Repository navigation
New WSL local models: settle the generation and embedding slots by their preregistered measurement - #647
Conversation
seathatflowsinourveins
added
the
lane:foundation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b9d0bf6298
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
SOURCE ACCEPT for bounded historical evidence publication at b9d0bf6, against published main d2777ee. The consequential Astra/max source read returned that disposition; the native review process closed 0. This permits publishing the limited measurements. Final architecture, destination and SDK role qualification remain open. The decision expressly labels both results as local integration checks on one workstation/throwaway distribution, one run per case and arm, and excludes destination acceptance or a claim to the best available model. The settlements become The publication map has 35 public copies (32 declared identical, three declared sanitizations) and 196 private metadata entries. I independently checked all 35 published pinned Git blobs against their declared byte counts and SHA256: 35 match, 35 native exits 0. Map SHA256 Independent processing of the exact head/main registry originals confirms 9420→9464 file rows, 187→189 receipts, 44 added rows, 19 changed bindings and no removed rows; all previous receipts and all 26 convergence records remain. Head registry SHA256 Failed generation arms, missing failure text, co-residency limits and unrun reranking remain evidence limits. Pinned Ollama Modelfile documentation supports the relative-GGUF/context configuration mechanism. The pinned keep-alive implementation supports the documented residency difference. Actual loading, served dimensions, GPU ownership/restoration, residency, destination install and broader role quality still require their own native evidence. No evaluated model, benchmark, GPU or destination operation was run for this review. The source review itself was one requested Astra/max model run; actual served model/effort, billing and upstream request count remain unknown. Required CI still governs merge eligibility. |
…d local-model rows ManifestRuleTests counted 38 installing rows (36 of the 64-row plan plus the two consensus rows). Settling local-generation-model and embedding-model adds two installing rows, so the plan installs 40 (check_plan.py: 40 installed, 38 by the default run, 2 only when named). The set equalities of the same test already held; only the count was stale. Found by the hosted validate run of #647 (run 37110319677). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
SOURCE ACCEPT carried to ae90cb2, one commit directly atop root-accepted b9d0bf629883e7719c76e0655aa59c4adcbd3e92. The previous bounded historical-evidence scope and all destination/runtime/role limits remain. The immutable delta changes only tests/test_new_wsl_client_config.py:438: comment plus expected plan-row count 38→40, accounting for the two settled local-model rows. The preceding set-equality assertions are unchanged. Diff +3/−2; no behavior, evidence or model result changes. Both heads bind the registry to Git blob The old-head validate failure, completed09:08:08, remains a failure. At09:28:59 the new-head validation was in progress and macOS jobs queued. The owner's 146-tests/validator0 report is distinct from a completed hosted check. Seven original native API captures returned0; the immutable compare and exact changed lines were read independently. No tests, models or host operations were rerun for this small source repair. |
|
FINDINGS / required macOS gate open, exact #647 head The existing mac job111173581589/run37112688327 failed11:20:17Z; step20's full suite failed11:20:06Z. Its original artifact11273275413, bound to this run and exact head, contains 9849 tests /1737.068s /two failures /1328 skips. The only failures are generation-files L1133 and embedding-files L1149, each the synthetic expected-state case returning1 when0 is required. These are integration fixtures, not actual model runs. Both execute the exact install-plan acceptance programs, embedding L846 and generation L1569: This is a source-supported portability lead. The failed fixture captures then discards inner stdout/stderr atL1105–1106, so original CI does not identify whether checksum or later grep failed. Do not claim the precise inner cause has been reproduced. Concrete owner repair: reuse that existing capability