivery adapter only |
| Delivery reserve | Signed covered delivery allowances / same delivery unit | Same shared ledger; proactive work cannot draw it | Unspent reserve as incident reserve, beside already charged usage |
A forecast grants/reserves nothing. Missing delivery coverage is unknown, not zero. The reserve is shown only once (unspent remainder plus usage already charged), never in discovery's different native unit. These Factory limits do not reserve provider subscription capacity; other owner sessions remain outside ledger coverage.
8. Discovery selection and cadence integration
AL4 proposes sf-schedule-commission/2 and sf-lifecycle-schedule/2 state, both with /1 readers. A Factory adapter derives per-Product lanes, rhythm days, calendar cadences, scope, pins and sub-ceilings from decisions; the signed commission remains the upper bound. Product zone must equal commission zone. Lowering needs no new signature; whether a raise may be rederived within an existing signature is a later owner decision in S1. Per-Product budgets/high-water marks and latest-full period live in state; fixed occurrence IDs and F1–F6/coalescing/replay remain.
Daily/weekly/monthly rhythms use their anchor night; a missed monthly anchor can run on a later eligible anchor in the same month, never catch up a past month. Full IDs stay full/ISO-week; state separately records calendar period. A cadence change uses the recorded instant and periods already run. Each Product has at most one occurrence a night.
Source class alone is insufficient. Brief /2 adds optional first-party/public class; /1 is unclassified, never implicitly first-party. Proposed collection /2 takes a versioned sf-discovery-selection/1 input: exact bounded Source IDs, due activity purposes, commissioned brief digest, approved rhythm selection and activity pins. It validates the union of first-party Sources for inbound, basic Sources for light and all Sources for deep. Each Source is fetched once. Inbound explicitly selects first-party Sources marked full on a basic maintenance night, within the signed selection scope and unchanged occurrence Budget. Selection outside the commissioned brief/approved purpose or wider than its upper bound refuses before fetching. A full deep occurrence replaces the basic one as today.
The selection digest is part of the collection claim, result and replay identity. Same occurrence plus identical selection replays; changed selection conflicts before fetching. Version every affected claim/result family together, retaining the old /1 reader and original rhythm semantics rather than silently acquiring selection. A first-party class is ownership/relevance metadata, not credentials or private-channel permission: the collector remains public HTTPS only. Private inboxes or owner-relayed feedback need separately commissioned adapters/inputs.
Conditional GET can re-emit cached Captures on 304. Persisted Capture identity and novelty (repeat/unchanged) plus consuming triage suppress duplicate work, not the status code itself. Overlapping due activities share one selection/result.
9. Proposals, owner surfaces and sunset
The core computes quiet clauses from complete supplied windows; absent coverage is unknown, never quiet. Proposed thresholds are 8 weeks for low/storming/engaged/scaling, 26 for maintenance, and 4 owner-quiet weeks for engaged. Triggers are quiet read-out, repeated budget pressure, a Supported bet outcome/stop, owner quiet, floor exposure, review date, proved sunset closure and policy revision. Quiet clauses are digest text, not interruptions. Level proposals are owner-required asks with attributed source evidence, one open per Product and at most one offered per four weeks (floor exposure exempt), inside the five-per-week owner bundle [A]. Unanswered is offered once more then parked; decline suppresses the same evidence for eight weeks. Silence never sets a level.
Queue order uses decision values: dated floor obligations, sunset obligations, Product queue rank, then oldest. No level-name mapping lives in the owner-policy consumer. Urgent exposure updates/reopens an existing sourced item immediately; it does not grant unbounded notifications or change the level. KR1 continues measuring owner minutes and asks with source coverage.
AL5's bounded digest line shows name, current health, level and at most one cited clause: recent change (explicit next-Monday date), proposal, quiet, owner quiet or work performed. Unset reads “level not set”; compaction drops the clause before the level. Dashboard is read-only. A future guarded set command is AL9, outside the estimate. Set CLI later saves a new 0600 request exclusively before execution, uses stable saved IDs on uncertain retries, refuses missing journals/irregular or oversized files, and opens read verbs read-only. It never creates a live journal implicitly.
