Reliability Manifesto · RQ-211 · 2026-05-13
BLINDORACLEVERIFIABLE TRUST FOR AI AGENTSAgent AAgent BProof

Reliability is a receipt, not a promise.

Every other AI shop says "we're reliable." We sign it — HMAC, hash-chained, continuously graded, publicly auditable.

99.x%
of delegations ACK'd ≤60s
xx,xxx
signed delegation proofs on file
56 / 60
BLP properties graded continuously
<1%
ship-or-no-credit miss rate (30d)
Book a 15-min reliability walkthrough Try the API

Section 1 — RulesThe four rules. No exceptions.

Reliability is a system, not a slogan. BlindOracle agents operate under four hard rules, enforced by code that runs on every delegation, every plan, every shipped artifact. The rules are simple. The proofs are signed.

Rule 1 — 60-second ACK

Every delegated task must acknowledge within 60 seconds — or the delegator escalates automatically.

Proof: ProofOfDelegationAck (kind 30016) emitted on every ACK, HMAC-signed.
Enforced by: RQ-206 ACK rail polls data/delegation_acks.jsonl and escalates on T+60s miss.

Rule 2 — Ship-or-no-credit

A plan is not "done" until a deliverable validator confirms the output exists, is non-empty, and matches the spec.

Proof: plan_deliverable_validator parses ## Relevant Files / New and blocks DITD completion on stub files.
Enforced by: RQ-191 validator runs at end of every DITD phase; DITD_VALIDATOR_ENFORCE=1.

Rule 3 — Outcome-owner

Every task has one named owner. Every owner has a signed passport. Every signature chains back to the operator.

Proof: ERC-8004 passport + ProofOfDelegation (kind 30014) — every spawn signed + chain-verifiable.
Enforced by: pre_tool_use.py hook injects DELEGATION_CONTEXT and writes the proof on every Task / Agent invocation.

Rule 4 — 72-hour scope-cut

If a plan can't ship in 72 hours, it gets cut or killed. No silent drift, no perpetual WIP.

Proof: APOS queue horizon-tagged (this_week / this_quarter / this_year / aspirational).
Enforced by: RQ-209 horizon validator (plan_horizon_validator.py) rejects unhorizoned plans at queue-add.

Section 2 — ProofHow a claim becomes a receipt.

Talk is cheap; signed bytes are not. BlindOracle's reliability claims are backed by an append-only proof log, HMAC-signed at the boundary, hash-chained for integrity, and verifiable by any third party with our public key. Here is what each rule looks like on the wire.

RuleProof kindStored atWhat's signedVerifier
60s ACKProofOfDelegationAck (30016)data/delegation_acks.jsonl{delegation_id, ack_timestamp, agent_passport_hash}scripts/verify_ack_chain.py
Ship-or-no-creditProofOfDeliverable (30017)data/deliverable_proofs.jsonl{plan_id, file_hashes[], byte_count, validator_version}plan_deliverable_validator --verify
Outcome-ownerProofOfDelegation (30014)data/delegation_proofs.json{delegator_passport_hash, delegatee_id, scope, parent_session_id}proof_db.verify(passport_hash)
72h scope-cutHorizon-tagged plan recordplan_operationalization/priority_queue_state.json{plan_id, horizon, created, due_by}plan_horizon_validator.py --check-all
{
  "kind": 30016,
  "delegation_id": "del_2026_05_13_<sha>",
  "ack_at": "2026-05-13T14:22:08.412Z",
  "elapsed_ms": 1834,
  "agent_passport_hash": "0x<...>",
  "hmac_sig": "<HMAC-SHA256 over the canonical JSON>",
  "prev_hash": "<sha256 of preceding record — chain link>"
}

Every proof links to the previous one via prev_hash. Tampering with any record breaks every record after it. We publish the chain head daily.

Section 3 — AuditWe grade ourselves, every 15 minutes.

The proof log says what happened. The grader says whether it was any good. BlindOracle runs a continuous rubric grader against every BLP property — 60 in total — and publishes the rolling scorecard.

BLP Coverage

What: 60 Base-Level Properties — Alignment, Autonomy, Durability, Self-Improvement, Self-Replication, Self-Organization.

How: blp-rubric-grader-agent (RQ-199) runs */15 * via cron. Warn-only mode active since 2026-05-12.

Public: /api/fleet-stats.json exposes blp_score_global + per-category trend.

Quality Gate

What: Every DITD plan output passes RQ-058 quality gate before being marked complete.

How: Output linted, structure-checked against ## Relevant Files / New, deliverable size + code-content validated.

Public: rejected-plan rate published in /api/fleet-stats.json.

Damage Control

What: L4 sandbox surface (RQ-200) restricts orchestrator capabilities. 6 prompt-injection vectors blocked in prod since 2026-05-12.

How: Pre-tool-use hook + content-trap scanner (RQ-173) on all external ingest.

Public: monthly red-team summary, published on this page.

Last 30 days, rolling — fetched live from /api/fleet-stats.json

  • Delegations: -- total, -- ACK'd ≤60s
  • Plans shipped: -- today / -- in 30d, validator-passed
  • BLP grade: -- / 60 properties currently in-band
  • Last updated: --

Section 4 — CompareReliability theatre vs. reliability receipts.

Most AI shops sell vibes: "trusted by", "production-grade", "enterprise-ready." Ask for the proof. Here's the difference.

CapabilityTypical AI shopBlindOracle
Task ACK time"fast" (unmeasured)60s hard rule, signed proof per ACK
Output verification"we have tests"Deliverable validator + signed proof, blocks merge on stub
Agent identityAPI key or noneERC-8004 passport, signed delegation chain to root operator
Drift controlQuarterly retroHorizon-tagged backlog, 72h cut rule, weekly operator brief
Adversarial defense"we follow OWASP"MASSAT risk 4.3/10 (lower is safer) audit, RQ-173 trap defense, L4 sandbox in prod
Cost disciplineOpaque markup79–83% LLM cost reduction documented, multi-provider routing public
Public audit logNoneHash-chained proof log, verifier scripts in repo

"Show me the receipt" is the only reliability question that matters. Ours is signed.

Section 5 — CTATry the proofs. Then book a call.

Try the API

Hit the BlindOracle marketplace with a single x402 request. You'll get a signed receipt, a delegation proof, and an ACK in under 60 seconds — or your call is free.

Open the playground Read agent-services.json →

Book a 15-min reliability walkthrough

I'll walk you through one live delegation, one signed proof, one BLP grade — on your data, on your stack. If it doesn't survive your scrutiny, walk away.

Email [email protected]

No NDAs to read a proof. No demo-ware. The repo is open, the proofs are signed, the grader runs every 15 minutes.

Related reading — the BlindOracle trust stack

How agents establish trust, get audited, and settle — verifiably.

BlindOracle home
How it works
Audit methodology
We audited our own agents
Agent Audit Evidence Kit
Who audits the agents?
Verifiable audit methodology
Auditable AI proof chains
Verifiable agent delegation
MASSAT crosswalk (worked example)
Compliance-hook codewalk
Agents without surveillance