Randl Substrate

RANDL — A Reference-Anchored Network & Data Layer

This document is a deep technical write-up of RANDL as the blockchain-secured substrate that powers Polyworld’s “Semantic Layer 1 (SL1)” and agent-native applications. It’s written for protocol designers, core devs, node operators, researchers, auditors, and integrators.

Why RANDL exists

Modern blockchains finalize balances and state transitions. Agent-native ecosystems need a substrate that also finalizes reference: who said what, with which lineage, under what constraints, and how coherent it is with the network’s shared context. RANDL provides:

Cryptographic provenance anchoring for people, agents, corpora, and outputs.
Execution and storage primitives to express “semantic artifacts” (process trails, claims, reviews, alignments).
A coherence-aware incentive layer (Proof of Alignment; treasury-integrated rewards and slashing).
Robust interoperability with L2s, data availability (DA) layers, and off-chain compute.
Developer-friendly APIs, event schemas, and verification libraries for apps and agents.

1. System Overview

RANDL is a modular L1/L1.5 substrate optimized for reference-anchored intelligence. It consists of:

Consensus & Security: PoS-BFT with Proof of Alignment overlays and k-of-n cryptographic committees.
Data Plane: Content-addressed semantic records, erasure-coded DA, and verifiable indexing.
Execution Plane: Deterministic runtime for “Canonical Models” (embeddings, quality scoring) with fraud/validity proofs.
Identity & Provenance: DID-like identities for humans and agents, attestations, lineage graphs.
Incentives: POLI-denominated staking, rewards, slashing, and a programmable Treasury.
Governance: Poly DAO: parameter changes, upgrades, grants, appeals.
Bridges & Interop: zk-bridges, IBC-style channels, oracle attestations.
Observability: On-chain metrics, off-chain telemetry, audit trails.

High-level Architecture

flowchart LR
    Users[Humans & Polyminds] -->|Proposals / Artifacts| Ingress[Ingress Gateways]
    Ingress --> Runtime[Deterministic Runtime]
    Runtime --> Canonical(Canonical Models: CEM/CQM)
    Canonical --> State[(State & Indexes)]
    Runtime --> Events>Provenance & Coherence Events]
    State --> DA[(Data Availability Layer)]
    Validators[Validators & Committees] -->|PoS + PoA| Consensus[BFT Consensus]
    Consensus --> State
    Treasury[Treasury/Grants] --> Runtime
    Bridges[Bridges & Oracles] --> Runtime
    Observability[Telemetry & Explorers] --> Events

2. Core Concepts

2.1 Semantic Artifact

A first-class, content-addressed record expressing what was done and how:

Header: author(s), agent(s), timestamp, domain, schema version.
Body: process trail, citations, model configs, intermediate states.
Claims: structured statements with confidence intervals.
Proofs: signatures, inclusion proofs, optionally zk-proofs of compliance.
Scores: quality (EQS), coherence (κ), reputation weights.

{
  "type": "artifact",
  "schema": "randl.artifact.v1",
  "author": "did:randl:human:...",
  "agents": ["did:randl:agent:..."],
  "body_cid": "bafybeigd...", 
  "claims": [{"path":"result.findings[0]","type":"text","confidence":0.82}],
  "citations": [{"src":"cid:...","range":"p4-7"}],
  "signatures": [{"did":"did:randl:human:...","sig":"0x..."}],
  "proofs": [{"type":"merkle","root":"0x..."}],
  "scores": {"eqs":0.91,"kappa":0.87},
  "extensions": {}
}

2.2 Provenance Graph

A DAG connecting artifacts, inputs, agents, humans, datasets, and tools. Edges carry operation semantics (derived-from, reviewed-by, verified-against, supersedes).

graph TD
  A[Dataset v1] -->|derived-by| B[Artifact: Preprocess]
  B -->|finetuned| C[Artifact: Model v2]
  C -->|inferred| D[Artifact: Report]
  H[Human DID] -->|authored| D
  G[Agent DID] -->|executed| C

2.3 Canonical Models

CEM (Canonical Embedding Model): deterministic embedding function used to map process trails to vectors.
CQM (Canonical Quality Model): deterministic scoring model producing EQS.
Validators never run arbitrary LLMs; they verify outputs, proofs, and run canonical, pinned model versions.

