Replace legacy static model governance with evidenced execution resources

kilo commented

2026-10-28 20:05:07 +00:00

Member

Goal

Implement the owner's 2026-10-28 architectural decision to remove repository-versioned static model governance and route authorized executable resources using actual evidence and local authority. Supersede the parts of D-082 and earlier decisions requiring bootstrap static catalog and rating-based routing, while preserving historical decisions and valid security and provenance protections.

Why and Baseline

selection_app.py loads model-catalog.json and model-policy.json for selection, profiles and provenance; model_tracks.py and execution_sources.py use model-tracks.json for family-derived capability floors. Gateway, CLI, REST, MCP, packaging, persisted decisions and tests depend on these contracts. #176 delivered an explanation-only MI consumer with bootstrap preservation; it is closed and is not equivalent to this explicitly authorized removal. Open PR #233 addresses worker CA trust, not this work.

Scope

Remove all three JSON artifacts, packaging and runtime dependencies, obsolete loaders, static workflow model/role assignments, assumed task profiles and model-family rating floors. Do not replace them with another static catalog, profile set or track registry.
Kernel/harness owns task intent, workflow, reviewer selection and acceptance; MI owns external evidence and history; Router owns execution sources, inventory, local capacity, compatibility, authorization, selection and admission; Console presents authorized state and commands.
Reuse existing ModelInventory, SourceRegistry, ExecutionSource, bindings, adapter evidence, capacity, authorization, MIConsumer and selection/routing cores as applicable. Construct candidates from configured authorized sources, actual discovery, evidenced adapter capabilities, applicable trusted MI evidence and local policy, without conflating discovery, execution attestation, entitlement or public knowledge.
Honor exact physical model and effort pins, hard requirements, known/unknown distinctions, authorization, evidenced bounds, capacity and scarcity; explain exclusions and limitations. Unknown quality is not a fabricated score or entitlement. An exact authorized model request need not fail solely for lack of comparative quality evidence. Empty execution-source configuration must remain honest.
Preserve trusted MI references, applicability, freshness, conflict/withdrawal handling, immutable decision provenance, independently supported local capabilities and offline usability. MI must not alter private authority or imply quota.
Inventory public CLI, API, MCP, gateway, configuration, persisted state, decision and utilization schemas before implementation. Preserve meaningful compatible inputs and historical interpretation; otherwise provide explicit versioned migration or actionable refusal, without rewriting ratings into evidence or deleting user state. Remove obsolete interfaces only after testing migration/refusal.
Update architecture, selection and operating-policy documentation to describe actual design and record owner decision and explicit supersession; retain historical decisions.

Acceptance Criteria

All three files are absent from source, wheel and sdist; no hidden renamed static catalog/rating/track equivalent, workflow-model selection logic or obsolete runtime loader remains.
Deterministic tests exercise discovery and positive authorized selection, exact model/effort pins, no silent substitution, hard requirements, unknown capability refusal, runtime compatibility, bounds, capacity/quota/policy and explainable exclusions. Synthetic fixtures do not claim live acceptance.
MI absent, unavailable, stale, conflicting and withdrawn evidence is handled honestly without negating independently evidenced local runtime facts; no configured executable source yields no fabricated candidates.
Historical decision/utilization evidence remains interpretable; supported existing user configuration and installed upgrades are tested; unsupported legacy inputs fail actionably. CLI, REST, MCP and gateway parity is preserved or explicitly versioned/migrated.
Focused regression tests, full repository validation (uv sync --only-dev, make check, uv run basedpyright, git diff --check), package and installed lifecycle checks pass, including clean wheel/sdist installation without removed artifacts.
A fresh separate native Kilo pr-reviewer session using GPT-6.1 Sol high independently reviews the complete frozen PR for architecture, indirect static replacement, correctness, security, identity, authorization, migration, interfaces, tests and UX. Material findings are fixed and re-reviewed within the existing finite budget. Record actual session/model evidence without inferring runtime identity.
Integrate only through canonical Forgejo PR to develop after exact-currentness approval and all validation; verify integrated acceptance before closing this issue. No direct develop push, main promotion, release, deployment, live inference, or external-repository mutation.