Entering sunsetting requires an owner-approved immutable plan: audience/counts (unknown allowed), channels/notices, export window and method, retention/deletion including backups, account closure, alternatives/refunds as owner decisions, named Services/credentials (never values), shutdown/recovery and evidenced closure criteria. Proposed minimum notice is 30 days [A], subject to actual owner/legal requirements at commissioning. Before AL2b, set sunsetting refuses. Every notice is an owner-granted Action with a Receipt; plan approval alone grants none. Exit to none requires closure evidence or explicit owner acceptance of each gap. Exit to another level is an explicit withdrawal, preserving sent notices and permitting only its required withdrawal notice via its pinned record. Remaining dated obligations retain their original evidence/authority rules.
Proposed defaults, historical rationale only: Factory engaged, Tally exploring, JustSpeakToIt low, Pasta/Family Watch/Focus Sentinel maintenance (shares total 70%). These came from 29 September bet evidence and roadmap 85a5472, not a current release/health check. Refresh that evidence before presenting choices; no default is applied here. The four later owner choices are levels/wording, zero-token floors at none, each Product's default and share/reserve sizes. Q11 signature semantics stays an owner choice inside S1; sunset notice terms wait for an actual sunset. Europe/London and other thresholds are labelled proposals, not authority.
10. Proposed increments and acceptance
All are unimplemented. Independently reviewed source and meaningful normal/fault evidence must precede acceptance; historical receipts are immutable. Codex implements under a later bounded commission, Astra extra-high reviews independently and the coordinator accepts. Serialise final full checks; do not treat this prototype run as module acceptance. Order: AL1a →AL1b →AL5 →AL4 →AL7 →AL2a/AL3 →AL6 →AL8; AL2b when a sunset/outcome/new policy needs it. Owning-lane work counts once, not twice.
| Increment | Proposed scope | Required proof beyond ordinary parser tests |
|---|---|---|
| AL1a pure core | Neutral activity contract, policy data, pure resolver/period/quiet functions; file-driven matrix/resolve CLI, no journal/consumer | All 192 cells; all advertised subclasses across levels; static rule mutations/table/digest parity; neutral-only consumer for B2 feature and owner changelog purpose; evidence waivers never remove obligation/purpose pins; week coverage/future/current-head refusals; first Wednesday revision includes unset zero; hard cap absent/present/exhausted/raised/cleared; DST/ISO week/zone cases; unknown never zero |
| AL1b register | Caller Journal/identity/clock; bounded immutable set/current/revision/history/inForce/resolvePin/decide; saved request CLI | Sunday→Monday lock wait timestamps inside transaction, clock rollback refusal, equal-time revision order; replay after policy change/withdrawal/limits; concurrent expected-version races; 65th weekly revision refused; explicit keep/set/clear and stale resolved request; corrupt state/event/receipt refusal; bounded IDs/files; no live journal |
| AL5 read surfaces | Shared future digest revision with WI1c, quiet coverage, read-only dashboard | Digest word bound, citations, unknown coverage, unset not none; compaction retains level; no consumer level mapping |
| AL4 discovery | Commission/state /2, current fence, sub-ceilings, explicit collection selection and brief class | Mixed public/first-party maintenance; first-party full Source selected on basic night; overlapping due activities fetch once; 304 novelty avoids duplicate triage; changed-selection retry conflicts; /1 history replay; month/ISO/DST anchors, missed periods, zone mismatch; lowering with stale commission, unavailable port, current-check races, unknown/replay never dispatch |
| AL7 admission/ledger | Coordinated R15/E40 gate revision; authoritative shared ledger reservation/settlement/release; opportunity gate and work-item claims/owner requests; T4 | Grant required; current fence including content-identical revision and stale high-water mark; 3 concurrent Products never exceed reserve/per-Product limits; hard cap independent of share and charged by ordinary+reserve; cap raise/clear unchanged this week; crashes around reservation, gate, effect, settlement/release; conflicting/replayed/cross-Product receipts; no timed reclaim; two distinct same-month obligations and duplicate refusal; neutral-only class/purpose answers |
| AL2a first proposals | Floor exposure and review-date proposals; immutable create-once records | Complete source windows; one urgent item immediately updated/reopened, same-source deduplication; bounded offers, decline suppression, silence inert |
| AL3 external adoption | Standalone package with caller-owned Journal, identities, clock and substituted consumer port | Packed consumer, actual resource closure, corruption/restart/replay/isolation/fence proof through documented exports only |