probe/fallback for the synthetic fixture and expose failed subprocess diagnostics. Preserve its positive and planted-negative oracles and the production WSL acceptance programs. Then obtain native macOS evidence for the repaired head. No GPU/provider/model operation is needed for these synthetic checks. Root hands off the owner paths and has changed no Claude-lane files. One authorized isolated Linux native run at this exact head passed all five Root independently verified21 non-auth original command stdout/stderr bindings and seven pinned source payload identities, actual artifact run/head and ZIP-member equality. The packet retains22 commands:20exit0, two retrieval exit1s ( |
|
Codex exact-head SOURCE ACCEPT for the checksum-shim and diagnostic repair; FINDINGS before landing at The repaired fixture uses the existing Landing finding: the registry still carries eight foreign bindings that differ from current main: The original Mac run's two positive fixture failures remain retained. Their original inner stderr was not supplied, and source inspection does not authenticate the proposed checksum cause or establish that the repaired Mac run passes. The earlier passing Linux controls retain their synthetic/integration scope. A successful native repaired Mac result is still needed for that claim. Custody: |
…d local-model rows ManifestRuleTests counted 38 installing rows (36 of the 64-row plan plus the two consensus rows). Settling local-generation-model and embedding-model adds two installing rows, so the plan installs 40 (check_plan.py: 40 installed, 38 by the default run, 2 only when named). The set equalities of the same test already held; only the count was stale. Found by the hosted validate run of #647 (run 37110319677). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
febc6e8 to
bd52c9d
Compare
|
Root exact-head SOURCE ACCEPT for the landing-carry repair. Head bd52c9d, base/captured main 4ced292. Fresh native guard13:01:14 confirms OPEN/non-draft. This clears the eight foreign-binding landing finding in the febc source read. Root compared complete, non-truncated Git trees at the accepted febc6e8 and current head. Their literal delta is12 paths: the registry and11 carried current-main paths. Each of those11 matches main's Git identity exactly. All65 PR file paths other than the registry retain their accepted old Git identities. The three-dot API comparison uses the older d277 merge base and is retained as that comparison, rather than treated as a literal old-to-new diff. All9401 foreign current-main file bindings now match;9464 registry membership and current-main order are preserved. The eight previously stale file rows are the only old-to-new registry-row changes. Owned rows,189 receipt rows (including all187 current-main rows) and26 convergence records remain unchanged. The checksum test is byte-identical to accepted febc: Git blob b0941052293fc59d5bed34e586361cb2652630a8,78877B/SHA256 The prior checksum capability/shasum fallback source acceptance carries forward. Old macOS failure output and synthetic Linux fixture results remain distinct from this head's native macOS/host/GPU qualification. Required CI must pass at this head before owner merge; this read did not rerun tests or prove the cause of a historical failure. Original source custody: complete held/head trees SHA256 |
…records and receipts A sanitized copy of the preregistration (one replacement: the throwaway distribution's name) with its hash chain and copy notes; receipts for the generation slot (Part A and its extension) and the embedding slot (Part B); the per-case validity files of the four scored A2 runs and the per-query nDCG@10 files of E1 to E3, from which both decision statistics reproduce unchanged with the published scripts; the frozen scorers, runners and steps; a reconstruction of the measured Swift Modelfile from the Ollama library's blobs (sha256 8911245e... with the measurement's FROM line); files.json lists every file of the private measurement folder by sha256. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…l by their measurement Two settlements in the form of the model-server settlement, each with the preregistration copy and its receipt. Both slots are rows that the convergence decisions add, so the assembler now settles such a row right after those decisions (it must be a split of the rounds; its resolution stays as recorded; its installed job stays owned by one row). The record's tables show a returned measurement's settlement also on a split row. Tests follow the new state (58 rows install, 6 measurement, 5 split); one control refuses a settlement aimed at an added row that is not split. Manifest, tables, handbook and the handbook receipt's output hashes regenerated. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… sources local-generation-model downloads the Swift IQ3_S file at its Hugging Face revision with its sha256, places the repository's two Modelfiles beside it and creates swift-iq3s-s2o and swift-iq3s-s2o-64k (context 64,000 as the model's own parameter). embedding-model pulls qwen3-embedding:0.6b, stops unless the server lists the pinned library manifest digest, and creates qwen3-embedding-8k (context 8,192). No server-wide context is set. Both rows create their models through the running model server, which the plan does not start, so they install and are checked only with --only. Acceptance: post_install reads files only (the placed Modelfiles, the library manifest's digest, the created models' layers); service_health shows each model's own context and makes one short generation or one embedding call. check_plan.py now checks the dispatch of rows installed only when named in both directions; LocalModelAcceptance runs the four programs against stand-ins. README, SOURCES and VALIDATION say what was read and checked, and that no command of the two rows has run as a plan row anywhere. GPU ownership between the workstation's model services and the destination's server is left to the lifecycle design, as a stated limitation. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…preregistered measurement docs/decisions/2026-10-03-new-wsl-local-models.md records the deciding measurements as the manifest names them, the preregistration and its amendments, the arms, gates and rules of Part A and Part B, both results with their decision statistics, the four deviations and the recorded defects, the other model family's read (pull request 608, comment 5964331790) and the artifact map it asked for, what the install plan does, what is not established (the destination's acceptance, physical co-residency in pb2, one host and one task each, and the wider measurement the critic's own text named), the sources by sha256 and the overturn conditions. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…placed, server smoke check on the settled embedder) - .gitignore: exceptions to *.jsonl for the four per-case validity files and the three per-query nDCG@10 files, which both decision statistics reproduce from; the seven files are now committed (hashes unchanged, as files.json and both receipts record them). - scripts/steps/W-partb-window.sh: the published copy named the throwaway distribution on line 18; it now writes '<the throwaway distribution>' (quoted, still parses with bash -n). files.json marks it changed with both hashes; COPY-NOTES gives 32 identical and 3 changed copies, corrects the byte-identical list and records the distribution-name scan the first publication lacked. - files.json: the 231 files are described as listed; the copy-out added to the private folder afterwards is listed by its checksum list, with the measured Swift Modelfile (sha256 8911245e...). README and the decision record no longer say that file was not retained. - part-a receipt and the decision record: the control's A1b footprint is 12.97 GB (12,968,494,366 bytes, 12.08 GiB), not GiB. - Receipts' files.json hash, settlements, manifest, handbook and its receipt regenerated; the manifest's state counts are unchanged. - Install plan: the local-model-server row's after_sign_in smoke check is one /api/embed call to qwen3-embedding-8k instead of `ollama run embeddinggemma`, which would pull the losing arm; a stand-in test covers it. README and SOURCES: the embedder's 32,768-token load, the private step scripts listed by hash, and the measured server's NUM_PARALLEL/KEEP_ALIVE against v0.35.0's defaults. - Decision record: cites the Codex lane's binding read of the artifact map (file bindings only; an exact-head verdict is still required). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…smoke check The repair round cited server/routes.go:975-977 at cc4069396f for "/api/embed answers 404 for a missing model". Those lines answer 404 only when getExistingName fails, which happens when listing the manifests fails; for a model the store does not have it returns the name unchanged, GetModel fails with the manifest's not-found error, and handleScheduleError answers 404 "model %q not found, try pulling it first" (routes.go:981-984 and 3226-3227). EmbedHandler (928-1151) calls no pull. accept.sh's comment, the local-model-server row's notes and SOURCES.md now cite those lines; the stand-in test's error answer is that message. The check itself is unchanged. check_plan.py exit 0 (69 rows; 74 commands and 61 acceptance entries agree); bash -n accept.sh exit 0; LocalModelAcceptance: 5 tests OK. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d local-model rows ManifestRuleTests counted 38 installing rows (36 of the 64-row plan plus the two consensus rows). Settling local-generation-model and embedding-model adds two installing rows, so the plan installs 40 (check_plan.py: 40 installed, 38 by the default run, 2 only when named). The set equalities of the same test already held; only the count was stale. Found by the hosted validate run of #647 (run 37110319677). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… not GNU's, and report a failed program's output The required macOS job of run 37112688327 (job 111173581589) failed the expected-state cases of test_the_generation_files_check and test_the_embedding_files_check with exit 1. The fixture discarded the program's output, so the inner cause was not recorded. Both programs run `sha256sum --check --status`, and setUpClass only checked that a sha256sum was on PATH. The fixture now reuses tests/test_adoption_bootstrap.py's probe sha256sum_checks_like_gnu and, where it is false, writes the same scratch sha256sum that runs `shasum -a 256` as that module's run_install_pin; setUpClass requires shasum only then. run_program returns the finished process, and each case's assertion message carries the program's exit code, stdout and stderr. The case tables, the expected exits and the install plan's programs are unchanged. The test file's re-registration in manifests/evidence.json is in the branch's last commit (docs/lanes.md, hot-file protocol). Linux only: the module passes, and with a stub sha256sum that rejects --check --status first on PATH (a simulation, not macOS) all five LocalModelAcceptance tests pass through the shasum shim, where the previous head failed the same two expected-state cases. Native macOS evidence comes from CI on this head. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… skip the server smoke check until the embedder exists install.sh's model_server_answers, which both local-model rows run before their first command, accepted any server that answered `ollama ls`. It now also stops the row unless GET /api/version reports 0.35.0, the measured server (docs/api.md:1824 and server/routes.go:2023 at cc4069396f), before anything is downloaded, pulled or created. The local-model-server row's after_sign_in check calls qwen3-embedding-8k, which only the embedding-model row creates, and step F9 runs that stage without any model row, so it failed there every time. accept.sh now prints skipped, which is not a pass, until that model's manifest is in the store the embedding row's post_install check reads; the check itself is unchanged. No command or acceptance entry of install-plan.json changes; its notes, the README, SOURCES, VALIDATION and the decision record describe both repairs. LocalModelAcceptance runs accept.sh's stage itself (skipped, 0, 1), and ModelServerGuard runs the guard against stub ollama and curl (0.35.0 passes; another version, no version, a failed request and no server fail; both rows run it before their first command). The local-model evidence README notes that M6-a1b.py's docstring repeats A1's order while its code and record follow amendment 2's; the hash-bound copy is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bd52c9d to
e471dee
Compare
|
Exact-head FINDINGS on Three of the five review concerns have source corrections:
Two historical evidence concerns remain open. Please preserve the scripts that ran and add an impact/qualification disclosure bound to their existing original evidence; editing an as-run script would not repair a past trial.
