Factory docs, home
Page navigation

Pre-release documentation. Proposed designs are labelled in the page; acceptance evidence is maintained in the build review. Local paths appear as placeholders. Current state: Build review.

Proposed M01–M04 r2, design only. No runtime, scheduler, model or release commission. A Product needs durable owner intent for its next major outcome, with achievement guidance, independently admitted plans, evidenced progress and honest approximate timelines. Maintenance, security, recovery and improvement continue alongside that trajectory under explicit current caller policy. All 38 existing epics, D4/R7, Knowledge, independent adoption, Fable/Opus and release gates remain unchanged. This optional extension has its own estimate/tracker; it does not silently become an E35 prerequisite or rewrite sealed estimates.

Authority: 1 October owner request U-M01; project intent r11, entity r1 and the task capsule. Tally reaching App Store deployment/review is an example of intent, never a live Tally instruction. Product values and security/release criteria remain in their existing owners; milestone achievement guidance cannot change them.

Actual gap and reuse

Source basis: a41686c. Portfolio stores explicit focus objective/slice text, not milestones, revisioned plans or completion criteria. ProductGuidance stores Product values; it is not a milestone store. The actual evaluateTemporalPlan/parseTemporalPlan and forecastCapacity/parseCapacityForecast functions produce passive dated evidence and conditional capacity scenarios. Capacity's result counts possible additional increments; it is not an ETA. Material Timing concerns communication/readiness windows, not generic work scheduling.

LifecycleSchedule owns discovery occurrences and their non-renewable call allowances. It starts no provider work and is not a general milestone dispatcher. W30/W31/W32 remain separate incomplete Workflows. Reviewed W31 design publication 7ac019c additionally makes external Candidate admission and original-dispatch orphan-proof discovery explicit missing contracts; milestone planning must not rely on the older packet's apparent readiness. D4/R7 42d2700 retains delivery ownership; Knowledge 52096ff retains its own prerequisites.

Reuse CommandJournal for intent transactions and the existing pure planning outputs. M01 has no Intelligence/model decision; M02 uses only the separately qualified Research Workflow and its existing Intelligence policy. Add one bounded Milestone objectives responsibility; do not build a second work ledger, authority store, budget ledger, dispatcher or generic inbox. Proposed public names are software-factory/milestone-objectives and a neutral contract; neither exists now.

R2 closes the independently reviewed r1 design gaps: explicit current-pin drift decisions, a typed passive construction/physical binding seam, and lossless native Temporal attribution on conflicting or unknown dates. R1 and its REVISE evidence remain retained; this amendment is not source acceptance.

M01: useful first increment

One caller-commissioned Product aggregate stores multiple stable milestone IDs and all admitted historical revisions, plus an explicit nullable primary selection. Its scope is (namespace, physical store identity, Product). The caller persistently binds that original scope; reopening, cancellation, namespace changes or a new milestone may not restore consumed external allowances. The exact proposed identity preimages and passive construction seam are specified below; these are not existing APIs.

Each immutable revision contains a title, objective, optional achievement guidance, zero to eight observable criteria, desired target if stated, owner/source provenance and exact Product-values/operating-policy/original-allowance references. An owner may start with just title/objective and optional achievement guidance. No criteria means completion is unavailable/unknown; an empty set can never satisfy completion vacuously. Research or a caller may later propose criteria, but a current-authorised new intent revision must admit them before a plan is admitted or completion recorded. Criteria describe the outcome and name independently commissioned evidence rules. The ordinary form never exposes physical-store/fence/allowance fields: its qualified backend supplies those fixed commission/current provenance pins, or returns explicit unavailable without inventing defaults. Text is bounded data, never a tool instruction. The caller owns a private intent store; owner text is not copied into public receipts, diagnostics or model prompts. Provenance is an opaque bounded identity/digest, not credentials, private endpoints or raw research payloads. Required Product identity comes from a current trusted registry/fence, not the request's claim. The original allowance-scope pin is fixed by the persistent commission and cannot change through revision, plan admission or a new milestone ID; changed current policy/value pins never replenish that allowance.

