M0-01R2: Separate knowledge lineage from effective temporal projection #10

Merged

adrian.tkacz merged 2 commits from issue-9-knowledge-effective into develop

2026-10-04 22:14:43 +00:00

kilo commented

2026-10-04 19:20:05 +00:00

Member

Refs #9
Predecessor evidence: #3 / PR #4 STOP_REVISE at 303f5fe118a592f3cf5e7ea47c5955625850ea28; #7 / PR #8 STOP_REVISE at c70bbd380a0728bad04112f1c02f076fb94ee1e4. Neither predecessor was patched or cherry-picked; no projection implementation carried forward. One normal checkout /home/adrian/workspace/model-intelligence, ordinary branch from freshly fetched canonical develop.

Frozen Delivery

Base/merge base: 6edd31d491a41a92e73ebfc6d9ee81c284129f8a.
HEAD: db74252119ea7013f223e55abae7ab661bcae3a1.
Normal Git commit/push verified under effective author/committer Adrian Tkacz adrian.tkacz@creatidy.com. Clean status, full delta, whitespace, ancestry and canonical origin verified. PR targets develop, issue stays open, no merge/main/release/deploy.

Separated Proof

Explicit immutable public observations with independent acquisition/freshness; typed model, baseline, override and surface claims with supplied IDs/revision links. No value/source/timestamp inference of lineage; history retention required for evidence transitions.
knowledge() reads all retained history, latest-known heads independently of applicability, future versions and stale observations. Branches remain unresolved explicit heads.
effective_at() independently restricts claims to explicit effective version boundaries, then removes replaced ancestors. Revision boundaries are monotone within an explicit lineage (equal boundaries allowed); no arbitrary timestamp/source/authority winner. Future-known baseline/model/surface/override/withdrawal revisions cannot suppress current predecessors.
Durable BaselineClaim.effective_from versions have no finite interval. Overrides separately have an effective version boundary plus a bounded active period, public conditions and local zoned windows; an explicit withdrawal has no terms/period. Once a revision boundary arrives, predecessors do not resurrect after expiry or failed conditions.
Applicable baseline fields plus declared active override fields; omitted fields inherit. Absolute quota replacement, no implicit composition. Independent field conflicts, baseline conflicts beneath masking, branch diagnostics and relevant campaign boundary conflicts cite IDs; inactive disputed participants appear only in relevant diagnostics, never active overrides. Irrelevant historical campaigns excluded.
evidence_delta() exposes new observation/claim identities directly, including all revision/withdrawal announcements in a batch and future known evidence. No overloaded changes() or event engine. Retrieval-only silent, semantic identity/freshness rewrites rejected, missing history never cessation.
projection_delta() compares typed effective endpoint views, carrying before/after participants/terms/conflicts. It does not reconstruct invisible causal history. Equal-price replacement can change baseline participants without inventing a price-change event.

Proof Coverage

Required Jan1 price20 A / Nov1 price25 B learned Oct1: latest B/current A Oct15, explicit evidence revision immediately, fixed-time projection delta silent, effective B and A/B endpoint transition Nov1. Past-effective learned revision also covered. Same-term future/past baseline revisions cannot manufacture unknown gaps.
Override activation/inheritance/expiry/fallback, future revision/withdrawal announcements versus effective boundaries, shortening/extension/term change, delayed active periods, history retention/current exclusion, later independent campaigns, conditions/windows/DST/half-open boundaries.
A/B/C: contextual model/benchmark/price evidence and future revisions; plan generation/baseline/campaign; separate advertised/observed/selectable/enforceable modes and no invented provider-managed physical ID.
Predecessor adversaries: missing rows, old closed history, cross-source equal-value revision suppression, stale/retrieval-only silence, branches including nested/future/equal-value branches, baseline conflict masking, batched price-changing then equal-price/reverting revisions, relevant inactive boundary diagnostics, future same-term baseline gap. 120 input permutations and socket-blocked combined A/B/C proof.

Complexity and Limits

4 production modules: evidence.py 353 LOC, projection.py 153, deltas.py 63, init 1 = 570 physical LOC. 25 public frozen dataclasses, 1 type alias, 1 private top-level helper, 0 runtime dependencies. PR8 final production was 647 LOC with overloaded knowledge/projection/changes; this proof separates new identities from effective endpoint comparison and replaces causal reconstruction with supplied lineage. No generic EAV/graph/DSL/event machinery/provider hierarchy or infrastructure. Synthetic proof, not real market verification/product GO. No copied upstream code/data; upstream catalogs remain outside core.

Executed Validation

uv run --no-sync python -m unittest discover -s tests -p test_separated_proof.py -v: PASS, 33 proof tests with lifecycle/branch/delta/adversarial subcases.
UV_OFFLINE=1 make check: PASS, 35 tests including packaging, locked sync, Ruff lint/format, basedpyright 0 diagnostics and staged/unstaged whitespace.
UV_OFFLINE=1 uv build: PASS, sdist/wheel.
uv run --offline --no-cache --isolated --no-project --with /home/adrian/workspace/model-intelligence/dist/model_intelligence-0.1.0-py3-none-any.whl python /tmp/kilo/mi-r2-wheel-smoke.py: PASS. Imported installed wheel outside src, knowledge/current future baseline split, evidence/projection delta, py.typed and no Requires-Dist verified.
Full staged/intended/base-range whitespace: PASS.

Automatically entering repository-native /finish-pr now: fresh foreground independent whole-PR reviewers, direct frozen JSON handoff, maximum three remediation rounds and a fresh final review. Additional adversarial valid-input scrutiny required within inspected read-only permissions; do not infer approval from these implementation checks. Hard-stop if causal reconstruction heuristics or structurally harder temporal machinery become necessary.

