Model currency 2026-09-27: per-lane decision record, model-fallback guards, advisor template - #434

Summary

Records the 2026-09-27 model-currency check the user asked for ("only the latest SOTA models, even the latest ones
released in the past weeks") and closes three routes by which work could run on an older model or lose its advisor.

• docs/decisions/2026-09-27-model-currency.md: one decision table per lane and role. Every routed model is the
  newest in its family on 2026-09-27, so no row switches. Four rows hold a newer model behind an upstream gate:
  jina-reranker-v3.5 for qmd, Nemotron-3-Embed-1B for ai-memory, Claude Sonnet 5.5 and Haiku 5.5 (announced, no model
  ID), and Opus 5.5 on OmniRoute (the gateway lists no claude-opus-5-5, so Opus-role stages stay on native Claude
  Code).
• Fallback guards: .claude/settings.json and examples/claude-native/ultracode.settings.json, plus the recipe's
  embedded copy, now carry "switchModelsOnFlag": false and env.CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK = "1", the
  template's two guards. Before this, a session that loaded only this repository's files, or an adopter who ran
  claude --settings with the portable file, could re-run a flagged Opus 5.5 or Fable request on Opus 4.8 or Opus 5.
• Advisor: adoption/templates/claude.settings.template.json sets "advisorModel": "fable", so a new host that
  applies the template gets the Fable 5.1 advisor the user's rule assigns.
• Catalog: the isolated-builder label in lean-workflow-child-routing changes from "Sonnet/max" to "Opus/max",
  matching all three agent definitions.
• llama.cpp: the pin stays at b11057. v0.5.0 is b11146 (commit 7fe450e1), but the paired acceptance of the
  2026-09-25 re-pin gate has not run. The record notes the divergence, the review status of the 2026-09-26 receipts
  and that acceptance as the one gate still open.
• Workstation refresh record: one dated line closes its bin/vllm follow-up on this host, from the coordinator's
  host observation.

Scope

• Base commit: bfdc99cb (origin/main at the last replay, after #381 PR-A: token-adoption measurement kernel (M3/M4/M5, RTK eligibility, MCP attempt states) #432). The branch was cut from
  ec8a48927c5d2536a57884f7f163c2d13770e7e2, the commit that the failing-first runs below name.
• Lane: lane:foundation. Every changed path is foundation-owned (docs/lanes.md). The only shared hot file,
  manifests/evidence.json, changes by hash registration alone, which "does not by itself make a PR shared".
• Owned paths touched: .claude/settings.json, adoption/templates/claude.settings.template.json,
  catalogs/foundation/decisions.json, docs/decisions/2026-09-25-workstation-sota-refresh.md (one dated line),
  docs/decisions/2026-09-27-model-currency.md (new), examples/claude-native/ultracode.settings.json,
  recipes/claude-native-ultracode.md, tests/test_foundation_catalog.py, tests/test_install_claude_profile.py,
  and manifests/evidence.json (hot-file commit only).

Changes

Current shas, after the replay onto bfdc99cb, oldest first:

1. f21012c2 Model currency record: the method (three Opus research units, one adversarial refuter each and one
   synthesis), the decision table, the holds and their gates, the fallback guards, the advisor, the isolated-builder
   label, the llama.cpp divergence, evidence classes, alternatives, overturn conditions, unresolved items and
   limitations.
2. 94d5c607 Fallback guards in the committed project and portable Ultracode settings, and in the recipe's embedded
   copy and prose. New CommittedSettingsFallbackGuardTests: both files carry the template's value and form, and the
   recipe embeds the portable file without $schema.
3. 0f219fc5 "advisorModel": "fable" in the Claude profile template. New
   test_the_advisor_is_fable_and_accepted_for_the_main_model: the value is fable, the main model is Opus (the
   pairing table accepts Fable for it), and the template sets none of the variables that turn off feature-flag fetching
   or the advisor.
4. cda311e5 Foundation catalog label isolated-builder (Opus/max). New CheckedInAgentLabelTests: every
   "agent (Model/effort)" label in the foundation decisions must match that agent's frontmatter. The test reads the
   frontmatter with a regex, so it does not need PyYAML.
5. c5af2cc8 Hot-file commit: scripts/host_receipts.register_file for the eight changed files only (19 insertions
   and 14 deletions: two fields for each of seven updated entries and five lines for the new one). component_matrix.py --write and new_host_grand_list.py --write produced no content change. manifests/stack.json is unchanged.
6. 28a4b0b9 and 302fe71d Review repair (one round). The record states the review status of the 2026-09-26 llama.cpp receipts:
   independent agree for the current install and use receipts, needs_changes for the superseded first install
   receipt, and a stale "only self reviews" sentence in the serving-switch receipt. It keeps the paired acceptance as
   the one re-pin gate. It adds deniedModels as alternative 4, names it in the guards' Overturn, and closes the
   bin/vllm follow-up, in the record and with one dated line in the workstation refresh record. 302fe71d narrows
   the record's bin/vllm evidence row to "no receipt in the repository". Neither commit touches
   manifests/evidence.json; the two edited records need re-registration.
7. 152548f6 GPT-6 review repair (one round; see Independent review). The embedding row now says qmd 2.8.3
   formats Qwen3-Embedding as well as EmbeddingGemma and loads other models through QMD_EMBED_MODEL, and it
   cites the Gemma releases page for the 2025-09-04 date. The query-expansion row lists
   tobil/qmd-query-expansion-0.6B-gguf (2026-04-24) as a candidate for evaluation, not a switch. The llama.cpp row
   limits its claim to li26's arms and proxy workload. The main-session row gives the coordinator's xhigh.
   17d3d8fc re-registers the record in the hot-file commit, which stays last.

Evidence

No native run backs the guard or advisor changes. No request was flagged on purpose, because deliberately tripping a
safety classifier is not an acceptable test, and no session was started with the changed files.

The table uses this template's four classes. The committed record names the same evidence in its own terms: its
structural_validation rows are the synthetic rows here, its "independent observation (platform record)" rows are
the source_review (platform record) rows, its "native operation (read-only)" gateway call is the first
local_integration row, and its "the coordinator's host observation (no receipt in the repository)" row is the
bin/vllm row. Both describe the same claims.

Local commands run (unit outputs are kept in the unit's notes). The four new tests are
tests.test_install_claude_profile.CommittedSettingsFallbackGuardTests,
tests.test_install_claude_profile.ProfileTemplateSettingsTests.test_the_advisor_is_fable_and_accepted_for_the_main_model
and tests.test_foundation_catalog.CheckedInAgentLabelTests.

$ python3 -m unittest -v <the four new tests>   # ec8a4892 with the two test files from 75d4edaa
exit 1; Ran 4 tests; FAILED (failures=4)
  CommittedSettingsFallbackGuardTests (path='.claude/settings.json'): None != '1'
  CommittedSettingsFallbackGuardTests (path='examples/claude-native/ultracode.settings.json'): None != '1'
  test_the_advisor_is_fable_and_accepted_for_the_main_model: None != 'fable'
  CheckedInAgentLabelTests (lean-workflow-child-routing, isolated-builder): ('sonnet', 'max') != ('opus', 'max')
$ python3 -m unittest -v tests.test_install_claude_profile.CommittedSettingsFallbackGuardTests
  # the same, plus both settings files from 75d4edaa, recipe still at ec8a4892
exit 1; Ran 2 tests; FAILED (failures=1)   # test_the_recipe_embeds_the_portable_settings_file
$ python3 -m unittest -v <the four new tests>   # at 75d4edaa
exit 0; Ran 4 tests; OK
$ TMPDIR=/var/tmp/claude-models python3 -m unittest -q tests.test_install_claude_profile tests.test_foundation_catalog   # at 75d4edaa
exit 0; Ran 80 tests; OK (skipped=3)
$ node test-envelope.mjs            # examples/claude-native/workflows, at ec8a4892 and at 75d4edaa
exit 0 both times; SUMMARY passed=254 failed=0 total=254
$ node test-contract-mutations.mjs  # at ec8a4892 and at 75d4edaa
exit 0 both times; SUMMARY passed=74 failed=0 total=74
$ python3 scripts/component_matrix.py --write
exit 0; {"flip_rule_violations": 0, "rows": 32, "status": "written"}; no content change
$ python3 scripts/new_host_grand_list.py --write
exit 0; {"status": "written", "layers": 32, "winners": 66}; no content change
$ python3 scripts/validate.py
exit 0; {"components": 69, "hashed_files": 7347, "profiles": 4, "receipts": 159, "status": "passed"}
$ TMPDIR=/var/tmp/claude-models python3 -m unittest -q    # full suite at 75d4edaa
exit 1; Ran 6670 tests in 683.040s; FAILED (failures=1, errors=1, skipped=765)
  ERROR tests.test_order_throughput.CapacityRunTests.test_cli_refuses_live_base_url_from_env_file_without_network:
        SafetyError: credential_file_permissions:worktree
  FAIL  tests.test_gpt6_family_tiering_20260926.RunnerTests.test_sigterm_records_the_running_call_as_interrupted_without_counting_it:
        run_arm: the temporary directory /tmp must lie outside the state directory and every repository
$ TMPDIR=/var/tmp/claude-models python3 -m unittest -q <those two tests>   # at ec8a4892, then at 75d4edaa
exit 1 both times; Ran 2 tests; FAILED (failures=1, errors=1), with identical messages
$ python3 scripts/validate.py                          # at 75d4edaa, after the commit
exit 0; {"components": 69, "hashed_files": 7347, "profiles": 4, "receipts": 159, "status": "passed"}