Closed commands: submit, revise, select-primary, admit-plan, cancel, record-completion. Read is passive. Submission does not select or activate a milestone. select-primary explicitly chooses a current non-cancelled revision or null. Revising/cancelling a selected milestone retains the historical selection pin but makes it ineligible until an explicit selection confirms a current revision or clears it. No due date, plan arrival or missing selection auto-activates another milestone. Multiple milestones/history therefore remain visible without silently competing for primary authority.

A caller can submit a bounded plan proposal without W32 once the referenced intent revision has at least one admitted criterion; each admitted generation must bind at least one of those criteria. admit-plan rechecks current authority and an independently controlled review bound to the exact proposal digest, milestone revision, criteria, Product values and policy. A proposal's own claimed approval is ignored/refused. Admitted subobjective generations are immutable: stable IDs, exact parent revision, dependencies, acceptance criteria, provenance and work class. This is plan admission, not a ticket, Slice, job, dispatch or release receipt. No implementation task is created by M01. New plans for the same milestone revision may reference unchanged generations; replacing a generation appends its predecessor and never rewrites the old one. A new milestone revision requires newly reviewed generation bindings; old accepted/unknown work remains on its original generation, not silently relabelled or counted against changed criteria. First-profile dependency edges remain inside one plan and exact milestone revision; cross-milestone scheduling is deferred.

M01's useful read model shows current/old intent, selected next milestone, admitted plan, dependency order, desired dates, sourced expected ranges or explicit unknowns, completion evidence, and unavailable Research/Implementation. Its forecast projection validates actual Temporal/Capacity output with the existing parsers and re-evaluates freshness using their inputs/policies and caller time. Subject bindings must name the exact milestone/subobjective generations. It selects supported expected date evidence, preserving civil/instant domains, source disagreement, assumptions and expiry; it does not manufacture an overall date from unrelated endpoints. Missing overall evidence remains unknown even if some subobjectives have dates. Any displayed milestone/subobjective date is attributed to its exact caller/source scenario; M01 derives no completion date. One scenario must retain ONE account/window coverage, outstanding-exposure and reserve basis across all its subjects, so the same headroom cannot be reused for every node or for parallel upkeep. Incompatible or incomplete bases, heterogeneous work profiles, dependency/queue/review assumptions or external waits make the combined forecast unknown; independently sourced dates may still be shown as attributed evidence with that limit. Capacity models successive full-cycle releases and counts within a horizon, not a parallel dependency scheduler. It remains a separate native-unit/count scenario. Summing step envelopes or taking their critical path cannot establish a milestone date; a future aggregation method needs separate review against these shared constraints. No capacity-percent-to-token conversion, invented throughput, percentile, critical-path duration or third-party review average is introduced.

Passive construction, physical binding and identities

The proposed createMilestones({journal, commission, fence}) accepts the actual public nominal CommandJournal, a closed immutable sf-milestone-commission/1 / m01/1 commission, and a qualified synchronous fence. The commission contains the original commission ID, scope, host-binding ID and allowance-scope pin. Construction descriptor-validates bounded values and captures the exact Journal and fence binding; it does not open storage, call a port, inspect/create an aggregate, register a new scope or write anything. The sketch has a neutral generic binding plus a concrete facade declaration importing software-factory/command-journal; only a separately commissioned implementation could implement it.

The trusted caller has already persistently bound (physical Journal store, namespace, Product) to ONE original commission/allowance in its existing host registry. That same host owns the common non-expiring physical lock used by all Product/values/policy/authority writers and the actual Journal operations. Each fence carries the exact supplied Journal object and original commission; each returned frame must retain that captured binding by object identity, the exact frozen request and a module-created invocation token. Changed handles/bindings refuse before invocation. Reopening with another actual Journal handle requires a new service/fence pair qualified against the same original physical store/registry/commission; it never permits replacing an existing instance's handle or minting a new allowance. The module cannot authenticate a dishonest host or infer physical identity from a path/string. Both real host arrangements must prove that persistence/lock contract; no new authority store is introduced.

