Reading Hyperlex evidence

Researcher desk Static Pages · not live ~/.hyperlex/ Similarity ≠ Brier Settle before scoring

What each number means, what you may claim, and what you must not. Live operator state is never on this site — only method, samples, and sanitized history.

Keep these four rules in view

1. brier: null is correct until an operator settles forecasts.
2. Cosine / vector score is not a probability of adoption or virality.
3. Phase 5 is SPECULATIVE — scenario literature, not measurement.
4. Atoms, not bags — sigma · rizz · locked in are separate units.

---

Evidence stack

Read top → bottom. Higher layers do not “upgrade” lower ones into Brier.

#	Layer	What it is	Provenance	Brier?
1	Ingest	Route, source, query text	OBSERVED / mock	No
2	Lineage	Family + matched atomic terms	INFERRED	No
3	Vector neighbors	Embedding-near terms/receipts	INFERRED (hash default)	Never
4	Virality / typology	Analytic scores on receipt	INFERRED / SPECULATIVE	No
5	Forecasts	Open probabilities	INFERRED until settle	After settle only
6	Phase 5	Transmission / multi-agent / risk	SPECULATIVE	Always null
7	Brier series	Score from forecast↔outcome pairs	OBSERVED after settle	Yes

query (atom)
   │
   ▼
ingest ──► analyze ──► lineage + vector_neighbors + virality
   │              │
   │              └── forecasts (open) ──► settle ──► Brier series
   │
   └── Phase 5 (optional · SPECULATIVE · never Brier)

Atomic terms: free text like sigma rizz locked in is input, not one lexicon item. The engine expands to separate atoms. On archive cards, prefer listed atoms over a blended “seed.”

---

Provenance labels

Label	Means
OBSERVED	Measurement or network capture (or remote embed API)
INFERRED	Derived offline (hash embeddings, lineage match, hybrid re-rank)
SPECULATIVE	Scenario / model exploration (Phase 5)
brier: null	Correct default — numeric Brier only after settlement

Hash model id: hyperlex.hash_ngram_v1.d256 → always treat as INFERRED.

---

Vector evidence on Pages vs local

Where	What researchers get
This page + sample JSON	How to interpret scores; publish-safe table
Telemetry	Desk map + how-to-read
Slang lineage map	Interactive constellation · deep links ?term= / ?family=
Run history	Sanitized analysis / Phase 5 snapshots
Vector backend	Operator map (sqlite · chroma · promote)
Live local / Cloud	Not mirrored on Pages

Pages does not stream ~/.hyperlex/chroma or Chroma Cloud. To publish vector context: export a sanitized sample, or archive a receipt that already includes vector_neighbors.

Worked sample (local Chroma · publish-safe)

Specimen · vector search schema hyperlex.vector_search.v1 · brier null · INFERRED

query rizz sigma aura · backend chroma (local) · model hyperlex.hash_ngram_v1.d256
note: cosine over unit vectors — not a calibrated probability / Brier

Rank	Term	Family	Score	Source
1	rizz	brainrot-aura	0.59	lineage registry
2	rizz	brainrot-aura	0.59	YTD backfill
3	aura	brainrot-aura	0.50	lineage registry
4	aura	brainrot-aura	0.50	YTD backfill
5	aura	brainrot-aura	0.50	ingest auto-index

Machine JSON: samples/vector-search-rizz-sigma-aura.json

{
  "schema": "hyperlex.vector_search.v1",
  "query": "rizz sigma aura",
  "backend": "chroma",
  "model": "hyperlex.hash_ngram_v1.d256",
  "embed_provenance": "INFERRED",
  "n_hits": 5,
  "brier": null
}

Say this

Under offline hash embeddings, “aura” is a near neighbor of the multi-term query in the curated brainrot-aura family.

Do not say this

There is a 50% probability “aura” will go viral.
(confuses cosine similarity with Brier / forecasting)

Duplicate rows (registry + backfill + autoindex) mean multiple seed sources for the same surface form — not triple independent evidence.

---

Archive cards

On Run history:

You see	Means
analysis	Receipt-backed analyze / pipeline snapshot
phase5_scenario	Research sim only — SPECULATIVE
atoms listed separately	Preferred presentation
risk tier	Advisory only — not market advice, not Brier
Families	Lineage labels from receipt summaries
Brier null / open	Settlement not done (or not exported)

---

Claims matrix

Claim	Supported by	Not supported by
Term belongs to a lineage family (with confidence)	Lineage on receipt	Vector score alone
Terms are embedding-near in Hyperlex’s index	Vector neighbors / search	Brier, Phase 5 risk
Mutation candidates (next forms)	mutation_prediction (SPECULATIVE)	Brier, virality as certainty, market advice
Operator forecast skill over settled events	score-series after settle	Open pipeline results
Counterfactual transmission / cascade	Phase 5 (SPECULATIVE)	Vector DB, Brier
Live web prevalence	Live ingest + external methods	Pages static archive alone

---

Pipeline ↔ index (method)

pipeline / ingest / run
   → analyze (vector_neighbors if index warm)
   → receipt
   → vector_index (fail-open · local sqlite or chroma)
   → forecasts · optional Phase 5

bulk:   vector-seed (registry + YTD packs + receipts)
promote: vector-sync → Cloud  (explicit · not auto on ingest)

---

Reproduce the sample

export HERMES_SKILL_DIR="${HERMES_SKILL_DIR:-$HOME/.hermes/skills/hyperlex}"
HLX="python3 $HERMES_SKILL_DIR/scripts/hyperlex.py"
export HYPERLEX_OFFLINE=1

$HLX vector-seed --backend chroma --db ~/.hyperlex/chroma \
  --through 2026-08 --include-home --include-golden
$HLX vector-search "rizz sigma aura" --backend chroma --db ~/.hyperlex/chroma \
  --kind term --top-k 5

Related

Topic	Page
Atomic multi-term	atomic-terms.md
Brier design	brier-calibration.md
Lineages	slang-lineages.md
Phase 5	phase5.md
Vector backends	modules/vectordb.md