| AL6 owner queue | Decision before rules/asks, queue by neutral rank, attributed per-level minutes/asks | No level names in consumer; floor first, engaged-before-storming via rank only; overflow/unknown coverage preserved; rules get no level scope |
| AL8 other consumers | Branch-health list/guardrails, renewal sources, delivery capacity adapter, versioned Marketing purpose admission, repair | Owning contracts preserved; forecast missing coverage unknown/no authority; permitted owner changelog needs shipped-release pin; promotion refused at maintenance/sunset; correction allowed with pins; no purpose/missing pin refused; launch stops on sunset; withdrawal notice pinned |
| AL2b sunset/remaining proposals | Plan, closure, withdrawal, outcome/budget/policy proposals | Invalid/incomplete plan refusal; each closure criterion proved or owner-accepted; sent notices immutable; policy revision only by explicit set |
| AL9 optional | Guarded dashboard set | Stable request/unknown retry protocol; separate future scope, outside estimate |
AL1a source candidates are src/activity-level-contract.ts, src/activity-policy.ts, src/activity-level-core.ts, scripts/factory-level.ts; AL1b adds src/activity-level.ts. They do not exist in this recovery. Future touch-points include Intelligence's no-model declaration, observability protocols, graph/neutral-import checks, reference/guide/how-to documentation, documentation map and generated site. Add the module to docs/composition.json then regenerate cards/maps. Artwork remains pending unless separately commissioned. No current package export, CLI entry, qualification pin, source receipt, grant or operating-policy file is changed here.
11. Registration plan and limits
Proposed E42: Govern each Product's activity and investment through one owner-set level. Kind adoption, work type new-module, prerequisite E28; related E14/E21/E22/E23/E24/E25/E26/E27/E29/E32/E33/E36/E37/E38 and architecture/composition E39–E41. Proposed estimate 7–10 reviewed increments, low confidence; re-estimate after packed core adoption and first consumer wiring. No date promise.
The issue's four acceptance items are E42's proposed success criteria. Independent design review passed and issue #82 was filed before adding E42 and its estimate to docs/epics.json. E39–E41 remain unchanged; generated implementation, epic assertions, current summaries and AL roadmap rows now reflect 42/15/57–89. Registration closes no epic and establishes no runtime acceptance.
Later standing-commission wording and any live owner choices belong in a separately scoped policy/implementation change; this recovery changes no operating policy. Existing historical receipts retain original bytes and status; new accepted behaviour gets new evidence.
Not established: runtime level support, any level or ceiling signed, live journal/consumer/schedule change, private-channel access, new model qualification, reserve enforcement, current Product exposure/default evidence, independent adoption, CI acceptance, release or customer value. Unknown remains unknown. Owner choices and independently verified implementation remain gates; design review and issue registration are recorded above.
12. Eight finding dispositions
| Original finding | Corrected source contract | Future proof |
|---|---|---|
| Astra r2 1: pre-lock timestamp | §5 transaction-time sample, rollback refusal and equal-time ordering | AL1b Monday lock wait/rollback/tie/replay |
| Astra r2 2: owner cap not exposed separately | §6 neutral hard-cap union; §7.2 independent weekly minimum/provenance | AL1a/AL7 absent, present, exhausted, raised, cleared |
| Astra r2 3: no shared transaction owner | §7.4 scoped authoritative ledger and reservation/crash protocol | AL7 concurrent cross-Product gates and every crash boundary |
| Astra r2 4: missing evidence/purpose request answer | §4.2 resolved evidence and class/purpose union; §6 neutral API | AL1a/AL7/AL8 consumer-only feature and changelog fixtures |
| Astra r2 5: Source class cannot select | §8 versioned exact selection and rhythm semantics; claim/result/replay/novelty | AL4 mixed scope, full inbound, changed selection, /1 history |
| Astra r2 6: routine major impossible | §4.2 explicit tech-debt/improvement route | Inert subclass fixtures now; AL1a all advertised subclasses |
| Astra r2 7: monthly cadence blocks obligation | §4.1/§7.3 periodic sweep versus event reaction, urgent single-item update | AL2a/AL7 two distinct monthly obligations, duplicate source |
| Fable r2 material: incomplete week history | §7.1 complete window and implied unset, future refusal, current-head trust | AL1a truncated suffix/first Wednesday/future snapshot |
This table records the correction disposition. Fresh Astra extra-high review passed the recovered revision 3 design before registration; the original failures remain preserved in the provenance manifest. That design verdict establishes no runtime implementation acceptance.