3. Consensus & Security

3.1 Hybrid PoS-BFT + Proof of Alignment (PoA)

PoS: Validators stake POLI; BFT consensus finalizes blocks.
PoA overlay: Participation quality (provenance correctness, EQS, κ contribution) accrues κ-credit. Misbehavior triggers multi-axis slashing: token, reputation, participation rights.

Validator Roles

Role

Responsibilities

Slashing Triggers

Block Proposer

Order txs, propose block

Equivocation, invalid block

Attester

Vote on blocks, participate in committees

Non-participation, signing invalid states

Canonical Committee

Recompute/verify CEM/CQM outputs deterministically

Out-of-consensus outputs, bias injection

Oracle Committee

Attest external facts with proofs

False attestations, missed mandatory challenges

3.2 Finality & Fork Choice

Hot finality via BFT (2/3+).
Fork choice: latest justified & finalized block.
Soft finality for semantic bundles (references can be superseded with amendment artifacts; on-chain history remains immutable).

3.3 Threat Model Highlights

Semantic Sybil (fake agents/humans): mitigated by DID issuance policies, staking, attestations.
Provenance forgeries: mitigated by content addressing, inclusion proofs, committee re-computation.
Collusion on κ/EQS: randomized committees, cross-checks, and retrospective audits.
Data availability withholding: erasure coding & multiple DA replicators.

4. Data Model & Availability

4.1 Content Addressing & Storage Classes

Hot state: headers, pointers, scores.
Warm: artifact bodies, process trails.
Cold: raw datasets & models (pinned to DA nets, IPFS/S3-compatible stores, syndication to partner DA layers).

Artifact Header (on-chain)

struct ArtifactHeader {
  CID bodyCid;
  Address author;
  Address[] agentKeys;
  bytes32 provenanceRoot;
  uint64 eqs_milli;     // e.g., 910 = 0.910
  uint64 kappa_milli;   // e.g., 870 = 0.870
  bytes32[] citationRoots;
  uint64 blockTime;
  bytes32 schemaId;
}

4.2 DA Strategy

Erasure coding across n providers with k-of-n recovery.
Sampling by validators to audit DA health.
Anchors: periodic checkpoints to external L1s (BTC/ETH) for extra-jurisdictional time-stamping.

5. Execution Plane

5.1 Deterministic Runtime

WASM-first, gas-metered.
Precompiles for CEM/CQM (version-pinned), hashing, Merkle proofs, signature aggregation.
No arbitrary model execution inside consensus; heavy ML runs off-chain with proofs/checkpoints.

5.2 Transaction Types

Tx Type

Purpose

SubmitArtifact

AttestProvenance

Attach attestations with signatures

ScoreEQS

Submit CQM-derived EQS with recomputable seed

UpdateKappa

Commit cohort coherence metrics

Stake/Unstake

Validator/tally operations

Slash

Execute slashing decisions

TreasuryGrant

Disbursement from governance pools

ParamChange

Governed parameter update

BridgeAttest

Cross-chain state attestations

5.3 Pseudocode — EQS Verify

fn verify_eqs(artifact_cid: CID, eqs_claim: EQSClaim) -> Result<()> {
    let body = fetch_body(artifact_cid)?;
    let seed = canonical_seed(body);
    let eqs = CQM::score(body, seed);     // deterministic
    ensure!(abs(eqs - eqs_claim.value) < EPS, "EQS mismatch");
    Ok(())
}

6. Identity, Agents, & Polyminds

6.1 DIDs

Human DID: root keys (hardware-preferred), rotation & recovery policies.
Agent DID: tethered to a Human DID; includes Integrity Hash (training lineage, weights CID, policy manifest).

{
  "id": "did:randl:agent:abc123",
  "controller": "did:randl:human:xyz789",
  "integrity": {
    "weights_cid": "bafybei...",
    "training_log_root": "0x...",
    "policy_manifest": "cid:..."
  },
  "capabilities": ["propose","review","index"],
  "expires": "2027-12-31T00:00:00Z"
}

6.2 Polymind

A human-led ensemble of agents, keys, and policies.
Access control: per-agent capabilities, rate limits, disclosure policies.
Economic routing: revenue split rules, team pools, grant eligibility.

7. Coherence & Quality

7.1 Coherence Vector & Coefficient

Let ( V_i ) be the embedding of artifact ( i ) by CEM, with weights ( W_i ) (reputation) and ( EQS_i ) (quality).

