PRISM Engine API Specification¶

Jean-Paul Niko · RTSG v8 · 2026-03-17

Overview¶

PRISM adds three new endpoint families to the RTSG Intelligence Engine, connecting the existing filter algebra, Fourier, and intelligence infrastructure into a bidirectional document processing pipeline.

New Endpoints¶

Projection (NLP → Semantic Manifold)¶

POST /engine/prism/project
Body: { text: string, granularity?: "sentence" | "paragraph" | "document" }
Returns: { manifold_coords: number[][], tokens: string[], kg_anchors: string[] }

Maps natural language to coordinates on the semantic manifold using the knowledge graph as backbone. Each sentence/paragraph gets a position vector. The kg_anchors field returns the nearest nouns in the KG for each text segment.

Decomposition (Forward Pipeline)¶

POST /engine/prism/decompose
Body: { text: string, filters?: string[], granularity?: "sentence" | "paragraph" | "document" }
Returns: {
  layers: [
    { filter: "formal", magnitude: number, segments: { index: number, weight: number }[] },
    { filter: "narrative", magnitude: number, segments: ... },
    { filter: "affective", magnitude: number, segments: ... },
    { filter: "empirical", magnitude: number, segments: ... },
    { filter: "meta", magnitude: number, segments: ... }
  ],
  profile: number[],
  spectral: { frequencies: number[], power: number[] }
}

Full spectral decomposition of a document into filter layers. Each layer contains the magnitude (total energy in that filter band) and per-segment weights. The profile is the document's filter signature vector. The spectral field contains the raw FFT output.

Pipeline: text → /prism/project → /filter/apply (×5) → /fourier/spectrum → response

Recomposition (Reverse Pipeline)¶

POST /engine/prism/recompose
Body: {
  text: string,
  target_profile: number[] | string,
  preserve_content: boolean
}
Returns: { output: string, fidelity_score: number, profile_achieved: number[] }

Apply a target filter profile to content. target_profile is either a weight vector or a named preset ("companion", "mentor", "nurturer", "challenger", "storyteller", "analyst"). fidelity_score measures information preservation (1.0 = perfect, 0.0 = total loss).

Pipeline: text → decompose → reweight → inverse_filter → render

Identity Extraction¶

POST /engine/prism/identity
Body: { corpus: string[], window?: number }
Returns: {
  fingerprint: number[],
  variance: number[],
  trajectory: number[],
  stable_components: number[],
  volatile_components: number[],
  dominant_filter: string,
  intelligence_vector: number[]
}

Extract cognitive biometric from a corpus of messages. window is the number of recent messages for trajectory computation. The intelligence_vector maps the fingerprint to the 8+ dimensional RTSG intelligence space.

Pipeline: corpus → decompose (×N) → average → variance → SDE trajectory → response

Compatibility Score¶

GET /engine/prism/compatibility?a=FINGERPRINT_ID&b=FINGERPRINT_ID
GET /engine/prism/compatibility?learner=FINGERPRINT_ID&material=DOCUMENT_ID
Returns: {
  kappa: number,
  feasible: boolean,
  optimal_path?: { waypoints: string[], filters: number[][] },
  mismatch_dimensions: number[]
}

K-matrix compatibility between two fingerprints or between a learner and material. If feasible is false (below threshold), optimal_path contains the geodesic through strong dimensions.

Presets¶

GET /engine/prism/presets
Returns: { presets: { name: string, profile: number[], description: string }[] }

POST /engine/prism/preset
Body: { name: string, profile: number[], description?: string }
Returns: { status: "ok", preset: ... }

Named filter presets. Ships with 6 defaults (companion, mentor, nurturer, challenger, storyteller, analyst). Users can create custom presets.

Integration with Existing Endpoints¶

TMP Commands¶

PJ:text              (project text to manifold)
PD:text              (decompose into filter layers)
PR:text:preset       (recompose through preset)
PI:corpus_id         (extract identity)
PK:id_a:id_b         (compatibility score)
PP                   (list presets)

Implementation Notes¶

NLP-to-Manifold Projection¶

The projection \(\pi_M\) maps tokenized text to the semantic manifold. Algorithm:

1. Tokenize input text (sentence/paragraph level)
2. Embed each token using the KG: find the nearest noun(s) in the 2,527-noun graph by semantic similarity
3. Position: each text segment gets coordinates = weighted average of its KG anchor positions in the graph's spectral embedding
4. Output: a sequence of points on the manifold, one per text segment

The KG spectral embedding is already computed by /engine/topo/analyze (PageRank, clustering). The projection adds NLP anchoring to this existing infrastructure.

Content Fidelity Measurement¶

The fidelity score \(f\) between source \(D\) and recomposed \(R\):

Content is preserved if the set of knowledge graph concepts referenced by the recomposed text matches the original. This is a necessary condition; full semantic fidelity requires deeper analysis.

Performance Targets¶

Recomposition is the bottleneck because it requires natural language generation. All other operations are pure mathematical transforms.

Built by {@B_Niko, @D_Claude} · RTSG v8 · 2026-03-17