/delegate & /validate Versions
Status: v2 is live today across Router REST/MCP/x402 surfaces. v3 and v4 are roadmap stages that progressively add explicit consensus and programmable trust.
1. What /delegate and /validate Are For
/delegate and /validate Are ForTwo core primitives:
/delegate– Execution primitive“Send this job to the network.”
Offloads a bounded workflow to Cortensor (models, tools, routing, retries, redundancy).
Returns not just final text, but structured evidence of how work was done (steps, tools, configs in v2+).
/validate– Safety / trust primitive“Before I act or pay, check this for me.”
Re-runs or cross-checks a claimed result under a policy.
v2+ returns a standard verdict contract:
VALID/INVALID/RETRY/NEEDS_SPECplus reasons, evidence, and (later) consensus metadata.
Together they give agents:
A cloud-like execution fabric to offload heavy work (
/delegate).A neutral trust oracle to confirm results before irreversible actions (
/validate).
2. Current Surfaces (v1 + v2)
Today, the Router exposes:
v1 (prompt-oriented)
POST /api/v1/delegatePOST /api/v1/validateFields like
task_request_input,task_result_output,task_domain,task_type,evaluation_mode.Good for simple prompt → answer / prompt → judgment flows.
v2 (contract-first, policy-aware) – current default
POST /api/v2/delegate– payloads shaped around:objective,input,execution,policy, optionaltools/workflow/metadata.
POST /api/v2/validate– payloads shaped around:claim,policy,context, optionalactor/metadata.
Enforces clear separation between:
What is being asked (
objective/claim),How it should run (
execution),Under what guarantees (
policy/ tier / timeouts / basic redundancy).
All four combinations (/delegate v1/v2, /validate v1/v2) are available across:
REST, MCP, A2A / agent cards, and x402 (pay-per-call).
3. Router Evolution – v0 → v4
This section describes how /delegate and /validate evolve across router versions.
The same primitives are reused by Corgent (infra-facing, ERC-8004-friendly) and Bardiel (Virtual-first service agent) surfaces.
v0 – Bare LLM API (Pre-Router Era)
Single-shot
/completion-style calls to a single model or session.No explicit distinction between:
execution vs validation,
policy vs payload.
Any redundancy or checking is hand-rolled per app:
custom retries,
ad-hoc “LLM judge” prompts,
no shared contract or common verdict format.
Works for demos and one-off tools, but not enough for an agent economy.
v1 – First /delegate & /validate (Prompt-Oriented)
/delegate & /validate (Prompt-Oriented)Introduces the two primitives as separate endpoints:
/delegate→ “do this for me”/validate→ “check this for me”
Still prompt-centric:
fields like
task_request_input,task_result_output, with light hints (task_domain,task_type,evaluation_mode).
Runs against a single
session_id:any miner redundancy is handled inside that session’s configuration.
Useful for:
early Corgent / Bardiel prototypes,
simple “send → get answer” and “send → get judgment” flows.
v1 remains supported as the simpler legacy path.
v2 – Contract-First, Policy-Aware Router (Today)
This is the current default for new integrations.
Key changes vs v1:
Structured payloads instead of free-form prompts:
/delegatev2:objective– what the agent wants done.input– data to work on (string, array, structured).execution– mode, model, max steps, tool usage, timeouts.policy– tier (fast,balanced,safe,oracle), latency/timeout bounds, basic redundancy, safety rules.
/validatev2:claim– what is being asserted (summary, transaction, tool output, plan step…).policy– tier, redundancy level, rules (e.g. non-hallucination, consistency).context– task/app metadata that matters for the verdict.
Verdict contract for
/validate:VALID,INVALID,RETRY,NEEDS_SPECplus reasons and evidence, so agents know how to react.
Policy & tiers become first-class:
tier⇒ high-level tradeoffs:fast(low latency, minimal redundancy),balanced,safe,oracle(higher redundancy / stricter checks).
Timeouts and basic redundancy (especially in factcheck-style flows) move from ad-hoc to structured policy.
Router view is still mostly single-session:
Clients hit a single
session_id.That session may internally fan out to multiple miners / models, but consensus aggregation is implicit, not an explicit router-level knob.
Unified surfaces:
v2 is available consistently across:
REST, MCP, agent cards (A2A), and x402 (pay-before-run).
Mental model: v2 gives you execution and verification primitives with structured contracts and policy tiers, but consensus still mostly happens inside a session.
v3 – Explicit Redundancy & Consensus (1 / 3 / 5 Sessions)
Goal: turn redundancy + consensus into explicit, configurable knobs at the router layer across /delegate, /validate, and factcheck.