An instance invokes its fence only once per operation. The request contains the commission, authenticated actor, exact normalised execute/list/plans/inspect operation and canonical request digest, including the particular revision/plan/assessment request and supplied planning evidence. The host cannot substitute a different request/frame. Callback use is synchronous, at most once and valid only during that exact run; any late/repeated callback refuses before Journal access. Check original Journal method descriptors before each read/use, after all caller work and immediately before execute; never invoke accessors or overridden methods. Native Promise observation grants no authority. Read permission is checked before disclosure; a current mutation grant is separately required before a new decision. These reference checks support the qualified host contract; they do not claim that an arbitrary object proves a physical lock.

Let J(v) be the existing canonical JSON UTF-8 representation, H(v) the 64 lowercase hexadecimal SHA-256 characters of J(v), and D(v) be sha256: followed by H(v). Freeze these preimages:

IdentityExact preimage or representation
Commission digestD(commission) over all six closed fields: protocol, profile, commissionId, scope, hostBindingId, originalAllowanceScope
Aggregate IDmilestones: + H({protocol:'sf-milestone-aggregate-key/1', scope}); 75 characters, excludes commission ID so a changed commission cannot find a fresh aggregate
Command IDsfmilestone: + H({protocol:'sf-milestone-command-key/1', scope, requestId}); 76 characters, distinct generated type, never the 63-character caller ID grammar
Journal payload{protocol:'sf-milestone-journal-command/1', commissionDigest, actorId, request}; exact request includes expectedVersion/action, so changed actor/payload/version with an original ID conflicts
Fence request digestD({protocol:'sf-milestone-frame/1', commissionDigest, actorId, operation}); operation is the complete normalised operation; no callback/time value
Intent/generation/plan digestsD({protocol:<kind>, commissionDigest, value}), using kinds sf-milestone-intent/1, sf-milestone-generation/1, sf-milestone-plan/1; value is respectively the complete IntentRevision/SubobjectiveGeneration/PlanProposal with only its own pin.digest omitted; all nested pins/predecessors remain exact
Receipt result digestD({protocol:'sf-milestone-result/1', value}), where value is the exact actual Journal result; replayed is the Journal envelope flag, not part of this value

Dynamic admission observations enter the first stored state/result, never a new replay identity. Receipt-result value never includes its own resultDigest; the facade adds that digest to the receipt envelope. Source-owner/native Temporal digests retain their original bytes and prefixes; they are not rehashed or given fabricated identities. Generated IDs are stored whole in selection/receipt/history. Exact-ID vectors and the one-event-per-accepted-transition grammar remain source acceptance gates.

Current pins and command decisions

Compare the entire selected revision's Product-values and operating-policy pins with the frame's current pins. Preserve the original selection even when invalidated. After ordinary absent/revised/cancelled checks, either difference yields pins-stale with product-values-changed, operating-policy-changed, or both; no auto-refresh/re-selection. Missing current facts gives authority-unavailable; missing READ refuses without exposing history. Allowance drift is a host/commission mismatch and refuses, never a stale intent that can be revised into new allowance.