Custody and landing: 90 old→new delta paths consist of 80 exact main carries and 10 owned modifications; the other 56 paths of the complete 66-path PR retain their accepted Git identities. All 35 publication-map members, both historical receipts and all three reviewed historical scripts are unchanged. I verified 28 complete repository originals, two pinned upstream blobs and 46 original native stream pairs, all exits 0, against byte/SHA256/Git bindings. No measurement, model, private pilot or test was executed for this read. All 9,478 foreign registry row payloads and 191 existing receipts in order are preserved, as are the 29 convergence records and unowned metadata. Two adjacent foreign decision rows are reordered into sorted order: the SDK-worker and PR320-retirement decisions, the same explicit normalization proposed in #666. This is a narrow, identified order change, not row loss. The owner should carry that accepted normalization consistently with #666 and check whole-result sortedness in addition to payload preservation against actual main before landing. Source acceptance does not substitute for the required current-head checks, GitHub thread disposition or landed-tree observation. |
…heir impact W-partb-window.sh's precheck (lines 18 and 28) returned the status of the cleaning pipeline rather than the remote precheck's, so a failed precondition could not stop a Part B window. The new section binds each admitted window's precondition to the precheck's own printed lines in raw/_gpu-window.txt (sha256 42c3758a...): pb1 at 23:13:16Z and 23:13:27Z, pb2 at 23:48:48Z and 23:48:55Z, each recording resident=[] llama-server pids=[]. M11p-precheck.sh exits 0 exactly when those two printed values are empty. Window pb1 ran the earlier revision amendment 4a hashed (688722f7...), which was not retained, so the defect is stated for the retained revision only. M11-g0-run.sh line 12 prints its checksum count without testing it. The section binds "prepared files unchanged" to each G0 record's printed count of 0 (raw/M11-g0-S1.txt 4849c444..., raw/M11-g0-S2.txt 09046e46..., raw/M11-g0-S1b.txt e85547ce...) and notes that every arm that ran G0 failed it, so no decision rests on a G0 pass. Both scripts stay as they ran. Answers PR #647 review threads r4172365849 and r4172365852. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…o receipts on main's copy Start from main's manifests/evidence.json (ecea286) and apply this branch's entries through the docs/lanes.md hot-file protocol: host_receipts.register_file for the 63 files the branch registers (44 new, 19 re-hashed, now including the review repairs to install.sh, accept.sh, install-plan.json, the install plan's README, SOURCES and VALIDATION, the local-model evidence README and tests/test_new_wsl_definitive_defaults.py) and the two receipts[] entries re-added at the end (part-a-receipt.json and part-b-receipt.json of new-wsl-local-models-20261002). convergence_records[] is unchanged. This commit replaces the previous evidence-only commits bd52c9d and e471dee. Against e471dee it changes one entry: the local-model evidence README, re-hashed after the commit before this one added its known-defects section (sha256 deeac61a..., 13,179 bytes). files[] keeps e471dee's order. Main's copy at ecea286 lists two files out of path order (docs/decisions/2026-10-03-omniroute-sdk-worker-0160.md after docs/decisions/2026-10-03-retire-pr320-loki-denominator-host-receipts.md), which scripts/validate.py rejects; scripts/evidence_manifest.py --write, the command validate.py names, swapped those two entries and changed nothing else. component_matrix.py --write and new_host_grand_list.py --write leave the manifest unchanged; scripts/validate.py passes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
e471dee to
4c033f1
Compare
|
Codex root exact-head SOURCE ACCEPT WITH NOTES for the two historical-script impact disclosures at The added README section states both defects and their consequences: the window wrapper masked the remote precondition exit; G0 printed a checksum diagnostic without gating the subsequent runs. It preserves the as-run scripts and names the printed observations on which the owner bases the impact assessment. It also expressly records that pb1's wrapper revision was not retained, so that revision's defect cannot be established from source. This is the needed historical disclosure, not a retroactive repair of an enforced gate. Evidence scope: I verified the public source, artifact metadata and bindings. I have not read the four private witness bodies or independently authenticated their printed observations. The statement at L68 that the defect