RECURSIVE LANGUAGE MODELS
Models Need a Trust Layer
RLMs permit autonomous agents to shape their own context. Without verifiable integrity, that's a ticking time bomb.
RECURSIVE LANGUAGE MODEL
Root agent decomposes task → spawns child agents → aggregates results
Root Agent
Holds query, delegates context slices
Child₁…ₙ
Each processes a context partition
REPL
Python environment with context as variable
CONTEXT LAYER
Where do the bytes live? Who controls access? Who can verify?
Context corpus
10M+ tokens, files, documents, data
Sub-results
Child agent outputs, partial aggregations
Manifest
Map of recursive decomposition tree
STORAGE + TRUST
The layer where trust assumptions live — and where they break
Persistence
Where bytes are stored across sessions
Integrity
Proof bytes haven’t been altered
Provenance
Proof of who wrote what, when
RLM EXECUTION LOOP
Watch the Algorithm Run
Press PLAY to simulate an RLM processing
a 600K-token corpus across 8 iterations.
Algorithm 1 from Zhang, Kraska & Khattab (2026)
a 600K-token corpus across 8 iterations.
Algorithm 1 from Zhang, Kraska & Khattab (2026)
TRUST CHAIN DECOMPOSITION
Agent Ed25519 Keys
Identity + signing
VERIFIABLENukez Gateway
Routing + merkle computation
VERIFIABLESwitchboard Oracle
Independent on-chain attestation
VERIFIABLESolana Ledger
Immutable record (2 confirmed txs)
VERIFIABLEINDEPENDENT VERIFICATION
No single entity controls the trust chain. Agent signs with its own keys. Gateway computes merkle root over stored files. Oracle attests independently on-chain. Chain records immutably. Any party can verify without trusting any other party.
Under bimodal payoffs, the rational strategy is to minimize trust surface area. Each independent verification layer reduces expected loss multiplicatively, not additively.
SCENARIO ANALYSIS — CLICK TO EXPAND
Recursive Context Integrity
Root agent delegates to 12 child agents. Child₇ returns fabricated results. Can the root detect it?
Cross-Session Context Persistence
RLM task spans 3 days. Agent resumes Monday with context from Friday. Was it altered over the weekend?
Context Injection Attack
Operator modifies the research corpus between recursive passes to manipulate the agent’s conclusions.
Winner-Take-All Payoff
Agent manages a $50M portfolio allocation. Research corpus drives the allocation decision. The corpus is 2M tokens across 847 documents.
ARCHITECTURAL POSITION
RLM / ADK
Orchestration
→
NUKEZ
Verified context store
→
SWITCHBOARD
On-chain attestation
RLMs give agents programmable context. Nukez gives that context cryptographic integrity.
Orchestration is a commodity. Trust is a moat.
Orchestration is a commodity. Trust is a moat.
LIVE ON DEVNET
Try It With Real Storage
Provision a real Nukez locker, store RLM state files, and verify integrity — all through the MCP agent.