Operation under current READAdditional new-write grant and pin ruleDrift/historical outcome
SubmitCurrent mutation grant; submitted values/policy equal current; allowance equals original commissionMismatched pins refuse STALE before write
ReviseCurrent mutation grant; exact current non-cancelled predecessor/version; NEW intent pins equal current, original allowance unchangedOld predecessor pins may be stale; append deliberate new revision, keep old selection/history and require explicit new selection/review
Select non-null / admit-planCurrent mutation grant; exact current non-cancelled revision and both current pins; plan review covers exact proposal/current revision/pinsRe-selecting an old stale intent does not repair it; refuse STALE. Old review never admits a new revision
Clear primary / cancelCurrent mutation grant; exact current selection/version or current milestone headObsolete intent pins do not prevent safe controls; current policy may still deny the mutation. No work termination or allowance release inferred
Record historical completionCurrent mutation grant plus current authorised assessment of the exact original revision, at least one original criterion, original rules/values/policy and attributable observationsZero criteria refuses UNAVAILABLE with no write. Original pin drift/cancellation does not erase or prohibit a genuinely attributable historical record; it cannot establish current completion/trajectory eligibility
Original receipt replayCurrent READ, exact actor/request identity, reconstructed original history and actual replayed:true; replay path's decision must unconditionally refuse any new writeWrite revocation or values/policy drift does not hide a permitted historical receipt. Missing receipt cannot fall through into new admission
List/plans/inspectCurrent READ, passive history validationReturn immutable original records. Current completion/combined forecast is unknown on pin drift; single-source date evidence remains explicitly historical/attributed. No historical record changes

A new current-pin revision and independently reviewed plan can restore eligibility only through explicit selection/admission. Upkeep remains separately admitted by its own current policy and original allowances throughout; no row above disables or reprioritises it. Recorded completion under old pins remains historical; its current validity is unknown with typed pin-drift reasons until a separately authorised current assessment can speak about the current revision (no retroactive criteria replacement).

Authority, replay and retained work

Every new mutation is admitted under one qualified physical authority/Journal fence shared by all applicable authority, identity and operating-policy writers. Authentication stays outside request data. The implementation must use the actual public CommandJournal and descriptor-validated original methods, following the corrected receiver pattern; no opaque alternate write target. Caller ports are synchronous and bounded. Qualification is a real host proof, not a boolean field or expiring lease.

Parse before ports; inside the outer fence capture current registry, policy, Product-values/review/criterion facts and one admission time, then invoke a pure Journal decision at the exact expected aggregate version. Do not perform port calls or nested Journal work inside decide. The admission sample is not a physical commit deadline. Current permissions are admission facts; later asynchronous work requires its own current admission. No promise result is awaited as permission. Bounded own string/symbol descriptor observation must observe native rejection and fulfilment data under one unchanged budget, without getters/thenables or changing the original refusal/possible-commit classification.

An intent mutation, event and original receipt share one actual Journal transaction. Same original ID/payload replays; changed reuse conflicts. Dynamic admission time is excluded from replay identity. Current read authority is required before any receipt/history disclosure. A recorded write may replay under read authority when write permission was revoked, but only if the actual Journal receipt confirms replay. Missing/corrupt receipts refuse; an already reconciled authorised historical witness survives an outer acknowledgement fault. A possible commit returns unknown and can only reconcile the original ID. No fresh request, milestone, revision or scope may be used to evade uncertain prior work.

External research, delivery, evidence and budget owners are separate transactions. M01 never claims their atomicity. Subsequent adapters must persist intent/binding before contact, reuse the owning Workflow's original ID/recovery rules, and treat unacknowledged admission as unknown. Accepted or uncertain work remains linked to its original milestone/subobjective generation and allowance after revision, replanning, cancellation, primary changes or restart. Removing a node from the current plan creates an explicit historical/orphaned-work view; it never cancels or forgets that work. Cancellation stops future milestone-driven admission only; cancellation of a worker is a separate current-authorised owner action and is not proof of termination.