Refs #9 Predecessor evidence: #3 / PR #4 STOP_REVISE at `303f5fe118a592f3cf5e7ea47c5955625850ea28`; #7 / PR #8 STOP_REVISE at `c70bbd380a0728bad04112f1c02f076fb94ee1e4`. Neither predecessor was patched or cherry-picked; no projection implementation carried forward. One normal checkout `/home/adrian/workspace/model-intelligence`, ordinary branch from freshly fetched canonical develop. ## Frozen Delivery Base/merge base: `6edd31d491a41a92e73ebfc6d9ee81c284129f8a`. HEAD: `db74252119ea7013f223e55abae7ab661bcae3a1`. Normal Git commit/push verified under effective author/committer Adrian Tkacz <adrian.tkacz@creatidy.com>. Clean status, full delta, whitespace, ancestry and canonical origin verified. PR targets develop, issue stays open, no merge/main/release/deploy. ## Separated Proof - Explicit immutable public observations with independent acquisition/freshness; typed model, baseline, override and surface claims with supplied IDs/revision links. No value/source/timestamp inference of lineage; history retention required for evidence transitions. - `knowledge()` reads all retained history, latest-known heads independently of applicability, future versions and stale observations. Branches remain unresolved explicit heads. - `effective_at()` independently restricts claims to explicit effective version boundaries, then removes replaced ancestors. Revision boundaries are monotone within an explicit lineage (equal boundaries allowed); no arbitrary timestamp/source/authority winner. Future-known baseline/model/surface/override/withdrawal revisions cannot suppress current predecessors. - Durable `BaselineClaim.effective_from` versions have no finite interval. Overrides separately have an effective version boundary plus a bounded active period, public conditions and local zoned windows; an explicit withdrawal has no terms/period. Once a revision boundary arrives, predecessors do not resurrect after expiry or failed conditions. - Applicable baseline fields plus declared active override fields; omitted fields inherit. Absolute quota replacement, no implicit composition. Independent field conflicts, baseline conflicts beneath masking, branch diagnostics and relevant campaign boundary conflicts cite IDs; inactive disputed participants appear only in relevant diagnostics, never active overrides. Irrelevant historical campaigns excluded. - `evidence_delta()` exposes new observation/claim identities directly, including all revision/withdrawal announcements in a batch and future known evidence. No overloaded changes() or event engine. Retrieval-only silent, semantic identity/freshness rewrites rejected, missing history never cessation. - `projection_delta()` compares typed effective endpoint views, carrying before/after participants/terms/conflicts. It does not reconstruct invisible causal history. Equal-price replacement can change baseline participants without inventing a price-change event. ## Proof Coverage Required Jan1 price20 A / Nov1 price25 B learned Oct1: latest B/current A Oct15, explicit evidence revision immediately, fixed-time projection delta silent, effective B and A/B endpoint transition Nov1. Past-effective learned revision also covered. Same-term future/past baseline revisions cannot manufacture unknown gaps. Override activation/inheritance/expiry/fallback, future revision/withdrawal announcements versus effective boundaries, shortening/extension/term change, delayed active periods, history retention/current exclusion, later independent campaigns, conditions/windows/DST/half-open boundaries. A/B/C: contextual model/benchmark/price evidence and future revisions; plan generation/baseline/campaign; separate advertised/observed/selectable/enforceable modes and no invented provider-managed physical ID. Predecessor adversaries: missing rows, old closed history, cross-source equal-value revision suppression, stale/retrieval-only silence, branches including nested/future/equal-value branches, baseline conflict masking, batched price-changing then equal-price/reverting revisions, relevant inactive boundary diagnostics, future same-term baseline gap. 120 input permutations and socket-blocked combined A/B/C proof. ## Complexity and Limits 4 production modules: evidence.py 353 LOC, projection.py 153, deltas.py 63, init 1 = 570 physical LOC. 25 public frozen dataclasses, 1 type alias, 1 private top-level helper, 0 runtime dependencies. PR8 final production was 647 LOC with overloaded knowledge/projection/changes; this proof separates new identities from effective endpoint comparison and replaces causal reconstruction with supplied lineage. No generic EAV/graph/DSL/event machinery/provider hierarchy or infrastructure. Synthetic proof, not real market verification/product GO. No copied upstream code/data; upstream catalogs remain outside core. ## Executed Validation - `uv run --no-sync python -m unittest discover -s tests -p test_separated_proof.py -v`: PASS, 33 proof tests with lifecycle/branch/delta/adversarial subcases. - `UV_OFFLINE=1 make check`: PASS, 35 tests including packaging, locked sync, Ruff lint/format, basedpyright 0 diagnostics and staged/unstaged whitespace. - `UV_OFFLINE=1 uv build`: PASS, sdist/wheel. - `uv run --offline --no-cache --isolated --no-project --with /home/adrian/workspace/model-intelligence/dist/model_intelligence-0.1.0-py3-none-any.whl python /tmp/kilo/mi-r2-wheel-smoke.py`: PASS. Imported installed wheel outside src, knowledge/current future baseline split, evidence/projection delta, py.typed and no Requires-Dist verified. - Full staged/intended/base-range whitespace: PASS. Automatically entering repository-native /finish-pr now: fresh foreground independent whole-PR reviewers, direct frozen JSON handoff, maximum three remediation rounds and a fresh final review. Additional adversarial valid-input scrutiny required within inspected read-only permissions; do not infer approval from these implementation checks. Hard-stop if causal reconstruction heuristics or structurally harder temporal machinery become necessary.

kilo added 1 commit

2026-10-04 19:20:05 +00:00

Separate knowledge lineage from effective temporal projection db74252119

Rows
Columns