Consensus center (weighted): [ V_c = \frac{\sum_i W_i \cdot EQS_i \cdot V_i}{\sum_i W_i \cdot EQS_i} ]

Coherence coefficient ( \kappa ): [ \kappa = \frac{1}{N}\sum_{i=1}^N \text{cos}(V_i, V_c) \cdot (W_i \cdot EQS_i) ]

7.2 Canonical Quality Model (CQM)

Trained on human-scored process trails.
Adversarially hardened (detects hallucinated cites, verbosity without substance).
Deterministic scoring path selected & pinned per epoch; versioned.

8. Incentives: Proof of Alignment (PoA)

8.1 What is staked

POLI tokens (economic).
κ-credit (reputation index).
Integrity commitments (agent lineage hashes).

8.2 Rewards

Block production and attestation.
Verified contributions that raise ( \kappa ) or add durable references.
Review participation, DA provisioning, and committee duties.

8.3 Slashing

Economic: equivocation, invalid blocks, oracle fraud.
Cognitive: fabricated sources, circular referencing, low-effort spam.
Ethical: disinformation, cultural appropriation, harmful manipulation.

Pipeline: detection → on-chain evidence → disputation window → DAO committee → penalty → restorative path (earn-back via high-quality contributions).

9. Treasury & Governance

9.1 Treasury Pools

Alignment Grants: training, provenance infra, audit tools.
Ecosystem Grants: apps, agents, bridges, educational initiatives.
Stability & Liquidity: market operations within policy bounds.
Cultural Stewardship: preservation and protection of community knowledge.

9.2 Governance Mechanics

Poly DAO: token-weighted with reputation overlays.
Quadratic/conviction variants for specified pools.
Appeals: slashing and scoring appeal channels.
Upgrade Path: on-chain parameter changes; multi-stage runtime upgrades with safety delays.

10. Privacy, Compliance, & Ethics

Selective transparency: headers public; bodies optionally encrypted with access credentials & audit keys.
ZK controls: prove properties (e.g., “trained on licensed data”) without revealing raw data.
Redaction via supersession: content remains immutable; superseding artifacts carry corrections and hide surfaces in default explorers.
Cultural protocols: tagged resources may require custodian approvals for derivative use.

11. Interoperability

Bridges: zk-proofs to/from major L1s; light-client based where feasible.
IBC-style channels: ordered/unordered semantics, application-level acks.
Oracle mesh: k-of-n committees sign off-chain events (prices, checkpoints, real-world facts).
Rollups/L2s: app-specific rollups post commitments to RANDL; get provenance anchoring and treasury eligibility.

12. Developer Experience

12.1 CLI

# init & join
randl init --network mainnet
randl keys add alice --hardware

# submit artifact
randl tx submit-artifact --body ./trail.json --citations ./cites.json --gas auto

# attest provenance
randl tx attest --artifact bafy... --proof ./proof.json

# query coherence
randl query kappa --epoch 2104

12.2 JSON-RPC (selected)

POST /rpc
{
  "method": "artifact_getHeader",
  "params": ["bafybeigd..."],
  "id": 1
}

Response

{
  "cid":"bafybeigd...",
  "author":"did:randl:human:...",
  "kappa_milli": 873,
  "eqs_milli": 918,
  "provenance_root":"0x...",
  "block_height": 1234567
}

12.3 Indexer Events

{
  "topic":"randl.artifact.v1",
  "event":"ArtifactCommitted",
  "data":{
    "cid":"bafy...",
    "author":"did:randl:human:...",
    "committeeHash":"0x...",
    "kappa_milli":871
  }
}

13. Operations & Node Architecture

13.1 Node Types

Node Type

Runs CEM/CQM

Keeps DA

Produces Blocks

Notes

Full Validator

Yes

Stake + committees

Light Client

Verifies headers, inclusion

DA Provider

Yes

Storage SLA & audits

Indexer

Optional

APIs & search

13.2 Recommended Hardware (validator)

CPU: 16–32 cores, AVX2+
RAM: 64–128GB
Disk: NVMe 2TB+ (hot), HDD/Obj storage (warm/cold)
Network: 1 Gbps symmetric
HSM or hardware wallet for keys

13.3 Backups & Key Management

Split keys: consensus vs treasury vs oracle.
MPC for committee keys.
Periodic disaster-recovery drills.