Both full-suite failures come from the worktree's location under /tmp: the unit's worktree sits there. Neither
touches a changed path, and both fail the same way at the base commit, so they are not regressions.

Repair round, at b7dfe498 (the same results at 56426d69), with TMPDIR=/var/tmp/claude-models-repair:

$ git diff --check          # working tree, 75d4edaa..b7dfe498 and ec8a4892..b7dfe498
exit 0 all three times
$ python3 -m unittest tests.test_foundation_catalog tests.test_install_claude_profile
exit 0; Ran 80 tests; OK (skipped=3)
$ python3 scripts/validate.py   # before re-registration
exit 1; "SHA-256 mismatch" and "byte count mismatch" for exactly
  docs/decisions/2026-09-25-workstation-sota-refresh.md and docs/decisions/2026-09-27-model-currency.md

Replay onto e82e6be7, at a127cc2c (hot-file protocol, docs/lanes.md):

$ replay (rebase, then re-register the branch's own files in the last commit)
registered 9 (own set A 8, changed-and-hashed-on-main 8); dropped foreign set-A entries: 0
content lines (ex evidence.json) IDENTICAL; scripts/validate.py FAILS=0
$ TMPDIR=/var/tmp/claude-models-repair uv run --no-project --with pyyaml==6.0.1 python -m unittest tests.test_foundation_catalog tests.test_install_claude_profile
exit 0; Ran 80 tests; OK   # no skips: the three skips above are the PyYAML-dependent frontmatter tests
$ git diff --check origin/main HEAD
exit 0

After the GPT-6 repair, at d5a6a7b2 (still on e82e6be7; the diff from the reviewed head a127cc2c is the
record and its manifests/evidence.json entry):

$ python3 scripts/validate.py
{"components": 69, "hashed_files": 7349, "profiles": 4, "receipts": 159, "status": "passed"}
$ git diff --stat a127cc2c d5a6a7b2
 docs/decisions/2026-09-27-model-currency.md | 8 ++++----
 manifests/evidence.json                     | 4 ++--

Replayed again onto 55fc8d17 (#433 merged) as 89c6d9b3 (d5a6a7b2 and 80694cf4 before the replay): content
lines identical, 9 files re-registered, scripts/validate.py FAILS=0. Replayed onto bfdc99cb (#432 merged) as
17d3d8fc: content lines identical, 9 files re-registered, FAILS=0.

Those four lines report the two edited records' stale registrations in manifests/evidence.json; the hot-file commit
re-registers both.

