The architecture behind Context Lattice.

1. Context Supply Chain

The Context Lattice supply chain has nine stages, from raw source material to memory writeback. Every stage carries a correlation ID; every transition is auditable.

2. Six-Layer Agent Framework

The agent framework is the independent variable. Benchmark rankings shift significantly from framework-only changes — making framework design a primary engineering lever, not a model-selection afterthought. The six layers compose in strict order: Safety wraps Orchestration wraps Execution; Context Assembly and Affective Sensing run in parallel; Memory sits beneath all of them.

3. Three Memory Tiers

Production knowledge systems require tiered memory. A single vector store returns semantically proximate but factually incorrect answers for ownership and relationship queries. Context Lattice routes every retrieval through P1 → P2 → P3 in strict precedence.

4. Agent Work Packet Schema

ContextCurator emits agent work packets. Each packet is the structured artifact that wraps a piece of human intent in everything an agent needs to act safely: source context, constraints, acceptance criteria, and readiness score.

// agent_work_packet.json
{
  "packet_id": "awp_2c91",
  "issue_id": "PROJ-482",
  "intent": {
    "goal": "Add rate limit to /api/contact endpoint",
    "actor": "engineer:michael",
    "motivation": "Prevent abuse of strategy session form"
  },
  "constraints": [
    "must not modify auth middleware",
    "backwards compatible with existing clients",
    "latency budget: < 5ms p99"
  ],
  "source_context": [
    "atom_8f3a", // rate limiter design decision
    "atom_a02d"  // API style guide
  ],
  "acceptance_criteria": [
    "returns 429 with Retry-After header",
    "unit tests cover burst + sustained",
    "OTel span emitted on throttle event"
  ],
  "readiness_score": 0.82,
  "readiness_gate": 0.75,
  "agent_authorized": true,
  "trace_id": "7c4f...a912"
}

5. Context Atom Schema

Context atoms are the unit of governed knowledge. Every atom carries provenance, maturity, permissions, and source links. Atoms compose into context packs.

// context_atom.json
{
  "atom_id": "atom_8f3a",
  "title": "Rate limiting standard: token bucket",
  "body": "Internal APIs use token bucket with...",
  "maturity": "verified", // draft | curated | verified | promoted
  "source": {
    "doc_id": "arch-spec.md",
    "section": "3.4 Rate Limiting",
    "hash": "sha256:c4f8..."
  },
  "owner_role": "platform-eng",
  "permissions": ["role:engineering", "role:platform-eng"],
  "links": {
    "supports": ["atom_a02d"],
    "supersedes": ["atom_4421"]
  },
  "confidence": 0.94,
  "last_verified": "2026-04-22T18:11:00Z",
  "reuse_count": 17
}

6. OTel Observability

Every retrieval, every agent session, every committed action carries a trace ID. The v_issue_full_trace view joins Plane issue IDs, Claude session IDs, git commit trailers, and tool call spans into a single chain.

The result: from any deployed line of code, you can trace back through commit → trailer → Claude session → agent work packet → ContextCurator card → source context atoms → originating documents and decisions.

-- v_issue_full_trace view
SELECT
  issue.id              AS plane_issue_id,
  session.id            AS claude_session_id,
  commit.sha            AS commit_sha,
  packet.id             AS work_packet_id,
  atoms.ids             AS source_atoms
FROM plane_issues issue
JOIN agent_sessions session ON session.issue_id = issue.id
JOIN git_commits commit     ON commit.trailer_session = session.id
JOIN work_packets packet    ON packet.session_id = session.id
JOIN packet_atoms atoms     ON atoms.packet_id = packet.id;

7. Safety Architecture

Unconditional constraints belong in framework code, not prompts. Safety guarantees that live in instructions can be reasoned around by the model. Context Lattice enforces action gates, scope isolators, and intervention ladders at the middleware layer.

8. Integration Surface

Context Lattice does not replace your AI platforms. It supplies governed context to them through four interfaces:

# Retrieval API
POST /v1/context/retrieve
  body: { role, query, scope, max_tokens }
  → returns: { atoms[], pack_metadata, trace_id }

# Context pack export
GET  /v1/packs/{pack_id}/export?format=mdx
  → returns: assembled context pack in MDX, JSON, or Markdown

# MCP server
mcp://contextlattice.local/v1
  tools: context.retrieve, atom.lookup, packet.next

# Memory writeback
POST /v1/memory/candidate
  body: { session_id, outcome, candidate_atom }
  → returns: { candidate_id, review_status }