Deterministic native date projection

First parseTemporalPlan verifies the supplied historical plan; then evaluateTemporalPlan re-evaluates its exact inputs/policy at the qualified read time. Both exact plan digests/as-of values, the bound subject ID and every expected claim reference are retained in TemporalAttribution, including unknown/conflicting output. Each native claim reference resolves the original primary and derived source refs through that exact native plan; no locator/raw quotation is copied into the projection. Direct attribution uses native {id, revision, contentDigest} and {claimId, revision} shapes, never inventing storeId/recordId to convert a Temporal ref into a milestone SourcePin. A caller must retain the supplied plan if it wants to dereference those exact references later; M01 does not claim to archive external source bodies.

Real evaluated Temporal caseM01 projection
Invalid digest/shape, scope or subject bindingFixed refusal, no selected date or silent provenance repair
Missing plan/groupUnknown; null attribution only when no plan was supplied; otherwise exact native-plan/subject attribution with an empty expected-claim list
Native source-disagreement between usable expected claimsconflicting-evidence/source-disagreement, no date/winner/union window; retain all expected claim refs, including both Q4 and January in the retained r1 counterexample
Incomparable civil/instant domains or civil zonesconflicting-evidence/incomparable-domain, no timezone conversion or preferred source; retain references
Multiple active expected claims, even overlapping/identical supports without a native disagreementUnknown multiple-expected-claims; no unreviewed combining or tie-break rule
Incomplete coverage, active unusable/expired/stale/unvalidated/unknown-date claim, unresolved relevant hard constraint, current pin drift or incompatible shared scenarioUnknown with fixed applicable reasons and exact attribution; no arbitrary usable subset selected
Withdrawn/superseded claimsRetain their native refs; do not select them. If none remains usable, unknown inactive/missing evidence. An unusable active alternative still blocks selection
Exactly one active usable expected claim, complete coverage, no applicable conflict/unknown and compatible declared scenarioOne sourced-scenario with its exact native support, claim/source/basis refs, read time and exclusive freshness cutoff; combined-resource assessment still not derived

The cutoff is the earliest non-null claim expiry, every primary/derived source validUntil, and each source checkedAt + policy.maxSourceAgeMs; asOf >= cutoff is unknown, exactly matching Temporal's exclusive checks. Do not extend it to a desired target or later milestone revision. Constraint relevance starts at the bound subject, recursively adds predecessors of native after-constraints whose target is already in that set, and includes all hard constraints targeting that set; any hard violated/unknown result blocks a selected forecast and retains the native plan pin. Announced/observed dates remain reachable in that pinned plan but never substitute for an expected date. Capacity's original digest and native increment-count result are copied unchanged; it never provides a date.

Completion, timelines and parallel upkeep

Keep four fields separate: desired target, forecast, explicit commitment, and observed completion. M01 creates no commitment; its field says not commissioned. Completion is recorded only for the exact revision after current qualified assessment covers every original criterion with pinned attributable observations. A model judgement, completed ticket count or copied provider response cannot satisfy that rule. Historical completion remains historical when evidence expires, policy changes or a newer revision is created; current validity can become unknown or contradicted without deleting that history.

Illustrative local release facts are distinct: submitted, review approved, and publicly available. Submission does not prove approval; approval does not prove the selected build is live. The owner chooses which observed outcome the milestone requires. Apple/other third-party waiting time is unknown unless a specifically sourced, comparable, expiring scenario exists. No live connector or universal review duration is proposed. A later local fixture can simulate those separately attributed transitions.

Primary means the owner's next trajectory, not exclusive worker access. A trusted current operating policy names eligible upkeep lanes, WIP/priorities/pre-emption and reservation/allowance owners. There are no default percentages or invented reserves. Existing security/release policy cannot be overridden by milestone text. Any future combined forecast must account for already admitted work and unresolved effects at their original commissioned exposure, and for explicit shared reserves once across Products. M01 cannot prove that aggregate scheduling calculation and labels its sourced dates accordingly. Missing policy/reservation coverage makes admission unavailable and forecasts unknown; changing the milestone cannot stop independently commissioned maintenance/security/recovery/improvement or appropriate their allowance. Scheduling fairness and conflict admission belong to a later qualified existing owner, not this intent aggregate.

