M0-01R2 — Separate knowledge lineage from effective temporal projection #9
Labels
No labels
bug
duplicate
enhancement
help wanted
invalid
question
wontfix
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
Creatidy/model-intelligence#9
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Owner Authorization and Predecessors
Adrian explicitly authorized ONE successor issue and ONE implementation PR to develop; this issue is owner-selected upon creation. Canonical develop freshly fetched and verified at
6edd31d491a41a92e73ebfc6d9ee81c284129f8a. Use one normal checkout and ordinary issue branch; no git worktree. No merge/release/deploy/main changes.Refs #3 / unmerged PR #4 (
303f5fe118a592f3cf5e7ea47c5955625850ea28), the heuristic-snapshot experiment ending STOP_REVISE: payload, timestamp and row presence cannot reconstruct revision or cessation. Refs #7 / unmerged PR #8 (c70bbd380a0728bad04112f1c02f076fb94ee1e4), the explicit-lineage experiment ending a second STOP_REVISE: explicit IDs and baseline/override are useful, but global knowledge heads must not also govern effective applicability; one overloaded changes() mixed evidence and projected-state deltas. Use behavioral/adversarial evidence, not cherry-picked projection code. Do not patch either predecessor.Hypothesis and Accepted Scope
Deterministic typed Python proof, preferably zero runtime dependencies:
historical evidence -> knowledge/lineage viewindependently ofhistorical evidence -> effective-at-T -> projection delta; new explicit claims separately produce evidence delta. Success is conceptual simplicity, not minimum LOC.Observation: explicit observation/source IDs, authority/type/reference, observed/published/retrieved times, independent freshness and distribution/license. Retrieval is acquisition metadata; refresh cannot change semantic identity or automatically renew freshness.
Claims: explicit claim IDs referencing observations, separate typed model/capability/benchmark/price, plan/offer and execution-surface families. Equal payloads and cross-source matches do not establish identity or lineage. Explicit revises links describe knowledge/version lineage. Reject impossible lineage; preserve branching and unresolved independent evidence.
Knowledge view: retained history, supplied relationships, latest-known head(s), future known versions and stale observations, separate from current projection. Future revisions may become latest-known immediately without suppressing currently applicable predecessors.
Durable baselines: explicit effective_from versions, not arbitrary temporary finite intervals. A -> B -> C durable terms persist until applicable explicit revisions replace them; future baseline B cannot create a current-state gap. Temporary deviation belongs to bounded typed override fields.
Overrides: independent campaigns with declared fields, bounded active terms, explicit effective revision boundary, public conditions/zoned windows and effective withdrawals. Future revisions/announced withdrawal do not prematurely deactivate predecessors. Expired/revised versions stay historical but not active/current.
Effective projection: applicable baseline versions plus applicable override versions, omitted fields inherit, disjoint fields compose, incompatible applicable branch/source evidence surfaces conflicts with IDs. No retrieval/source/lexical/authority/timestamp winner or implicit multiplier. Unknown != false. Boundary disputes visible when current/relevant, not irrelevant historical noise.
Evidence delta: new IDs and explicit links announce claims/revisions/withdrawals directly, including future versions; retrieval-only silent. No projection-based causal reconstruction.
Projection delta: compare deterministic before/after effective views and participating IDs, not invisible knowledge events. Future-only knowledge at a fixed T cannot fabricate current price change. Historical/future claims must not be shoved into current views merely for causal attribution.
Required Proofs
Future baseline: Jan 1 price20 A, Nov 1 price25 B revises A, B learned Oct1. Oct15 latest-known B/effective A; evidence delta cites B->A; same-time projection delta silent. Nov1 effective B/projection delta baseline+price change with A/B participants. Also cover learning B after its effective start.
Override lifecycle: baseline only; future known campaign; activation; declared replacement/inheritance; expiry/fallback; future revision announcement/current predecessor; revision effective; shortened/extended/changed terms; advance withdrawal/effective withdrawal; retained historical versions/current exclusion; later independent activation.
A: model/provider/capability/contextual benchmark/price/revisions/independent disagreement, no universal winner; future-dated revisions do not prematurely apply.
B: durable baseline, bounded conditional/zoned campaign projection, no private remaining quota.
C: advertised/observed/selectable/enforceable across headless, structured output, workspace editing, model selection, physical-model observability/enforceability. Provider-managed surface need not know physical model ID; future evidence uses consistent lineage/applicability separation.
Port predecessor invariants: missing rows never cessation; older history cannot close newer evidence; equal cross-source values never lineage; superseded historical evidence not current correction; retrieval-only silent and stale stays stale; no timestamp revision inference; branch conflicts/no winner; historical expiry retained/current exclusion; relevant boundaries preserved; unknown != false; fallback after expiry.
Regressions from all PR8 finish findings: baseline conflict must persist through temporary masking; batched linked price revisions must not lose evidence signals; inactive boundary knowledge is an evidence delta rather than fabricated effective causality; future/same-term baseline replacement must not create unknown gap. Do not preserve predecessor's erroneous immediate-global-supersession expectation.