Constraints

This issue reflects the direct owner authorization of 2026-10-28. Implementation assignment is GPT-6.1 Terra medium; independent review is GPT-6.1 Sol high, with no silent model or effort substitution. Respect existing credentials, authorization and source boundaries. No new review launcher, reviewer environment or manual prompt relay. Preserve unrelated work. One issue, one branch, one PR.

Links

docs/architecture.md, docs/selection-policy.md, docs/security.md, docs/decisions.md, docs/llm-operating-policy.md
Historical related MI consumer: #176. Separate concurrent worker trust delivery: #224 / PR #233.
Canonical repository: https://forgejo.creatidy.com/BioMedical-IT/scarcity-router

## Goal Implement the owner's 2026-10-28 architectural decision to remove repository-versioned static model governance and route authorized executable resources using actual evidence and local authority. Supersede the parts of D-082 and earlier decisions requiring bootstrap static catalog and rating-based routing, while preserving historical decisions and valid security and provenance protections. ## Why and Baseline `selection_app.py` loads `model-catalog.json` and `model-policy.json` for selection, profiles and provenance; `model_tracks.py` and `execution_sources.py` use `model-tracks.json` for family-derived capability floors. Gateway, CLI, REST, MCP, packaging, persisted decisions and tests depend on these contracts. #176 delivered an explanation-only MI consumer with bootstrap preservation; it is closed and is not equivalent to this explicitly authorized removal. Open PR #233 addresses worker CA trust, not this work. ## Scope - Remove all three JSON artifacts, packaging and runtime dependencies, obsolete loaders, static workflow model/role assignments, assumed task profiles and model-family rating floors. Do not replace them with another static catalog, profile set or track registry. - Kernel/harness owns task intent, workflow, reviewer selection and acceptance; MI owns external evidence and history; Router owns execution sources, inventory, local capacity, compatibility, authorization, selection and admission; Console presents authorized state and commands. - Reuse existing ModelInventory, SourceRegistry, ExecutionSource, bindings, adapter evidence, capacity, authorization, MIConsumer and selection/routing cores as applicable. Construct candidates from configured authorized sources, actual discovery, evidenced adapter capabilities, applicable trusted MI evidence and local policy, without conflating discovery, execution attestation, entitlement or public knowledge. - Honor exact physical model and effort pins, hard requirements, known/unknown distinctions, authorization, evidenced bounds, capacity and scarcity; explain exclusions and limitations. Unknown quality is not a fabricated score or entitlement. An exact authorized model request need not fail solely for lack of comparative quality evidence. Empty execution-source configuration must remain honest. - Preserve trusted MI references, applicability, freshness, conflict/withdrawal handling, immutable decision provenance, independently supported local capabilities and offline usability. MI must not alter private authority or imply quota. - Inventory public CLI, API, MCP, gateway, configuration, persisted state, decision and utilization schemas before implementation. Preserve meaningful compatible inputs and historical interpretation; otherwise provide explicit versioned migration or actionable refusal, without rewriting ratings into evidence or deleting user state. Remove obsolete interfaces only after testing migration/refusal. - Update architecture, selection and operating-policy documentation to describe actual design and record owner decision and explicit supersession; retain historical decisions. ## Acceptance Criteria 1. All three files are absent from source, wheel and sdist; no hidden renamed static catalog/rating/track equivalent, workflow-model selection logic or obsolete runtime loader remains. 2. Deterministic tests exercise discovery and positive authorized selection, exact model/effort pins, no silent substitution, hard requirements, unknown capability refusal, runtime compatibility, bounds, capacity/quota/policy and explainable exclusions. Synthetic fixtures do not claim live acceptance. 3. MI absent, unavailable, stale, conflicting and withdrawn evidence is handled honestly without negating independently evidenced local runtime facts; no configured executable source yields no fabricated candidates. 4. Historical decision/utilization evidence remains interpretable; supported existing user configuration and installed upgrades are tested; unsupported legacy inputs fail actionably. CLI, REST, MCP and gateway parity is preserved or explicitly versioned/migrated. 5. Focused regression tests, full repository validation (`uv sync --only-dev`, `make check`, `uv run basedpyright`, `git diff --check`), package and installed lifecycle checks pass, including clean wheel/sdist installation without removed artifacts. 6. A fresh separate native Kilo `pr-reviewer` session using GPT-6.1 Sol high independently reviews the complete frozen PR for architecture, indirect static replacement, correctness, security, identity, authorization, migration, interfaces, tests and UX. Material findings are fixed and re-reviewed within the existing finite budget. Record actual session/model evidence without inferring runtime identity. 7. Integrate only through canonical Forgejo PR to `develop` after exact-currentness approval and all validation; verify integrated acceptance before closing this issue. No direct develop push, main promotion, release, deployment, live inference, or external-repository mutation. ## Constraints This issue reflects the direct owner authorization of 2026-10-28. Implementation assignment is GPT-6.1 Terra medium; independent review is GPT-6.1 Sol high, with no silent model or effort substitution. Respect existing credentials, authorization and source boundaries. No new review launcher, reviewer environment or manual prompt relay. Preserve unrelated work. One issue, one branch, one PR. ## Links - `docs/architecture.md`, `docs/selection-policy.md`, `docs/security.md`, `docs/decisions.md`, `docs/llm-operating-policy.md` - Historical related MI consumer: #176. Separate concurrent worker trust delivery: #224 / PR #233. - Canonical repository: https://forgejo.creatidy.com/BioMedical-IT/scarcity-router

