WARP

Plain search finds what's close to the query. Warp reshapes the query vector at retrieval time to steer search toward, away from, or between concepts — without modifying stored data.

Three operating regimes (empirically validated, not tuning parameters):

Regime	What It Does	Example
Routing	Gentle steering. Preserves query intent.	"Find memory research, weighted toward neuroscience."
Saturation	Cross-domain bridging. Finds structural parallels between distant regions.	"What's the equivalent of natural selection in economics?"
Bounded (default)	Both target and suppression work without interference. Safe general-purpose mode.	Any warp query where you want protection against drift.

Key validated properties:

• Cross-domain recovery: An embedding model that gets 6% on cross-domain relational triplets at zero warp recovers to 94% under calibrated saturation warp. The encoder's blind spots become queryable.
• Suppression alone is useful: Pushing the query away from already-seen content delivers 47% retrieval modulation. Anti-repetition and anti-hallucination at zero cost.
• Seeds-on-demand: Warp targets don't require pre-committed seed infrastructure. Any encodable text direction works. "Things from yesterday," "things I've decided," "concepts related to fluid dynamics" — all valid warp targets without pre-planning.
• Multi-target composition is mathematically exact at the vector level (measured cosine = 1.000000 vs. predicted superposition). Consumers can reason about multi-seed warps additively.
• Deterministic: Same query + same warp spec + same store = same result. Compliance and audit friendly.

Warped neighborhoods expose the analogy signal. Warp reshapes the geometric population while lineage stays unchanged. Large geometric-only population under saturation warp = "the warp found connections the store's history didn't anchor." That's a candidate cross-domain analogy region, surfaced structurally.

Note, the named regimes were removed, as each encoding model has its intrinsic shape, this led to confusion, it may be reintroduced with a later version as the calibrate function provides a measured bound for the WARP but for now it is 0-1.5f.