Bounds and refusal semantics

Proposed first profile: 8 lifetime milestones per original Product scope; 16 intent revisions total (8 per milestone); 16 admitted plans; 32 stable subobjectives/128 lifetime generations (4 per ID); 0–8 criteria per revision (at least one required for admitted plans or completion); 8 prerequisites per generation; depth 16; 32 source pins per plan. Same-scope references only; dangling or duplicate IDs, inconsistent predecessor identity, cycles, forward-generation claims and cross-Product dependencies refuse. Global counters include retired/cancelled/superseded entries; no eviction, reuse or reset. Cancellation is terminal for that milestone ID; revise/admit-plan refuse thereafter, and no new ID adopts its accepted or uncertain work. Historical inspection and recording attributable completion for a pre-cancellation revision remain possible under current authority; cancellation is not evidence that an effect did not occur. Expansion requires a separately reviewed profile/migration preserving the original lifetime counts and allowance identity, not a new namespace workaround.

Canonical UTF-8 bounds (serialised JSON bytes, including escapes; never raw character counts): title 160 bytes, objective 2,000, guidance 4,000, each criterion requirement 512, cancellation reason 512, whole generation 2,048; request/qualified-frame/command/event/result 64 KiB; state 1 MiB; depth 32. Growth stops at 768 KiB/96 accepted growing transitions, with at most 128 lifetime transitions. The remaining 256 KiB/32 slots reserve closure: at most eight cancellations, 16 first completion records and one final clearing of a non-null primary after growth admission stops (25 required transitions, seven spare). Ordinary selection/clearing before that stop consumes the growth allowance; replay consumes neither. No command may consume a reserve slot without reducing its corresponding retained obligation. Prove exact worst-case serialised state/receipt/event reserves before accepting any growth; counters and bytes both govern admission. Each accepted transition writes exactly one typed event (version/index/command identity checked); replay writes none. Reads capture one version and replay at most eight pages of 16 events, followed by one bounded one-row tail/sentinel check that refuses extra history. Sparse global sequence numbers are not treated as missing per-aggregate events; every captured aggregate version must match one complete event. Forecast evaluations are passive, not an unbounded observation log. Bounds intersect; refuse before any write, never truncate. A compact list/read API pages original IDs with version binding rather than returning a full catalogue. A separate bounded plan-summary listing keeps superseded plans discoverable; inspect may select one exact historical plan or the latest for that revision. Inspect returns one exact revision, at most one retained completion record for that revision, an admitted-plan summary and at most eight subobjective definitions/forecasts. A continuation pins the aggregate version, exact revision and plan; any drift refuses stale rather than mixing pages. A completion has at most 16 witnesses (two per criterion), is recorded at most once per revision, and occupies at most 8 KiB; current validity refers to that record without duplicating it. No recorded completion, or zero admitted criteria, means current completion is unknown, not silently recorded by read. Each forecast is at most 4 KiB, including up to 16 native expected-claim refs; this is a projection-only bound change from r1 to preserve attribution. Pages adapt downward from the requested maximum of eight whole subobjective items to fit the complete 64 KiB success envelope. Never truncate a claim ref/item, silently omit a row, or change admission limits to hide an unreadable historical record. A source proof must establish that maximal admitted fixed fields plus at least one full item always fit; otherwise this proposed result layout requires review before implementation acceptance. Empty scopes and unset initial selection return explicit absence, with no fabricated selection receipt. Every accepted revision/plan must remain readable through these pages; prove the composite 64 KiB result cap at maximal admitted fields. M01 work coverage explicitly says not bound; live/orphan work projection belongs to M03.

Fixed failures distinguish invalid/limit, stale/conflict, forbidden/unavailable, corrupt retained state and unknown commit. Absence is not zero, cancellation is not completed, unavailable is not failed, and malformed caller data is not proven stored corruption. Constructor/import/read do not initialise, dispatch, repair, fetch or migrate anything.