kilo commented

2026-10-28 21:53:03 +00:00

Author

Member

Implementation in progress on local issue-234-evidenced-resources, based on 2002d348a8d6d5f3d1b742a93ba3adf944d4b24e. D-084 records the owner supersession; the static artifacts and track loader are removed in the working tree. Runtime-source admission now requires exact administrator grants, preserves unknown model bounds and rejects legacy track grants pending explicit migration. Synthetic tests cover exact pins, authorization, local resource bounds, MI receipt binding, and source lifecycle. A wheel built from the pre-change develop revision was installed and replaced with the current wheel: synthetic legacy track policy and user state were retained, and old track adoption was refused. Full validation, remaining legacy test migration, PR, independent Sol High review and integration are still outstanding; no live provider execution has been asserted. Terra Medium is owner-requested; actual effort remains unverified.

Implementation in progress on local `issue-234-evidenced-resources`, based on `2002d348a8d6d5f3d1b742a93ba3adf944d4b24e`. D-084 records the owner supersession; the static artifacts and track loader are removed in the working tree. Runtime-source admission now requires exact administrator grants, preserves unknown model bounds and rejects legacy track grants pending explicit migration. Synthetic tests cover exact pins, authorization, local resource bounds, MI receipt binding, and source lifecycle. A wheel built from the pre-change develop revision was installed and replaced with the current wheel: synthetic legacy track policy and user state were retained, and old track adoption was refused. Full validation, remaining legacy test migration, PR, independent Sol High review and integration are still outstanding; no live provider execution has been asserted. Terra Medium is owner-requested; actual effort remains unverified.

adrian.tkacz referenced this issue from a commit

2026-10-29 01:08:56 +00:00

Replace static model governance with evidenced resources (#234)

kilo referenced this issue

2026-10-29 01:09:08 +00:00

Replace static model governance with evidenced resources #235

kilo commented

2026-10-29 01:09:29 +00:00

Author

Member

Implementation published as PR #235 at 5ae0afe6175965c3d8a89477d32dad2264054b27, base/merge-base 2002d348a8d6d5f3d1b742a93ba3adf944d4b24e. Full make check passed (3,072 tests, 18 skipped; basedpyright 0/0); separate basedpyright, baseline-wheel installed upgrade, package check, Docker build/offline artifact smoke and whitespace checks passed. Source execution tests use synthetic worker/attestation, not live-provider proof. Independent Sol high review is pending; issue remains open.

