Examples

Every example here is traceable to a real run on the repo that ships it — commands you can copy, outputs with real numbers. When the tooling changed, the example gets re-run (a stale example is worse than none: agents copy them verbatim).

Example	What it shows
systemone-rerank	the local decision-model tier reordering retrieval
freshness-cycle	scheduled upstream drift → evidence → team corpus
local-model-routing	privacy tiers: ollama tags vs on-device HF vs cloud
context-receipt-loop	receipts ↔ outcomes: delivery you can audit

System One fusion rerank

The problem: lexical scoring gives two candidates the same overlap score when they collide on words; only one actually answers the question. The System One tier (a local decision model — ollama serves the same wire protocol the Jev cloud uses, no auth, no egress) judges which top-k candidates genuinely answer.

Measured live on this repo's fabric (10,419 pages, tev1:latest keep-warm on localhost):

$ wf query "why do concurrent readers see torn state" --verbose
Query type: concept
Pages loaded: 10419
Scored pages: 618
  0.576 [claim   ] claim-alpaca-agents-...-workflow-analysis-md-003
  ...
  systemone rerank: {...}          # ← P(page answers) per top-10 candidate
  hop-gate reranked 3 edges        # ← graph hits ordered by P(edge serves query)
After graph expansion: 621

The separations are multi-order on diverse candidates (same batch, one state, one call):

candidate	P(answers)
direct answer (single-writer, non-atomic rewrites)	0.970
vocabulary page mentioning "torn write"	0.665
scoped-adjacent (ontology: atomicity is global)	0.329
word-collision page (Ollama endpoints)	0.026

The set-overlap scorer gives the two word-collision pages identical scores; the rerank tier is what kills the false positive. Contracts:

• silent fall-back — judgment disabled, ollama down, or >2s ⇒ None ⇒ lexical order stands (test: tests/test_systemone.py::TestDegradeContracts)
• never above lexical evidence — the fusion weight is bounded (RETRIEVAL["fuse_lex"]/["fuse_sem"]), and graph hits (hop-gated) can never exceed graph_seed_score (0.1)
• explicit opt-out — wf query ... --no-rerank

Also on the same tier: wf verify-effects gains a pairwise relation head (supports / contradicts / supersedes + probabilities + confidence) — proposed contradictions flow to the review queue; humans decide, contested states are the representation.

The freshness cycle

Hooks capture local commit drift on active machines — but a team whose every machine is dormant for a week can carry silently stale claims. The freshness cycle makes upstream freshness independent of anyone's habits.

$ wf freshness --dry-run                    # all connected projects
=== alpaca-agents ===
  git-history: clean                        # sha256-gated: nothing new upstream
=== aperiodic ===
  git-history: captured                     # 60 new PR/issue records detected
  reverify:    clean (0 overdue+stale)

For real (writes evidence + the .last-capture state marker):

$ wf freshness                              # per connected repo
$ wf gate                                   # what needs a human

Scheduled on the corpus CI (scaffolded automatically by wf sync init / wf sync setup — daily, opt-in via repo variable WIKI_FABRIC_FRESHNESS=1):

freshness.yml (corpus CI)
  → capture-git --since-state per repos: entry   (0 tokens, sha256-gated)
  → review --auto-reverify                        (mechanical, 0 tokens)
  → commit + push refreshed evidence
  → teammates: wf sync pull

Evidence-plane only: the cycle never rewrites claims; stale/contested states are the representation; re-extraction stays opt-in per routing tier.

Local-model routing

Privacy tiering is per-stage, per-repo — the fabric carries the decision, each machine may override. Three local tiers + one cloud tier:

# fabric.yaml — the decided defaults
llm:
  base_url: http://localhost:11434/v1
  local_model: gemma4:e4b-fixed            # ollama-served (tier 1: server-local)
  compiler_model: deepseek-v4.1-flash:cloud  # tier 4: cloud (compiler policy)

repos:
  auth-service:        # sensitive repo → everything on-device
    extract: local
    synthesize: local
  docs-repo:           # public docs → cloud is fine
    extract: cloud

Tier	Model shape	Runs where	Egress
1. ollama-served	bare tags (gemma4:e4b-fixed, tev1:latest)	the local ollama server	none
2. on-device HF	mlx-community/* (Apple), *GGUF	in-process (mlx-lm / llama-cpp)	none (model download once, human-gated)
3. decision models	tev1/nimble tags	local ollama — judging only (rerank, review triage, mining gates)	none
4. cloud	ids (deepseek, jev-latest)	hosted APIs	yes — never on the local tier

Two traps the tooling now catches for you:

• :cloud tags are NOT local — ollama's hosted farm egresses. looks_like_local_model refuses them as local_model, and lint flags a fabric.yaml that sets one: LLM-CONFIG llm.local_model: ... routes over the network (hosted farm).
• Gemma4 tags can't be judges — ollama's System One endpoint serves only the Tev/Nimble family; wf models ensure --check reports OK: gemma4:e4b-fixed (ollama-served) for extraction, and the judge falls back to tev1:latest with a notice.

Verification loop (0 LLM tokens until extraction):

wf models ensure --check          # presence probe (ollama list / HF cache)
wf ingest docs/auth.md --extract-claims   # routes via repos tier (local ⇒ gemma4)

Context receipts: the auditable loop

Delivery must be reviewable against what was known — receipts make the manifest a durable record:

$ wf context --task "Fix the torn-read window in catalog rebuild" --write-receipt
receipt: <fabric>/corpus/registry/receipts/receipt-6b965337825f.json (receipt-6b965337825f)
## Context Manifest
... (selected/excluded, every item with a reason) ...

$ # ... solve the problem, then log the outcome against that receipt:
$ wf log --project catalog --receipt receipt-6b965337825f \
    --problem "concurrent readers see torn state" \
    --intervention "single-writer mutex + atomic rewrite" \
    --outcomes "torn reads gone; rebuild latency unchanged"

Receipt ↔ outcome linkage is the evidence plane mining consumes: patterns promote only with ≥2 independent projects, dossiers need human review, and wf utility shows which patterns actual sessions used. The full contract lives in schemas/frontmatter.md.