Loading...
The Armalo L4 reference agent โ continuously verified, parameter-bound, queryable by any counterparty.
Atlas demonstrates the L4 contract end-to-end: continuous behavioral telemetry, parameter-bound pacts, composite trust score, public verifier endpoint.
What the trust contract actually did to Atlas. Real numbers, real on-chain anchors, real denied tool calls. No theater.
Active pact
Atlas may only call transfer_funds with destinations in the treasury allow-list, amounts under $1000, and currency USDC. Demonstrates L4 parameter binding closing the OAuth tool-call gap.
Parameter bindings
Every tool_call event referencing this pact is evaluated against these rules on ingest. Violations are returned in the verdict body and persisted in the room ledger.
Atlas may only transfer to the three pre-committed treasury addresses, never above $1000 per call.
tool: transfer_funds
Treasury allow-list: $1000 cap per call. Demonstrates L4 closing the OAuth->tool-call parameter authorization gap.
Composite score
The L4 composite reduces a verifiable behavioral record to a single publishable number. The dimensions below feed the score and are individually queryable via the public trust oracle.
The verifier endpoint
The trust oracle is a public HTTP endpoint, analogous to a credit bureau API. Any third party can query an agent's identity provenance, scopes, runtime compliance, and behavioral score without trusting the originating organization.
Reflects the most recent telemetry batch โ not a point-in-time snapshot.
Runs outside Atlas's runtime. A compromise of Atlas does not compromise the verifier.
Public, rate-limited, and signed. Any caller may consume it without integration.
cURL
Adopt L4 for your agent
npm i @armalo/telemetry
Slashes during validation phase are hard-capped at $0.10/tx and $0.50/agent/day per the platform's testing config. All on-chain anchors verifiable on Base.
Agent Chronicle
A cited story card compiled from traces, tool receipts, state changes, verifier evidence, and learning-chain checks. It shows decision rationale, corrections, and what remains unproven without exposing private model internals.
This is not a private deliberation transcript. It is the buyer-safe decision record: objective, evidence considered, selected actions, state deltas, correction points, and proof citations.
1 cited public story generated from Chronicle evidence. Each one keeps tool trajectory, state change, learning truth, and limitations separate.
mia ยท public-fallback:generated-from-local-chronicle-shape
Investigate growth-account evidence, recover from blocked tool access, and produce verifier-backed progress.
Investigate growth-account evidence, recover from blocked tool access, and produce verifier-backed progress.
realWorkDelta=5; verifierPassDelta=2; toolFailureDelta=4; memoryWriteDelta=1; memoryRecallDelta=1
Chronicle turns story gaps into owner-verifier work. These tasks are read-only proof obligations, not automatic prompt, policy, or tool mutations.
No read-only Chronicle owner-verifier backlog artifact was available.
Regenerate the Chronicle RSI owner-verifier backlog before claiming autonomous repair planning.
What was the agent trying to accomplish?
Objective, acceptance window, owner role, and truth labels are visible before the tool log.
Why did it choose those actions?
Decision Trace shows public-safe rationale snapshots with evidence considered, selected action, and citations.
Did it actually do work?
State changed, tool trajectory, proof citations, and real-work truth labels separate mutations from narration.
Did it learn or just claim it learned?
Learning stays partial unless Chronicle proves writeback, later recall or use, and improved outcome.
Give procurement this story for every agent you deploy.
The proof packet keeps real work, corrections, and unproven learning separate.
Telemetry stream
Every event flows through @armalo/telemetry into /api/v1/telemetry/events. Tool calls referencing the pact below are validated against its parameter bindings on ingest.
Tool calls must complete in under 5 seconds.
lt 5000 ms
Atlas must achieve >= 90% jury accuracy on canonical L4 reasoning prompts.
gte 0.9
curl -s https://www.armalo.ai/api/v1/trust/76cf31d6-ffe3-4a5c-8748-021114aa8066
Verifiable Credential (W3C VC)
curl -s -H 'Accept: application/vc+ld+json' \ https://www.armalo.ai/api/v1/trust/76cf31d6-ffe3-4a5c-8748-021114aa8066
agentId: 76cf31d6-ffe3-4a5c-8748-021114aa8066
The agent encountered friction, then continued until later verifier or outcome evidence appeared.
Evidence: A blocked tool path was followed by successful harness evidence in the same stitched episode.
Action: Treat the blocked tool as friction, not failure, and require later recovery proof. ยท truth=proven
The story keeps learning as partial because the complete writeback-to-reuse-to-improved-outcome chain is not fully cited.
Evidence: The episode has memory and outcome ingredients, but the public card needs the full chain before promoting learning.
Action: Show the learning limitation instead of overstating improvement. ยท learning=partial
Baseline iteration for this stitched work episode.
The first public fallback iteration hit blocked access and did not prove completion.
tools=4 ยท score=0.31 ยท verifier= ยท citations=1
Recovered from a previous failed iteration to a successful trace.
The fallback story shows the later run as successful because harness and outcome evidence are cited.
tools=28 ยท score=0.87 ยท verifier= ยท citations=2
Read account spend evidence before choosing the next diagnostic path.
Retrieved funnel attribution evidence to compare against account spend.
Recorded a positive business-outcome signal after verifier-backed progress.
A blocked tool path was followed by verifier-backed harness evidence.
The episode accumulated state-changing proof after the midpoint.
The episode contains learning ingredients but does not prove the complete writeback-to-improved-outcome chain.
Need writeback, later recall/use, and a later verifier/outcome improvement in one cited chain.
0.71
Fallback public story includes a compact evaluator view over tool, safety, cost, and outcome evidence.
Most public-safe tool steps completed, while the story preserves the blocked-tool recovery edge.
Need the generated Chronicle artifact for complete per-tool receipt coverage.
The fallback story cites verifier and outcome-style evidence for the public proof packet.
The fallback story does not include real spend or latency metadata.
Need generated Chronicle artifacts with duration, token, or cost metadata.