Implementation published as PR #235 at `5ae0afe6175965c3d8a89477d32dad2264054b27`, base/merge-base `2002d348a8d6d5f3d1b742a93ba3adf944d4b24e`. Full `make check` passed (3,072 tests, 18 skipped; basedpyright 0/0); separate basedpyright, baseline-wheel installed upgrade, package check, Docker build/offline artifact smoke and whitespace checks passed. Source execution tests use synthetic worker/attestation, not live-provider proof. Independent Sol high review is pending; issue remains open.

kilo referenced this issue

2026-10-29 01:09:58 +00:00

Replace static model governance with evidenced resources #235

adrian.tkacz referenced this issue from a commit

2026-10-29 02:22:49 +00:00

Scope worker telemetry and preserve target effort (#234)

kilo referenced this issue

2026-10-29 02:23:23 +00:00

Replace static model governance with evidenced resources #235

adrian.tkacz referenced this issue from a commit

2026-10-29 02:57:25 +00:00

Preserve executable pin effort and versioned provenance (#234)

kilo referenced this issue

2026-10-29 02:58:01 +00:00

Replace static model governance with evidenced resources #235

adrian.tkacz referenced this issue from a commit

2026-10-29 03:33:46 +00:00

Scope logical model resolution to authorized resources (#234)

kilo referenced this issue

2026-10-29 03:34:16 +00:00

Replace static model governance with evidenced resources #235

adrian.tkacz referenced this issue from a commit

2026-10-29 04:28:21 +00:00

Preserve live tool capability and exact runtime evidence (#234)

kilo referenced this issue

2026-10-29 04:29:20 +00:00

Replace static model governance with evidenced resources #235

kilo commented

2026-10-29 04:43:02 +00:00

Author

Member

Integrated acceptance verified via PR #235. Forgejo merge commit 516c155963c3cd06cb9c3dd74323fe4ddaf1b6ed is current develop, with approved base 2002d348a8d6d5f3d1b742a93ba3adf944d4b24e and reviewed HEAD cd9a74fbd9ce378b59ccc5668aa3e07e16c27bd5 as its two parents. The integrated tree 0a46f16c8fcc7e4421c3e0515db6c4a347177903 exactly equals the approved/validated HEAD tree; approved HEAD is an ancestor. Root tree contains none of the three removed artifacts.

Review #5 independently APPROVE with no findings in native pr-reviewer session ses_e7a1424b2ffep2G2PGXBweaq81; dispatch resolved GPT-6.1 Sol/openai/high, backend runtime identity independently unobserved. make check passed 3,090 tests (18 skipped), standalone basedpyright 0/0, baseline-code installed upgrade/package checks, Docker build and offline image smoke passed. A clean exact-head evidence checkout was used because an unrelated Agent Manager temporary file in the original checkout was preserved. Synthetic worker/source tests are not live-provider acceptance; context evidence remains explicitly administrator-attested. No main promotion, release or deployment.

Integrated acceptance verified via PR #235. Forgejo merge commit `516c155963c3cd06cb9c3dd74323fe4ddaf1b6ed` is current `develop`, with approved base `2002d348a8d6d5f3d1b742a93ba3adf944d4b24e` and reviewed HEAD `cd9a74fbd9ce378b59ccc5668aa3e07e16c27bd5` as its two parents. The integrated tree `0a46f16c8fcc7e4421c3e0515db6c4a347177903` exactly equals the approved/validated HEAD tree; approved HEAD is an ancestor. Root tree contains none of the three removed artifacts. Review #5 independently APPROVE with no findings in native `pr-reviewer` session `ses_e7a1424b2ffep2G2PGXBweaq81`; dispatch resolved GPT-6.1 Sol/openai/high, backend runtime identity independently unobserved. `make check` passed 3,090 tests (18 skipped), standalone basedpyright 0/0, baseline-code installed upgrade/package checks, Docker build and offline image smoke passed. A clean exact-head evidence checkout was used because an unrelated Agent Manager temporary file in the original checkout was preserved. Synthetic worker/source tests are not live-provider acceptance; context evidence remains explicitly administrator-attested. No main promotion, release or deployment.

kilo closed this issue

2026-10-29 04:43:13 +00:00

kilo referenced this issue

2026-10-29 18:02:14 +00:00

Persist origin-bound worker-local private CA trust (#224) #233

kilo referenced this issue

2026-10-29 19:25:19 +00:00

Persist origin-bound worker-local private CA trust (#224) #233

kilo referenced this issue

2026-10-29 19:27:05 +00:00

Persist origin-bound worker-local private CA trust (#224) #233