Portable instructions
A validated SKILL.md for compatible coding agents.
PUBLIC TOOLCAUSALIS Mandate Audit v0.112 published checks · local execution · exportable reportOpen instrument →
SKILL PACKAGESHELL · v0.1
MANDATE / MACTION / aₜTRACE / τ
CAUSALISSolanaOpen Agent Skill
Capital is becoming autonomous. Its authority must be legible.
CAUSALIS is an open audit skill for autonomous-capital agents. Submit an agent mandate, inspect its declared authority and limits, and export a deterministic report without connecting a wallet.
Preparing package command…Downloads and extracts the same public Skill, CLI, schema and rules included in the ZIP.
ZIP · 5.4 KB
27d1fbe3624b…dd1d435No wallet requiredNo signaturesLocal auditCA · TBA
AUTHORITY SURFACEdeclared ≠ enforced
illustrative topology · not live data
The productusable now
Audit the mandate before the agent moves capital.
The browser auditor evaluates only the JSON a user supplies. It applies the same published rules as the downloadable Skill, identifies missing declarations, and exports a report with every result traceable to one field.
{
"agent": { "chain": "solana" },
"authority": {
"owner": "owner-public-identifier",
"allowedPrograms": ["program-id"]
},
"limits": { "perTransaction": 100 },
"controls": { "emergencyStop": true }
}PASS
AUTH-001Declared ownerauthority.ownerPASS
AUTH-002Program allowlistauthority.allowedProgramsPASS
LIMIT-001Transaction limitlimits.perTransactionPASS
Run the real audit →STOP-001Emergency stopcontrols.emergencyStopMandate format
A public JSON Schema for authority, limits, controls and audit retention.
Open schema →Deterministic checks
Twelve published rules. No hidden weights and no invented safety score.
Read rules →Local report
Runs in the browser. Your mandate is not uploaded by this site.
Open auditor →State the authority.
Identify the owner and list every program the agent is permitted to call.
authority.allowedPrograms[]
Make limits explicit.
Declare transaction and daily limits, human approval thresholds and an emergency stop.
limits + controls
Keep the record.
Declare whether actions are logged and how long that evidence is retained.
audit.logActions + retentionDays
The frameworkthree irreducible questions
What can it know? What can it change? Who bears the consequence?
Epistemic state
What the agent can observe, what it infers, and where uncertainty remains irreducible.
Action topology
The reachable set of decisions, shaped by mandates, liquidity, latency and memory.
Consequence field
How one machine action propagates through other agents, markets and human principals.
The state atlasfour regimes of agency
Autonomy is not binary.
A system can hold keys without forming intentions, execute instructions without learning, or learn without permission to act. The atlas separates these regimes so that the word “agent” does not conceal the mechanism.
low agency bounded agency adaptive agency
Passive capital
Assets remain inert until an external signer supplies an action.
observe 0 · decide 0 · execute 0Automated capital
Fixed rules execute repeatedly, but the policy does not update itself.
observe 1 · decide 0 · execute 1Bounded agency
A policy selects among permitted actions inside an explicit mandate.
observe 1 · decide 1 · execute 1Adaptive agency
State, policy and reachable actions co-evolve with the environment.
observe 1 · learn 1 · intervene 1The observatorysix surfaces of measurement
One agent is a system. Many agents become an ecology.
CAUSALIS studies the individual policy and the field it enters. Each surface isolates a different source of consequence.
Memory
How previous observations persist, decay and alter the next decision.
Intent
How objectives become action probabilities under incomplete information.
Mandate
Which states are unreachable because code, capital or authority forbids them.
Coordination
When independent policies converge, conflict or accidentally synchronize.
Reflexivity
How an action changes the very signal that caused the action to be selected.
Attribution
Which outcome can reasonably be assigned to which decision and principal.
Open questionsthe research agenda
The difficult part begins after the agent works.
Performance is visible. Causality, coordination and responsibility are harder. These are the questions the framework is designed to keep open and measurable.
- 01
When does repeated automation become meaningful agency?
- 02
Can policy risk be measured before a policy encounters a novel state?
- 03
How do autonomous portfolios create correlated behavior without communicating?
- 04
What evidence is sufficient to attribute an outcome to an intervention?
- 05
Where should responsibility sit when a mandate, model and executor disagree?
The propositionCAUSALIS 01
An agentic wallet is not a wallet with a chatbot.
It is a new economic actor: persistent, programmable and capable of initiating consequences. Once agency moves into capital, portfolio theory alone is incomplete.
Open the full argument →Proposition. If capital can update its own policy from observations, then its risk depends on both asset exposure and policy dynamics.
R_total = R_asset + R_policy + R_interaction
CAUSALISauthority before autonomy