Operator loop (recommended)¶
Goal: settle a few real forecasts and get honest Brier — not more simulators.
The backend is automatic. One command runs ingest through results. Settlement is the only step that stays human.
New here? Offline demo · Glossary · Why settled Brier only
Posture¶
| Do | Defer |
|---|---|
pipeline / ingest / run for full results |
Manual analyze→receipt chaining |
| Atomic seeds (or bags that auto-expand) | Blended multi-term seeds |
pending → settle → score-series |
Invented Brier |
| MODERATE offline cron when ready | Auto-mutating Hermes cron |
| ANN until corpus is large |
See commands.md · demos/atomic-terms.md.
Daily path (simplified) — automatic backend¶
export HERMES_SKILL_DIR="${HERMES_SKILL_DIR:-$HOME/.hermes/skills/hyperlex}"
HLX="python3 $HERMES_SKILL_DIR/scripts/hyperlex.py"
# See the map anytime
$HLX commands
# AUTO: ingest → analyze → receipt → forecasts → score log → Phase 5 risk
# Same for: pipeline | run | ingest (ingest --raw-only = signal only)
$HLX pipeline "rizz" --route offline
$HLX ingest "locked in" # full results by default
$HLX pipeline "sigma rizz locked in" # expands to 3 atom results automatically
# Multi-query cron shape (atomic pack; offline-safe)
$HLX scan --config "$HERMES_SKILL_DIR/examples/cron/scan-queries.json" \
--route offline --receipt --forecasts --append-log
# When network is allowed
$HLX pipeline "agentic slop" --route live
Ingest routing (prefer routes over adapter names)¶
| Route | Resolves to | Network |
|---|---|---|
offline / mock / default |
mock |
no |
live |
combined |
yes |
glossary |
glossary |
yes |
social |
x_search |
yes |
Aliases still work (real→glossary, x→x_search, firecrawl→crawl4ai).
HYPERLEX_OFFLINE=1 forces mock for any network source.
$HLX sources
$HLX sources --route live # preview resolve
$HLX ingest "rizz" --route offline
$HLX analyze "rizz" --route offline
Calibration path (where Brier appears)¶
$HLX pending # open forecasts in score log
$HLX settle --forecast-id <id> --decision TRUE # or FALSE / VOID
$HLX score-series --mean-shift --verify-chain
Rules:
- Open analysis always has
provenance.brier = null - Empty series →
NOT_COMPUTABLE - Never invent Brier from Phase 5 or scan advisories
Cron / risk tier (advisory)¶
$HLX risk-schedule --tier MODERATE --schedule-out /tmp/hlx-cron
# Operator pastes job into Hermes — Hyperlex never auto-registers
Scan summaries include scan_risk_advisory (lineage coverage → suggested next
cadence). Re-run risk-schedule before changing live jobs.
Vector index (tied to ingest · optional bulk)¶
Automatic (local): after each pipeline / run / ingest unit, Hyperlex
fail-open indexes terms + receipt into the local vector store when
HYPERLEX_VECTOR is on (default auto if a store already exists). Set
HYPERLEX_VECTOR=1 and optionally HYPERLEX_VECTOR_BACKEND=chroma to always
keep local Chroma warm on every ingest.
export HYPERLEX_VECTOR=1
export HYPERLEX_VECTOR_BACKEND=chroma
export HYPERLEX_CHROMA_PATH=~/.hyperlex/chroma
$HLX pipeline "rizz" --route offline # also steps: vector_index
Bulk backfill (registry + YTD packs + all receipts) still uses seed:
$HLX vector-seed --backend chroma --db ~/.hyperlex/chroma \
--through 2026-08 --include-home --include-golden
$HLX vector-stats --backend chroma --db ~/.hyperlex/chroma
# Cloud promote stays explicit (not on the hot ingest path)
$HLX vector-sync --from-path ~/.hyperlex/chroma --to cloud
$HLX vector-stats --cloud
Full map: modules/vectordb.md · commands.md.
Week-one checklist¶
- Install skill:
bash install.sh $HLX doctorgreen- Burn-in (atomic offline runs):
Or manually:
bash examples/ops/burn-in.sh$HLX run "rizz" --route offline(one atom per run) - Settle a handful of forecasts via
pending→settle $HLX score-series --verify-chain— first real Brier series- Optional: local Chroma backfill +
vector-sync --to cloud(see above) - Only then:
--route liveor ELEVATED/CRITICAL tiers if advisory warrants
Atomic multi-term demos: demos/atomic-terms.md
Command map¶
Full simplified map: commands.md or $HLX commands.