knowledge: distill lithos-metal into source-grounded candidate skills #115

Activity

1. Dong1017

   added

   area:knowledgeKnowledge line: packs, skill library, knowledge admissionKnowledge line: packs, skill library, knowledge admission

   kind:taskBounded delivery task with a delegated sliceBounded delivery task with a delegated slice

   on Oct 23, 2026
2. Dong1017 commented on Oct 23, 2026

   

   Dong1017

   on Oct 23, 2026 – with ChatGPT Codex Connector

   Last edited by Dong1017

   Collaborator

   More actions

   Issue-owner takeover: complete the existing Lithos Metal contribution

   The requester explicitly delegated takeover of Issue115 and its existing PR #117 to this Codex parent on 2026-10-29, through source-scope acceptance and evidence-based closure. I am the accountable Issue owner for this continuation. The original contributor's work and revision-bound evidence will be preserved; the stopped local overlapping draft remains inactive.

   The accepted scope remains the pinned Lithos Metal candidate graph: compiler/fusion, GPU-resident generation and DSpark state, and serving/cache guidance. Native full-model, numerical, quality or performance qualification is outside this source-only acceptance.

   1. Check the pinned source 6653b7df5f2c4392ec7b038960f2796b1499923f, retained licenses/readsets, the four-node graph, existing consumers and the original source review. Keep earlier CPU and synthetic Metal reports attributed to their original author and tested revisions.
   2. Reconcile PR knowledge: add source-grounded Lithos Metal candidate skills #117 with actual current main 7d1d82f55cda89842bd23ee2da91b521ee60eb08, preserving all accepted canonical owners, registry/matrix additions and resource contracts. Commit the combined canonical source before regenerating native projections from that actual source. Resolve real inventory-test changes without relaxing per-record checks.
   3. Address the existing PR-body automatic-closing pattern, retain plain Refs #115, and publish truthful final-head author/declaration and verification records. New material/integration authorship uses an actual fresh Sol/max instance; the original Sol/xhigh authorship remains historical and is not relabeled. The PR enables maintainer edits; the branch-specific publication path will be checked before updating it.
   4. Obtain a fresh, independent Sol/ultra review of the complete final head and all applicable current Ready, installed-package and admission checks. Only then perform the authorized squash integration, verify the accepted tree/post-merge checks, publish the Delivered / Verified / Limits acceptance card, and complete Issue closure if every source-scope criterion is met.

   This continuation uses PR #117 and creates no duplicate delivery PR. No model downloads, dependency upgrades, new Metal/GPU/model/provider experiments, engine integration or evidence promotion are included. Existing incomplete or unrun claims retain their actual limits.

   Original claim and duplicate-avoidance update (retained)

   Update (2026-10-23): PR #117 now covers this accepted source-distillation scope. Following the requester's duplicate-avoidance rule, we have stopped our local overlapping draft and will not open a second contribution PR for this target. The PR author's verification reports have not been independently rerun in this task. The Issue remains open for owner acceptance.

   ---

   Claiming this source-distillation contribution at the requester's direction. I will deliver an independent PR using lithos-ai/lithos-metal@6653b7d and AREX-Skill@ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6, covering the three accepted areas: Metal task compilation/fusion; GPU-resident generation and DSpark state; serving/cache validation.

   The deliverable will retain English candidate guidance, pinned readsets and license/notices, integrate the existing owners/catalog/matrices, and include applicable CPU/static and packaged-resource checks plus independent revision-scoped review. Unrun Metal/full-model checks will remain explicit; no model downloads or performance qualification are included. The contribution PR will use a non-closing reference, and this issue will remain open for owner acceptance.
3. Dong1017

   self-assigned this

   on Oct 29, 2026
4. Dong1017 commented on Oct 29, 2026

   