leaves no qualification unknown is therefore the owner's assessment from those named records, not new native qualification by this root. The same boundary applies to the reported three G0 failures and unchanged-file diagnostic. No new run or causal/performance/default acceptance follows. The complete E471→4c03 delta changes only README and its registry row: README13,179B, SHA256 Original native captures remain retained, including the initial GraphQL1 from querying #608 as an issue and its corrected PullRequest recovery0. The fresh root guard returned0 and this exact head/main. No root test, model, GPU, service, private-log or runtime execution occurred; CI, prospective merge preservation and installed/default qualification remain separate. |
…utation runs in manifests/evidence.json Hot-file protocol (docs/lanes.md): this is the branch's last commit, on main d2fc380 (#647, #661), and takes main's manifests/evidence.json. It re-registers the closure record, the receipt and docs/harness-defaults.md, and the 71 files under evidence/artifacts/frozen-variant-guard-mutations-20261003/. component_matrix --write and new_host_grand_list --write changed nothing else. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…k is regenerated on top of main's #647 version (only this PR's research-state.json input-hash line) and the receipt's two output hashes are refreshed (standing ACK); registry carries main's rows plus the owned bindings Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…otocol) #647 settled local-generation-model and embedding-model by their preregistered measurement; this branch adds the wave-2 records, interim installs, statusline and the client-configuration map. Both survive. 15 paths conflicted. Regenerated with the repository's generators, not hand-merged: - definitive-manifest.json: evidence/artifacts/new-wsl-definitive-defaults-20261001/assemble_manifest.py (90 slots, 31 definitive, installed 59, interim 3; by_state resolved 23, measurement 6, split 5; consensus 6), then --check. - docs/decisions/2026-10-01-new-wsl-definitive-defaults.md: its only conflict was inside the generated tables; render_tables.py --write, then --check. - docs/new-wsl-handbook.json and .md: scripts/build_new_wsl_handbook.py --write. - evidence/artifacts/new-wsl-handbook-20261001/receipt.json: no generator writes it; the two output digests come from the builder's own digest function (as --write printed them); generator, profile and inventory were unchanged. Hand-merged, keeping both sides: - install.sh: both header notes; both functions (interim_acknowledged, model_server_answers); --list takes this branch's code-search line and main's embedding-model line; the dispatch loops take this branch's selected lists and main's named-only list (local-generation-model, embedding-model). - accept.sh: both header notes; the skipped-slot loop drops all five rows that now install (code-search, memory-owner, context-supply, embedding-model, local-generation-model), in plan order. - install-plan.json, owners.json: revised_at_utc set to the merge's time (2026-10-03T18:26:54Z); the plan's status keeps main's sentence, then this branch's. Rows auto-merged (disjoint: 13 rows here, 3 there). - README.md: counts from the merged plan (70 rows: 44 installed, 42 by the default run with three interims and two only when named, 3 measurement-only, 23 not installed; post_install 28/15/1, service 10, after sign-in 7; route counts with model-server 2); the default-run rule keeps the interim clause and main's named-only sentence. - README.md, SOURCES.md, VALIDATION.md: both new sections, in the order written (the local-model sections first written 05:25Z, wave 2 at 10:16Z), which VALIDATION's opening line requires. VALIDATION's opening line now says 70 rows and names the Wave 2 section, and Wave 2's checks list records the merged plan's check_plan.py line. - tests/test_new_wsl_client_config.py: 44 planned rows. - tests/test_new_wsl_definitive_defaults.py: the merged counts above; both sides' test classes (StatuslineInstallAndAcceptance; LocalModelAcceptance, ModelServerGuard). manifests/evidence.json: 54a96ff's copy (git checkout --theirs), then host_receipts.register_file for the branch's 55 registered files (17 of them re-hashed by this merge); receipts[] and convergence_records[] are main's. component_matrix.py --write and new_host_grand_list.py --write change nothing; scripts/validate.py passes (9542 hashed files). check_plan.py: OK: 70 rows: 44 installed (42 by the default run, 2 only when named), 3 measurement-only, 23 not installed; 94 commands and 66 acceptance entries agree with the scripts. bash -n install.sh and accept.sh: exit 0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-pin-20261003 Refresh onto main cac8700 (#651, #628, #647, #661, #648) under the docs/lanes.md hot-file protocol. - Handbook: main's docs/new-wsl-handbook.