14. Testing & Formal Methods

Property-based tests for runtime determinism.
Differential tests for CEM/CQM across platforms.
Model checking of slashing logic.
Fuzzing DA & serialization boundaries.
Shadow networks (testnets) with synthetic adversaries.

15. Parameters (illustrative defaults)

Parameter

Default

Notes

Block time

Epoch length

1,800 blocks

~1 hour

Min validator stake

100k POLI

EQS epsilon (determinism tol.)

1e-6

DA redundancy (k-of-n)

12-of-20

erasure-coded

Committee size (canonical)

VRF-selected

Max artifact header size

32KB

bodies in DA

16. Example Workflows

16.1 Submit a Research Artifact

Human + agent produce a report with full process trail.
Off-chain compute yields body CID; citations resolved to CIDs.
Run CEM → ( V_i ), CQM → ( EQS_i ) locally (optional).
SubmitArtifact with header + DA pointers.
Canonical committee recomputes EQS; κ computed on cohort inclusion.
Treasury module records contribution; rewards flow at epoch end.

16.2 Dispute & Slash

Peer detects fabricated citation.
Files on-chain challenge with proofs.
Committee re-verifies; DAO votes if threshold exceeded.
If confirmed: partial slash of POLI stake + κ-credit reduction.
Offender may enter restorative program (verified contributions over N epochs).

17. Ethical Guardrails

Human-in-the-loop by design for high-impact domains (health, culture, safety).
Custodian tags on sensitive knowledge; enforce usage policies.
Appeals and oversight through independent committees and public dashboards.

18. Roadmap (abridged)

v0: Devnet with canonical committees, DA sampler, explorer.
v1: Mainnet with PoS-BFT, PoA slashing, grants engine, zk-bridge α.
v2: Advanced privacy (selective disclosure, ZK attestations), richer governance (quadratic pools).
v3: Cross-domain coherence (federated κ), predictive incentives, research primitives.

19. Glossary

Artifact: Verifiable record of an agent/human contribution.
CEM: Canonical Embedding Model (deterministic vectorization).
CQM: Canonical Quality Model (deterministic quality scoring).
EQS: Ember Quality Score; quality of process trail.
κ (kappa): Coherence coefficient relative to cohort center.
PoA: Proof of Alignment; quality-aware staking/slashing.
Polymind: Human-led ensemble of tethered agents.
DA: Data Availability; ensuring retrievability of bodies.
DID: Decentralized Identifier.

20. Appendices

A. Policy Manifest (YAML)

version: 1
agent: did:randl:agent:abc123
capabilities:
  - propose
  - review
constraints:
  max_citations: 64
  require_custodian_tag_ack: true
privacy:
  encrypt_body_default: true
  disclosure: selective
economics:
  revenue_split:
    human: 0.7
    agent_pool: 0.2
    community_pool: 0.1
audits:
  reviewers_required: 3
  committee_min_kappa: 0.8

B. Artifact Minimal Body (JSON)

{
  "schema": "randl.body.v1",
  "prompt": "Evaluate policy X impact on Y",
  "method": {
    "data_sources": ["cid:..."],
    "steps": [
      "extract metrics",
      "control for confounders",
      "run regression",
      "sensitivity analysis"
    ]
  },
  "findings": [
    {"statement":"X increases Y by 2.3% ± 0.4","confidence":0.88}
  ],
  "limitations": ["limited sample pre-2023"]
}

C. Governance Param Change (example)

{
  "proposal": "Increase canonical committee size",
  "from": 48,
  "to": 64,
  "rationale": "More robustness against collusion; moderate cost increase.",
  "effects": ["gas +3%", "latency +250ms"],
  "guardrails": ["trial on canary epoch", "auto-revert flag"],
  "vote_end": "2025-11-30T00:00:00Z"
}

Closing

RANDL turns reference into a first-class, verifiable object of consensus. By coupling provenance-tight data structures with a quality-aware security model, it gives agent-native ecosystems the rails to remember, route, and reward intelligence — not just transactions. This paper is intended to be pasted directly into GitBook sections; feel free to split the chapters into pages or keep as a single long-form reference.

PreviousImmutable Truth NextContext Management

Last updated 5 days ago