3.1 Redundant Sessions as Policy
Requests gain a consensus block, for example (shape, not final schema):
replicas:1/3/5session_pool: list of candidatesession_ids (e.g.[201, 202, 203])aggregation:"majority"(later"weighted","median", etc.)disagreement_policy:"return_all"/"fail_hard"/"best_effort"
Router behavior:
replicas = 1→ behaves like v2 today.replicas = 3or5:router fans the same logical request to 3 or 5 independent sessions, each with its own miner sample and infra slice.
each session runs independently (Cortensor principle: redundant, independent runs).
router aggregates the results and returns:
a final answer / verdict, and
per-replica evidence + agreement metrics.
Applies to:
/delegate– redundant execution runs./validate– redundant verification runs.factcheck– redundant world/fact checks (generalizing today’s 1/3 pattern into a standard “1/3/5 replicas” contract).
3.2 Consensus Metadata in Responses
Every v3 router response includes a consensus block, e.g.:
replicas: how many runs were executed.agreement: fractional agreement (e.g.0.67for 2/3).verdicts: list of per-replica verdicts or high-level statuses.strategy: aggregation strategy used ("majority"initially).confidence: derived confidence score.
Examples:
/delegate: “2 of 3 runs agree on this structured result.”/validateorfactcheck: “3 of 5 runs agreed the claim is TRUE under this policy.”
Agents don’t just get “an answer” – they see how many independent runs agreed, and with what confidence.
3.3 Uniform Semantics Across Surfaces
The same consensus semantics apply across:
REST, MCP, A2A, and x402.
Agents can tune reliability vs cost using:
tier(fast vs balanced vs safe vs oracle),replicas(1 / 3 / 5),aggregationanddisagreement_policy.
v3 tagline: “Agent-ready consensus” – explicit redundancy knobs + structured consensus metadata for
/delegate,/validate, andfactcheck.
v4 – Programmable Trust & Reusable Artifacts
If v3 turns consensus into a knob, v4 turns it into a programmable trust fabric and a pool of reusable artifacts, not just one-off responses.
4.1 Pluggable Aggregation Strategies
Beyond simple majority:
aggregationchoices like:"majority","weighted","median_of_means","model_ensemble".
Optional hints such as:
weight_by:"reputation","stake","latency","cost".
Router still hides miner-level complexity, but agents decide how consensus is formed, not just how many replicas run.
4.2 Reusable Validation / Fact Artifacts
High-value /validate or factcheck calls can emit stable validation artifacts, e.g.:
id: unique artifact identifier (e.g.val_0xabc123...).kind:"validation"/"factcheck".scope: logical scope (e.g.trade:1234).ttl: suggested time-to-live.
Other agents or flows (ERC-8004, A2A, x402, on-chain actions) can then:
reference this artifact instead of re-running full consensus,
express policies like:
“Only execute this action if
val_0xabc123...isVALIDwith confidence ≥ 0.9.”
The agentic web becomes:
A graph of delegated calls + shared validations, not just isolated API calls.
4.3 Cost-Aware, Dynamic Redundancy
Redundancy becomes dynamic, driven by policy + observed risk:
Low-risk paths:
auto-downgrade to
replicas = 1once enough similar calls have consistent validations.
High-risk paths (payments, on-chain writes, critical financial or safety actions):
auto-upgrade to
replicas = 5with stricter aggregation.optionally require multi-model redundancy (different backbone families must agree).
All controlled by high-level policy, not per-call plumbing.
4.4 Deep Integration with Cortensor’s Core
v4 explicitly ties consensus and artifacts into:
Proof of Inference (PoI) and Proof of Useful Work (PoUW).
Miner reputation / slashing signals.
Optional on-chain anchoring of critical validations via Corgent in ERC-8004 markets.
v4 tagline: “Programmable, reusable trust” – pluggable consensus strategies, reusable validation artifacts, and dynamic redundancy driven by cost, risk, and PoI/PoUW-powered reputation.
4. Snapshot Summary
Use this block for high-level roadmaps:
v0 – Bare LLM API Single-shot completions; no shared contracts; redundancy and validation hand-rolled per app.
v1 – First
/delegate&/validatePrompt-oriented, single-session endpoints; early agent flows; simple “do this / check this” semantics.v2 – Today: Contract-first, policy-aware router Structured payloads for
/delegateand/validate, explicit verdict contract, tiers, and basic redundancy inside a session; unified across REST, MCP, A2A, x402.v3 – Next: Explicit consensus-aware router 1/3/5 replicas as policy; standardized
consensusblocks in requests/responses; explicit redundancy knobs and consensus metadata across/delegate,/validate, andfactcheck.v4 – Later: Programmable trust fabric Pluggable aggregation strategies, reusable validation artifacts, and dynamic redundancy tied to PoI/PoUW, reputation, and risk – making
/delegate+/validatethe core “intelligent delegation” layer for agents.
Last updated