Four bounded packets and estimate

PacketUseful resultPrerequisites and boundaryRough reviewed increments
M01 intent and passive plan projectionDurable submission/revisions, explicit nullable primary, manually/caller-reviewed immutable plan, criteria/history and honest forecast/unknown viewAccepted contract; actual Journal/fence/registry plus existing passive planners; W32/W30 explicitly unavailable2–3
M02 research proposal adapterOne bounded research request and attributed decomposition proposed for current admissionSeparately accepted W32 or another explicitly reviewed public research job; preserve its original job/Budget and unknown recovery; no new research engine1–2
M03 admitted work and schedule projectionExact plan-generation → existing ticket/Slice/job references, orphan work, dependency/overall forecast scenarios and current policy/reserve inputsSeparately accepted owning Workflow/D4/R7 interfaces; actual admission/reconciliation and scheduling contract. No fictionally available generic dispatcher2–3
M04 honest Factory UI and compositionProject next-milestone form, guidance/criteria, explicit primary, progress/dependency/forecast display and parallel upkeep statusBasic owner form/intent/manual-plan view follows M01 directly; Research/ticket actions remain unavailable. Later sections consume M02/M03 only after acceptance. Actual caller authority, installed package and independent responsive UI proof1–2

Estimate 6–10 reviewed increments, low confidence, for these adapters/UI after their separately tracked Workflow/owner prerequisites. This is not a calendar promise and excludes completing W30/W31/W32 or qualifying a live provider/release system. M01 is independently useful and does not wait for them. The initial M04 form/view can follow M01 immediately; its useful owner capability does not wait for Research or ticket dispatch. M02/M03 continue as separate later composition gates. If a first packet cannot retain its complete external consumer/fault proof, split its implementation internally without claiming the remaining gates passed.

Proof and implementation gate

Before source commission, independently review this exact proposal, neutral sketch and actual-source compatibility probe. M01 implementation then needs a real private-tarball consumer with two distinct qualified caller registry/fence arrangements over actual Journals; no Factory checkout, UI, default store or live provider. Proposed source closure is four small responsibilities: milestone-objectives-contract.ts (neutral values, no runtime imports); canonical state/replay validation using existing strict-input/canonical JSON helpers; passive projection using actual Temporal/Capacity evaluators; and a facade using the actual Journal and intrinsic native-Promise observation. There are no Portfolio, Guidance, Workflow, worker, provider or LifecycleSchedule runtime dependencies or edits. Compiler-derived closure and no-ambient neutral declarations must be proved, not assumed.

Required proof matrix: two-process duplicate/version races; current revoke before/inside the physical fence; null/changed/cancelled primary and multiple histories; revised criteria unable to retroactively complete old work; exact ID replay after crash/lost/late acknowledgement; orphan work and unresolved effects retained; source/plan/allowance mismatch; dependency cycles/depth/bytes/one-over limits and reserved closure; corrupt/missing/replaced receipt/history with no repair; constructor/read/passive imports; hostile accessors/thenables/native fulfilled/rejected nested/symbol data at every port and post-commit boundary; cross-Product/store identity; source expiry and historical versus current completion; all third-party stages separate; no invented forecasts under missing data/policy; upkeep still eligible with primary present/cancelled; actual installed parser/type/private-import tests. Later adapters add actual research/ticket owner substitutions and process crash/reopen proofs. Later UI reuses ValuesPage edit/save/stale/lost-ack history patterns and the dashboard command-token/origin/disabled-edit guards, while keeping milestone overview separate from Product values and Portfolio focus. It displays literal untrusted text and unavailable/progress states at narrow and wide sizes.

No source or public entry is commissioned by this document. The task capsule, four proposed packets and existing owners retain the next decision; historical receipts and existing module policies stay unchanged.

Source: docs/milestone-objectives-proposal.md