{md,json} and evidence/artifacts/new-wsl-handbook-20261001/receipt.json (#647) are the base; scripts/build_new_wsl_handbook.py --write regenerates both outputs with this PR's profile, and the receipt's two outputs hashes and profile_sha256 follow the regenerated files and the profile. Nothing else in the receipt changes. - manifests/evidence.json: three-way registry merge (main's rows kept, this branch's rows applied, the 87 branch-touched files rehashed from the merged tree). - #628's Codex SDK worker pins openai-codex and openai-codex-cli-bin 0.160.0 with the same artifact hash sets as adoption/sdk/requirements-linux-x86_64-py313.lock and cites rust-v0.160.0 a956835d; it changes none of manifests/stack.json, adoption/pins-linux-x86_64.json, tools/adoption/apply_codex_lane.py or adoption/templates. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1 participant
Scope
local-generation-model: Swift-1.5-Qwen3.8-27B IQ3_S on Ollama 0.35.0 asswift-iq3s-s2o-64k;embedding-model:qwen3-embedding:0.6basqwen3-embedding-8k). It publishes the measurement's sanitized records and two receipts, records both settlements in the definitive manifest (with the manifest, the definitive-defaults tables and the handbook regenerated), makes the two install-plan rows installable from pinned sources, and adds the decision recorddocs/decisions/2026-10-03-new-wsl-local-models.md.d2777ee76a6ca14e020856fcd76421ceebc1e4f0lane:foundation. Themanifests/evidence.jsonchange is the hot-file protocol's re-registration, which does not by itself make a PR shared (docs/lanes.md, "Labels and PRs").evidence/artifacts/new-wsl-local-models-20261002/(new): the preregistration's published copy, its hash chain and copy notes, the two receipts, the per-case and per-query records, the frozen scripts, the Modelfile reconstruction andfiles.json.evidence/artifacts/new-wsl-definitive-defaults-20261001/:settlements.json,assemble_manifest.py,controls.pyandrender_tables.py;definitive-manifest.jsonis generated.evidence/artifacts/new-wsl-install-plan-20261002/:install-plan.json,owners.json,install.sh,accept.sh,check_plan.py,models/*.Modelfile, README, SOURCES and VALIDATION.evidence/artifacts/new-wsl-handbook-20261001/receipt.json,tests/test_new_wsl_definitive_defaults.py(classLocalModelAcceptance), and.gitignore(two exceptions to*.jsonl).docs/decisions/2026-10-03-new-wsl-local-models.md(new). Generated: the tables ofdocs/decisions/2026-10-01-new-wsl-definitive-defaults.md,docs/new-wsl-handbook.jsonanddocs/new-wsl-handbook.md.manifests/evidence.json.SOTA sources
v0.35.0oncc4069396f3ad2c370c53eed2e4a42ac13adab84. Each command of the two rows is read at that commit (evidence/artifacts/new-wsl-install-plan-20261002/SOURCES.md, "The two local-model rows"):docs/cli.mdx#L85andserver/images.go#L1082: pull and the layer digest check.cmd/cmd.go#L2422anddocs/modelfile.mdx#L120,#L126,#L146: create from a GGUF file, a relativeFROM, andnum_ctx.envconfig/config.go#L112,#L277andmanifest/paths.go#L29: the model store and theOLLAMA_NUM_PARALLELdefault.docs/api.md#L48,#L1351and#L1659: generate withthinkfalse, list, and/api/embed.server/routes.go#L928-L1151,#L981-L984and#L3226-L3227: the embed handler answers 404 for a missing model and pulls nothing.d74895bbe5db4bec1e0024e7cc87d59c02d7631a. The file isSwift-1.5-Qwen3.8-27B-GSQ-RCO-IQ3_S.gguf: LFS sha2561333c6ea70ef348d4ac6d62732772e8ad6571ac5b3754c14ed54f1a0d904a786, 11,771,546,912 bytes. It is served with the Modelfile lines of the Ollama library modelqwen3.8:27b, whichevidence/artifacts/new-wsl-local-models-20261002/reconstruct_s2o_modelfile.pyrebuilds independently.qwen3-embedding:0.6b(Q8_0), manifest digestac6da0dfba84a81fdbfbaf330198c33cd77c4cdfc53e8bc50eb581914a15621d. Upstream: https://github.com/QwenLM/Qwen3-Embedding. Model card: https://huggingface.co/Qwen/Qwen3-Embedding-0.6B/blob/97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3/README.md#L36.bd59dd3b48974ad2e91219a6ceed41011e201163), taskinspect_evals/bfcl -T categories=simple_python.79857d39), taskCQADupstackUnixRetrieval, test split.evidence/artifacts/new-wsl-final-architecture-20261002/critics/added-critics-result.json(sha256ed444353fae463da52f5f461fb5975eb5d607260d6c219407e77567255b69622).evidence/artifacts/new-wsl-local-models-20261002/PREREGISTRATION-local-models.md: published copy sha256454f879d03d61fa5a8e9f0d877e62cd94878abcd32f8f09b16ac9514492c3a0b; a 16-entry hash chain whose final text isd161feaacd6b3082b3adf37e80a368b58bd6f68205efd399102b8d84d0d0f78b.manifests/evidence.jsonis main's copy, with this branch's files re-registered byscripts/host_receipts.pyregister_fileand its two receipts re-added. During the rebase, main's copy was kept by a merge driver given on the command line, using the custom-driver mechanism in https://git-scm.com/docs/gitattributes ("Defining a custom merge driver").Evidence-class table
swift-iq3s-s2o-64k) made 200 of 200 first-pass valid tool calls on 100 BFCL simple_python cases through the chat and responses surfaces, against 175 of 200 for the controlgpt-oss:20b. The difference is 0.125, with a 95% paired bootstrap interval of 0.065 to 0.19. Both arms ran entirely on the GPU with the embedder residentlocal_integration(native model runs on one workstation in a throwaway distribution, one run per case and arm; not the destination)evidence/artifacts/new-wsl-local-models-20261002/part-a-receipt.json(sha25655ac6ff05f4437dffb2623117c71fb8d6f786d71e63a44b46973b145b1d29577, kindnative_model_e2e)qwen3-embedding:0.6b(qwen3-embedding-8k) scored nDCG@10 0.49827 against 0.41394 for EmbeddingGemma on CQADupstackUnixRetrieval (difference 0.084331, interval 0.067738 to 0.100656). Nemotron-3-Embed-1B scored 0.5014, but its interval of -0.012194 to 0.018141 includes zero and it needs a second model server, so it does not take the slotlocal_integration(same scope)evidence/artifacts/new-wsl-local-models-20261002/part-b-receipt.json(sha2566f5dfbf83c5116e38b42d38ddd5fdbb1a27cf23b5ef7e8ac12d4071c9173a541, kindnative_model_e2e)local_integration(re-derivation from the records on 2026-10-03; nothing was re-measured)evidence/artifacts/new-wsl-local-models-20261002/README.md; outputs inpart-a/bootstrap.txtandpart-b/bootstrap.txtmeasurement, the decision countinstalledgoes from 56 to 58 anddefinitivestays at 31. The record's tables and the handbook are currentlocal_integrationassemble_manifest.py --check,render_tables.py --checkandbuild_new_wsl_handbook.py --check, each exit 0;controls.pyexit 0 (41 mutations killed, 53 tests OK)FROM, a per-modelnum_ctx, and a 404 from/api/embedfor a missing model, with no pullsource_reviewevidence/artifacts/new-wsl-install-plan-20261002/SOURCES.md, "The two local-model rows", read atcc4069396f3ad2c370c53eed2e4a42ac13adab84local-model-serverrow'safter_sign_incheck pass on stand-ins and fail on each planted defectsynthetic(stubollamaandcurlprograms, a scratchHOMEand model store)tests/test_new_wsl_definitive_defaults.py, classLocalModelAcceptance, within the 142-test run (exit 0)local_integration(static)check_plan.pyexit 0;bash -noninstall.shand onaccept.sh, exit 0 eachFROMline, the plan's Swift Modelfile equals the measured onelocal_integration(recorded by the branch on 2026-10-03; not rerun for this PR)reconstruct_s2o_modelfile.py --checkreturned 0 (SOURCES.md)manifests/evidence.jsonis main's copy plus this branch's 63 file entries (44 new, 19 re-hashed) and 2 receiptslocal_integrationscripts/validate.pyexit 0: 9464 hashed files, 189 receiptslocal_integrationscan_private.py, the coordination lane's scanner (outside the repository): its 13 matches in three files are the same, file by file, onorigin/mainand at the pre-rebase headNot established (decision record, "What is not established"):
Local commands run
From the branch's worktree at head
b9d0bf629883e7719c76e0655aa59c4adcbd3e92, 2026-10-03 (checks from 08:28:15Z):Decision record
docs/decisions/2026-10-03-new-wsl-local-models.md. It names the deciding measurement, the alternatives (gpt-oss:20b, the Bonsai and PrismML arms, EmbeddingGemma, Nemotron-3-Embed-1B, WeMM-Embedding-2B) and the comparisons that would overturn each slot.Host evidence
Not applicable: nothing under
evidence/hosts/changes.Checklist
permissions: contents: read(or a narrower, explicitly justified addition). (none changed)Merge slot: after session 0c's #630, #637, #626 and #628, in the order agreed on #608.
🤖 Generated with Claude Code