Hard Stops and Non-Goals
Last custom temporal-model proof before stronger NO decision. STOP_REVISE without further patches if causal inference needs substantial payload/timestamp/row heuristics, generic graph/rules DSL, broad provider subclasses, mixed evidence/projection deltas, arbitrary finite baselines for promotions, or architecture harder to explain than PR8. No _corrected/_ceased equivalents, global heads interchangeable with applicability, generic patch/event engine, database/temporal database/event sourcing/REST/MCP server/scheduler/ingestion/provider API clients/expression language/scoring/trust ranking/UI/downstream router or private runtime/accounts.
Validation and Delivery
Network-free deterministic lineage, knowledge-vs-effective, future/past-effective baseline, override lifecycle, evidence/projection deltas, conflict/branch, A/B/C and predecessor adversarial tests; full make check; package build; isolated installed-wheel smoke; full-range whitespace. Report production modules/physical LOC/public types/private helpers/runtime dependencies and predecessor comparison. Minimal dependencies/reuse boundary; no copied third-party datasets/code.
One branch from freshly fetched develop, Adrian-authored normal commits/pushes, ONE PR to develop with Refs (issue stays open). Concise successor comments on #7/PR8 allowed. Automatically run native /finish-pr: fresh foreground independent pr-reviewer tasks of frozen whole PR, direct JSON result; at most THREE remediation rounds after initial review, including fixes to fixes, then required fresh review. Reviewer should exercise additional valid-input adversarial probes within inspected permissions. No owner relaying/manual review/controller or extended loop. Final report only at terminal READY_TO_MERGE, STOP_REVISE, OWNER_DECISION_NEEDED or precise finite BLOCKED; never merge.
Implementation delivered as PR #10: #10. Base
6edd31d491; clean pushed HEADdb74252119. Separated knowledge/effective projection and evidence/projection delta APIs; 570 production LOC, 25 public dataclasses, 1 private helper, zero runtime dependencies. Focused proof tests PASS (33); full offline make check PASS (35), package build and isolated installed-wheel smoke PASS. Entering native bounded /finish-pr now; these implementation checks are not independent approval. Issue remains open; never merge.READY_TO_MERGE: PR #10 #10 received fresh native whole-PR APPROVE with empty findings at exact HEAD
177471b7bc/ base6edd31d491. One remediation round fixed both initially reproduced campaign-alternative/composition defects. Final full offline gate PASS37, build/isolated installed-wheel smoke PASS; final complexity587 LOC/4 modules/25 dataclasses/1 private helper/zero runtime deps. Knowledge lineage, effective-at-T, evidence delta and projection delta remain separate; future baseline proof has no unknown gap. Final clean/canonical/MCP currentness verified. PR remains unmerged and issue open; no main/release/deploy/product GO.kilo referenced this issue from Creatidy/scarcity-router2026-10-04 22:16:05 +00:00
kilo referenced this issue2026-10-04 22:37:23 +00:00
Verified completion of already-delivered #9 under the current owner-authorized loop's stale-open completion rule. No reimplementation, new branch/PR, remerge or history rewrite.
Actual PR #10 was merged by Adrian on 2026-10-04T22:14:43Z at
fb7299810fdc4612d4e0559465faef571662c6ef. Exact parents are original reviewed base6edd31d491a41a92e73ebfc6d9ee81c284129f8aand approved HEAD177471b7bc083e796d03096382740d1b2d80d676; merge tree equals approved tree6b8774423e4d468435e5c83edf3154b9a5b56f82. Both merge/approved objects are ancestors of current canonical develop98b270e387571decceb9be7d177d122656de9fa3.Canonical historical review13662/13664 remains: initial REQUEST_CHANGES, one reproduced/fixed remediation round, final fresh independent APPROVE/findings empty at exact177471/base6edd. No new review was dispatched or substituted.
evidence.py,projection.py,deltas.py,__init__.py,test_separated_proof.py, pyproject and lock have an EMPTY approved-to-current diff; the bounded proof's recorded four modules/587 LOC/25 dataclasses/one helper/zero runtime dependencies are preserved, not a claim about total current package size.Current verification: sanitized synthetic offline full
make checkPASS115/all Ruff/format/basedpyright/locked/whitespace gates; focused unchanged R2 proof PASS35; offline sdist+wheel build PASS; no-cache isolated installed-wheel smoke PASS from site-packages, covering knowledge/current future baseline separation, evidence/projection deltas, py.typed and no Requires-Dist. Original full35/37 counts are historical, not confused with current totals. Full original/base-to-current whitespace, fresh canonical ref/clean status and accepted source/test behavior verified.The complete supplied-history knowledge versus effective projection, baseline/override lifecycle, branch/independent conflicts, unknown/false, explicit evidence delta versus endpoint projection delta, A/B/C and predecessor regressions remain intact. STOP_REVISE predecessors #3/#7 and PR4/8 remain untouched. Receipt is the deterministic synthetic architecture proof only: no real-source value/full A/B/C, source acquisition, upstream copying permission, installed consumer acceptance, private routing/state, inference, release/deployment/main or product GO. Closing this stale-open issue records existing accepted integration, not a new feature/authority grant.