   Dong1017

   on Oct 29, 2026 – with ChatGPT Codex Connector

   Collaborator

   More actions

   Accepted: Lithos Metal source-grounded candidate knowledge

   flowchart LR
       A["Pinned Lithos source + AREX method"] --> B["Compiler / fusion"]
       A --> C["Generation / DSpark state"]
       A --> D["Serving / cache lifecycle"]
       B --> E["Owner-local candidate graph + packaged discovery"]
       C --> E
       D --> E
       E --> F["Source, CPU and package checks verified"]
   

   Loading

   • Delivered: PR #117 is squash-merged as ab03f3ff. One registered Lithos owner and three local subskills preserve pinned readsets, licenses/notices, navigation, both matrices and a bounded diagnostic.
   • Verified: Independent Sol/ultra final-head LGTM; four Ready jobs and current declaration admission passed. Both Python versions passed 2,509 tests with 11 skips, one warning and 284 subtests, and read 125 records from fresh installed wheels. Actual post-merge CPU smoke passed eight tests; post-merge declaration checks passed.
   • Limits: All four Lithos nodes remain candidate / empirical_validation: not_run. Full-model continuation, DSpark numerical/quality validation, native-backend qualification, service performance and knowledge efficacy remain unverified. The original Mac reports remain historical author evidence.

   Full acceptance record

   The delegated Issue owner accepts all four completion criteria in the current Issue's source-only scope, under the published takeover plan.

   Accepted criterion	Delivered artifact and decisive evidence
   Fixed upstream, readsets, licenses/notices and mechanism limits	Immutable source/provenance and notices, retained LICENSE, upstream 6653b7df5f2c4392ec7b038960f2796b1499923f, AREX method ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6. The independent reviewer matched the license/notice blobs and inspected the source contracts.
   Compilation/fusion, GPU-resident generation/DSpark and serving/cache graph	Root procedure and three owner-local subskills. Twenty-eight local links resolve; metadata and ownership remain explicit.
   Navigation, catalog, both matrices and packaged access / applicable CPU checks	Reference matrix, search-space matrix, native generated discovery, complete Ready full/wheel checks. The new author executed 13 contract and 27 focused tests, Ruff, real bundle/resource reads, and a noneditable isolated wheel check.
   Native/model gaps preserved and independent revision-scoped review	Final-head independent review linked above, with 13 CPU contract tests personally rerun by the reviewer and an explicit source-only disposition of the zero-comparison advisory. No native/model or efficacy result was invented or promoted.

   Revision binding: assessed main 7d1d82f55cda89842bd23ee2da91b521ee60eb08; real canonical source M 26743abcfae48cd6223bd757e74f193d97e3e5a7; native generated G eebd09469ca9a28fd50dc6da7c4dd90485ce17b1; final reviewed/public head T 2709a51bee9f5e27328d24b0a07dc30c6e13fe27; actual squash ab03f3ffb8f2e5bc02c304828c5c41890eef673e. The squash tree 76830cf6f5686d6800b85afcf1a2164ecbbdf3c0 equals T's reviewed tree and has the assessed main as its parent. All 540 committed source inputs match the native projection; all 76 prior discovery records retain their semantics.

   Execution boundaries: the local wheel and 27 focused tests ran on G; T only removed three inherited test-module descriptions, with independently verified executable-AST equality and identical package/source/generated bytes. The final T has its own complete Ready checks and clean-wheel runs. Local private validation reused the existing real PyYAML 6.0.3 distribution with byte equality; it installed no upstream runtime or new model dependencies.

   The current Skill evaluation advisory reports zero comparisons, zero model calls and 111 uncovered paths. Plan/publish succeeded; execute was conditionally skipped and is not counted as a pass. Pricing/budget limitations were not bypassed. These gaps provide no efficacy evidence and do not add a provider/device requirement to the accepted knowledge-only criteria.

   The original 153 upstream CPU, nine synthetic M4 Metal and 697-suite numbers are retained historical author reports, not reruns on this integration. Actual material authorship remains original Sol/xhigh plus new Sol/max; review-only Sol/ultra is excluded from material-author declarations. No campaign PASS, production readiness, model quality or performance acceptance is claimed.
5. Dong1017

   closed this as completedon Oct 29, 2026