SOTA sources

Every change follows these primary sources (fetched or re-checked 2026-09-27):

• Fallback guards: Claude Code settings reference,
  switchModelsOnFlag ("Scope: Any file";
  false pauses an interactive session and ends a -p request as an error) and the env entry ("Scope: Any file";
  project values apply after workspace trust or at -p startup; the ignore list does not name either key).
  Model configuration, automatic model fallback
  gives the fallback targets: Opus 5 for biology, Opus 4.8 for cybersecurity. Settings precedence.
  The template form follows adoption/templates/claude.settings.template.json and
  docs/decisions/2026-09-25-model-fallback-guard.md at ec8a4892.
• The undocumented variable: absent from env-vars, model-config, the
  settings reference and anthropics/claude-code
  CHANGELOG.md at 7779afb12e3635f46f56ec823979d68350ae000b
  (tag v2.1.283). Its behaviour comes from a source review of the installed npm @anthropic-ai/[email protected]
  client (jD, orn, IAr, Rhn, Qzn).
• deniedModels (alternative 4, not adopted): the same CHANGELOG.md at 7779afb1, lines 6-7 under ## 2.1.283
  (availableModelsMatch, and a "deniedModels managed setting to block specific models, even when availableModels
  allows them"); the settings reference,
  deniedModels (the summary row and the entry
  give "Scope: Managed", and the key is ignored "in user, project, and local settings and in --settings, with a
  warning"; "claude-opus-5" also blocks Opus 5.5 and "claude-opus-5-0" blocks Opus 5 alone);
  model-config, "Block specific models or versions"
  lines 350, 353 and 364, and
  "Automatic model fallback" line 517 (line
  numbers in the page's Markdown source, fetched 2026-09-27).
• Advisor: the settings reference, advisorModel
  ("Scope: Any file"); Advisor, "Choose an advisor model" (an Opus 5.5 main
  model accepts Fable, and without access or consent requests go without the advisor); env-vars, "Features that need
  feature-flag fetching".
• Catalog label: the model: and effort: frontmatter of isolated-builder in .claude/agents/,
  adoption/agents/claude/ and examples/claude-native/agents/ at ec8a4892, and
  docs/decisions/2026-09-27-claude-harness-settings.md item 1.
• Model currency: Claude release notes,
  models overview,
  model deprecations,
  Claude Opus 5.5 announcement,
  npm registry @anthropic-ai/claude-code (latest 2.1.283),
  OpenAI API changelog,
  openai/codex rust-v0.157.1 codex-rs/models-manager/models.json#L4,
  tobi/qmd v2.8.3 src/llm.ts#L86-L119,
  vLLM releases (v0.30.0),
  akitaonrails/ai-memory releases (v2.4.1) and
  #859,
  ggml-org/llama.cpp#26286, and the model cards
  nvidia/Nemotron-3-Embed-1B-BF16 at c0c9fea9,
  jinaai/jina-reranker-v3.5-GGUF at 884f7c67
  and Qwen/Qwen3.8-27B at 1d4bf0f2, with
  embeddings-benchmark/results at e0c9e75e.
• llama.cpp pin: ggml-org/llama.cpp tags b11057 (59657a613ab0fa4ab327d6c790123dff30bfbd67), b11146
  (7fe450e19305b828c199d602c23a8337aaa1f03b) and v0.5.0 (annotated c13fcbf684171d5e0bca3fc5c34be6a99174b05f,
  which points to 7fe450e1), read through the GitHub API. The re-pin gate is the paired acceptance in
  docs/decisions/2026-09-25-workstation-sota-refresh.md, "llama.cpp b11146 (v0.5.0)". The review status comes from
  the reviews and limitations fields of the three 2026-09-26 host receipts and of
  evidence/receipts/local-inference-c2-serving-switch-20260926.json, and the li26 runtime binding from
  blueprints/convergence-practice/local-inference-latest-20260926/eval_arm.py lines 114-116.
• The bin/vllm update: the coordinator's host observation of 2026-09-27 (repointed at 19:21:18Z).

Decision record

docs/decisions/2026-09-27-model-currency.md: evidence classes, eight alternatives, overturn conditions and unresolved
items.

Independent review

One independent evidence review returned repair. Its medium finding: the record said the 2026-09-26 llama.cpp
receipts carry only self reviews, but three carry independent_session reviews. Its low finding: the 2.1.283
deniedModels setting was missing from the alternatives. The single repair round the bounded review loop allows
(56426d69 and b7dfe498) fixed both, re-weighed the llama.cpp decision on the corrected review status (unchanged: the paired
acceptance remains the gate) and closed the bin/vllm follow-up from the coordinator's host observation.

