Palantir Foundry

Palantir Foundry Integration

Foundry remains Foundry; OVERWATCH governs calls into it. The VantageAdapter is one adapter — not the core.

Overview

DeepSigma integrates with Palantir Foundry through the DecisionSurface runtime layer. The architecture is deliberately adapter-based: the core claim-event evaluation engine is portable and environment-agnostic, while the VantageAdapter handles Foundry-specific storage and retrieval.

Core Engine (severity, seal_and_hash, memory_graph)
        |
  DecisionSurface Runtime  (ingest / evaluate / seal)
        |
  SurfaceAdapter ABC  (8 abstract methods)
        |
   +-----------+-----------+
   |           |           |
Notebook    CLI       VantageAdapter
(in-memory) (+ JSON)   (Foundry/Vantage)

Key principle: Foundry maintains its own autonomy. OVERWATCH governs the calls into it — not the other way around.

Current Status

The VantageAdapter is an honest stub. All 8 methods raise NotImplementedError with a roadmap reference. This is intentional — it defines the contract that a real Foundry integration must fulfill, backed by 11 contract tests that verify the interface.

Package: src/core/decision_surface/vantage_adapter.py

from core.decision_surface import DecisionSurface

ds = DecisionSurface.from_surface("vantage")
# All operations raise NotImplementedError until Foundry SDK integration

Adapter Interface

The VantageAdapter must implement all 8 SurfaceAdapter methods:

Method	Purpose	Foundry Mapping
ingest_claims(claims)	Store claims	Write to Foundry dataset or ontology objects
ingest_events(events)	Store events	Write Action Type instances
get_claims()	Retrieve claims	Query ontology objects (type=Claim)
get_events()	Retrieve events	Query Action Type instances
get_evidence()	Retrieve evidence	Build from claim/event matches
store_drift_signals(signals)	Persist drift detections	Write to drift tracking dataset
store_patches(patches)	Persist patch recommendations	Write to patch dataset
store_evaluation_result(result)	Persist evaluation summary	Write to evaluation dataset

Data Model Mapping

Object Types

Foundry Object Type	Envelope record_type	Notes
Ontology Object	Entity	Domain entities (customers, assets, cases)
Dataset Row	Claim or Metric	Analytical outputs, derived facts
Action Type instance	Event	Action executions, workflow triggers
Media Reference	Document	Attachments, reports, exported artifacts

Field Mapping

Foundry field	Envelope field	Transform
objectRid	record_id	uuid_from_hash('palantir', objectRid)
objectTypeRid	record_type	Map via object type table
__createdTimestamp	created_at	ISO-8601 from epoch ms
__updatedTimestamp	observed_at	ISO-8601 from epoch ms
__createdBy	source.actor.id	Foundry user RID via directory API
properties.*	content	Flatten property bag; preserve property names
__primaryKey	content.entity_id	Business key
Link types	links[]	Foundry link types map to envelope edge types

Link Types

Foundry link type	Envelope rel	Direction
relatedTo	supports	Forward
derivedFrom	derived_from	Forward
partOf	part_of	Forward
supersededBy	supersedes	Reverse (flip direction)
contradicts	contradicts	Forward

Ingestion Strategy

1. Ontology sync — Foundry Ontology SDK, list objects by type with pagination
2. Incremental — filter by __updatedTimestamp > last_sync_ts for delta loads
3. Transform — apply field mapping, resolve user RIDs, flatten properties
4. Link resolution — query linked objects and generate links[] array
5. Validate — run against canonical_record.schema.json
6. Ingest — write to canonical store

Provenance

Foundry records carry strong provenance through dataset lineage:

{
  "provenance": {
    "chain": [
      {"type": "source", "ref": "foundry://<enrollment>/<objectTypeRid>/<objectRid>"},
      {"type": "evidence", "ref": "<dataset_rid>", "method": "foundry_transform"}
    ]
  }
}

Confidence Scoring

Data Source	Confidence
Source-of-truth datasets	0.95
Derived/computed datasets	0.80
User-submitted data	0.60

Integration Patterns

Three core patterns govern how OVERWATCH interacts with Foundry:

1. Feature fetch wrapper — returns {value, capturedAt, sourceRef} so every read has provenance
2. Action adapter — requires a Safe Action Contract before any Foundry write
3. Verifier — reads authoritative ontology state after write to confirm postconditions

Permissions

Requires a Foundry service user with:

• ontology:read — query ontology objects and link types
• datasets:read — read dataset rows for ingestion

Connector Profile

Property	Value
Transport	HTTPS REST
Auth	OAuth
MCP Tools	-- (planned)
Exhaust	foundry adapter

Tests

• tests/test_vantage_adapter_contract.py — 11 tests verifying all 8 methods raise NotImplementedError and adapter properties (surface_name, isinstance, factory)

Roadmap

Full Foundry SDK integration is deferred to v2.1.1 "Institutional Expansion". The current stub + contract tests define the exact interface a real implementation must satisfy. See DecisionSurface for the adapter authoring guide.

Related Pages

• DecisionSurface — portable runtime and adapter ABC
• Integrations — all connector targets
• Event Contracts — routing table and event declarations
• AuthorityOps — authority enforcement for Foundry actions