Skip to content

Horizontal Slice — Brick + Factors (Core What-If)

Walk the pavement horizontally: from one raised brick (seed), compare features across linked cells — without re-scanning the whole context window like a vertical LLM.

Model

Concept INDB
Pavement cell Event (tokens + location + time)
Raised brick seed_id or resolved query
Horizontal factors Tokens extracted from what_if (any abstract phrase)
Slice Rescored resonance edges + optional Prism with observer overlay

Flow

1. ResonanceIndex → candidate event IDs (token/location overlap with seed)
2. Echo → score candidates only (not full O(N) memory when index is warm)
3. extract_factors(what_if) → ["meet", "girl", "dreams", ...]
4. apply_horizontal_shift → base_score, shifted_score, matched_factors per edge
5. Prism.synthesize(observer overlay from factors) → relational meaning under hypothetical

No LLM in this path. Hermes optional only for what-if?mode=llm|both.

API

POST /api/v2/slice

Dedicated horizontal slice (always available).

{
  "seed_id": "evt-…",
  "what_if": "today I meet the girl of my dreams",
  "radius": 0.3,
  "limit": 20,
  "observer_context": { "token_weights": { "love": 1.5 } }
}

Or resolve brick from text:

{
  "query": "birthday party",
  "what_if": "there is a fight"
}

Response (data):

  • factors — extracted modifier tokens
  • edges[] — base_score, shifted_score, delta, matched_factors
  • prism — Prism result under overlay (if cloud non-empty)
  • logic — human-readable trace of weight changes

POST /api/v2/what-if

Base query + modifier.

mode Behavior
core (default) base_result + slice (kernel only)
llm base_result + prediction (Hermes/Ollama)
both slice + prediction

Performance

  • Before: Echo scanned all events O(N).
  • After: ResonanceIndex narrows candidates (cap RESONANCE_INDEX_MAX_CANDIDATES), then score O(candidates).
  • Index maintained on ingest and fusion update; rebuilt on startup from memory.

Constants

See core/constants.py: RESONANCE_INDEX_MAX_CANDIDATES, SLICE_FACTOR_MIN_LEN, SLICE_SHIFT_MAX, WHAT_IF_MODE_*.

Observer profiles (config, not hardcoded)

Copy data/observer_profiles.example.json → data/observer_profiles.json or set OBSERVER_PROFILES_PATH.

POST /api/v2/slice
{
  "seed_id": "…",
  "what_if": "there is a fight",
  "observer_profile": "tactical"
}

List presets: GET /api/v2/slice/profiles

Explicit observer_context overrides observer_profile.

What-If seed resolution

POST /api/v2/what-if with mode=core returns seed_resolution (source: echo.center, search.top_hit, interpret.activation, etc.) before slice.

Index sync

ResonanceIndex rebuilds automatically after memory prune or TTL compaction (MemoryManager.on_compaction).

Persisted resonance graph

Echo materializes horizontal edges to {DATA_DIR}/resonance_edges.bin (encrypted, same format family as indb.bin).

  • Load on startup with ResonanceGraph.load()
  • Record after each Echo wave (record_wave)
  • Lookup before scan: graph ∪ index candidates (not full memory)
  • Prune when events leave RAM (compaction callback)

Env: RESONANCE_GRAPH_ENABLED (default true).

Inspect edges: GET /api/v2/slice/neighbors/{seed_id}

Slice responses include graph stats (resonance_graph, causal_index).

Semantic neighbors (optional ML)

When sentence-transformers is installed and RESONANCE_SEMANTIC_ENABLED=true, Echo adds embedding-similar events to the candidate pool (semantic in scan mode).

Env: RESONANCE_SEMANTIC_THRESHOLD, RESONANCE_SEMANTIC_TOP_K, RESONANCE_SEMANTIC_POOL_CAP.

CausalIndex links events in the same location within CAUSAL_TIME_WINDOW_SEC (default 24h). Persisted to causal_edges.bin.

GET /api/v2/slice/causal/{seed_id}?direction=both&limit=20

Echo scan mode may include causal alongside graph, index, semantic.

Subwave (unified broadcast)

One anchor, one frequency — all horizontal neighbors scored for phase lock:

POST /api/v2/subwave/broadcast

See SUBWAVE_RESONANCE.md.

POST /api/v2/slice runs subwave → slice by default (use_subwave=true). Response includes all slice fields plus top-level subwave.

Relation to LLM

LLM remains useful for natural language answers. Core slice owns structure: which links strengthen or weaken under a hypothetical. Use mode=both when you need both the graph shift and a prose summary.