A cross-family review followed. GPT-6 Astra at max ran through the packaged landscape-sweep codex_call.sh
lane, read-only, with live web search, at a127cc2c. Its seven checks included recomputing all ten files'
hashes and sizes, the Claude and GPT-6 release dates and IDs, the six-task RTEB(Code) means at e0c9e75e, the
2.1.283 guard behaviour, and read-only mutation checks of the two guard assertions. It returned repair with five
findings. The coordinator verified each against the primary source before editing:

Per the bounded review loop, there is no second review round. GPT-6's residuals: it ran no test suite, live
session, model load or GPU acceptance. The full settings-reference page exceeded its web tool. Some pinned
Hugging Face cards were intermittently unreachable. It did not observe the gateway inventory or the vllm symlink.

Residuals

• CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK is undocumented. Its effect on subagents and workflow children rests on a
  source review of the 2.1.283 client and was not probed, so a client update can drop it silently. The recipe says to
  re-check it after each update, but no automated watch exists yet. The documented deniedModels (2.1.283) works only
  from managed settings, which this repository does not write; adopting managed settings would make it the preferred
  guard (record, alternative 4 and Overturn).
• Behaviour change: in this repository a flagged request now ends in a refusal, or in a -p error, in the main thread
  and in every child, instead of an answer from Opus 4.8 or Opus 5. Coordinators check the child's transcript for a
  refusal before retrying. A contributor who wants the automatic switch opts out in .claude/settings.local.json.
• Workflow children are documented to inherit advisorModel, but that was not observed, and the resolved advisor ID
  has not been seen in a transcript.
• Template precedence: apply_claude_settings.py lets the template win for scalars, so re-applying the template resets
  a host that chose opus or no advisor. The record explains how such a host keeps its choice.
• Deferred: a workflow-contract check in test-envelope.mjs would change the portable contract, its SHA256SUMS and
  every adopter's vendored copy. The Python tests cover this repository's files.
• The llama.cpp re-pin to b11146 (v0.5.0) still needs the paired acceptance of the 2026-09-25 gate against a fresh
  b11057 baseline in an owned window; that is the remaining gate. The 2026-09-26 host receipts already carry
  independent reviews (agree for the current install and use receipts, needs_changes for the superseded first
  install receipt). Their own limitation sentences, and limitations[4] of
  local-inference-c2-serving-switch-20260926, still say they carry only self reviews; that text is stale, and this
  PR edits no receipt.
• Holds until their gates open: jina-reranker-v3.5 (stock llama.cpp rank-pooling; #26286 is open); Nemotron for
  ai-memory (2.5.0 with docs: refresh foundation layer pins and evidence dates #859); Sonnet 5.5 and Haiku 5.5 (no model ID yet). When those ship, three version-keyed places
  need an edit: the template's modelSettings, LEGACY_EXACT in adoption/hooks/claude/effort-default-guard.py and
  ALIAS_RESOLUTION in examples/claude-native/workflows/child-usage.mjs. Opus 5.5 on OmniRoute is held until the
  gateway lists claude-opus-5-5.
• Other items the record leaves unresolved: which stages use GPT-6 Sol, the OmniRoute provider path and whether it
  clamps max, unconfirmed release dates, workstation boundaries, and alias resolution on other hosts such as the
  Mac.
• The two known environment failures of worktrees under /tmp are not chased (see Evidence). A checkout outside
  /tmp should run the full suite once before merge.
• The branch is replayed onto bfdc99cb. The only shared path is manifests/evidence.json. After the replay,
  re-register the nine changed files (the eight above and docs/decisions/2026-09-25-workstation-sota-refresh.md,
  which main already registers), re-run both --write generators and scripts/validate.py.

Checklist

• No GitHub Actions or workflows changed.
• No secrets printed, logged or committed. No credential value was read. The committed record has no private
  paths or account data.
• No new paid hosting or subscription. The template does not change any host until it is applied. Billing note:
  once applied, each advisor call spends Fable tokens under the host's existing billing: API rates, plan limits,
  or usage credits on plans where Fable bills that way. If the account still needs the usage-credits consent, the
  advisor bills nothing until that consent is given
  (advisor, Cost). This host already runs with
  advisorModel: "fable".
• Peer-owned files and worktrees were preserved.

Recommended label: lane:foundation

🤖 Generated with Claude Code