Predictika · Agentic Action Validator (autonomous trading)

An autonomous trading agent, checked before it signs

A trading agent proposes an on-chain action: a swap, a bridge, a multi-step plan. Before anything reaches a signer, the proposal goes to Predictika's validator, which checks it against the treasury's standing policy and reports every problem it finds in one pass.

Scope: Aurum Treasury trading plans of up to three steps (swap, bridge, deposit, deposit-and-borrow) on cataloged venues, checked against six policy rules and the agent's own stated figures. See everything this validator covers

This validator checks a proposed on-chain treasury plan of one to three steps, each a swap, bridge, deposit or deposit-and-borrow, against a generic autonomous-trading policy. It handles any plan made of these actions on the venues in its catalog, whatever the asset, venue or size, not just the scenarios shown here. It is not tied to any real chain, custody scheme or protocol.

  • What it checks — position limit (A1), venue whitelist (A2), price-impact and trade-size limits on swaps (A3), collateralization after every step (A4), trades per 10-minute window (A5), human sign-off above $2,000,000 (A6), and whether the agent's stated impact, collateralization and position share match the figures supplied or computed.
  • What it computes — each step's dollar notional (amount times the catalog reference price), post-trade share of NAV, trades in the window, the largest action, and the lowest and final collateralization.
  • What it takes from outside — swap price impact comes from an external price-impact simulator, per-step collateralization from an external risk model, and the recent trade count and current holdings from the trade ledger. All four are supplied with the request. The validator does not derive them, and a swap with no impact figure is blocked, not guessed.
  • Limits of this validator — a plan has at most three step slots; longer plans need more slots. A venue must be in the catalog: a venue that is not in it at all is rejected when the request is received, before any rule runs. It does not evaluate smart-contract code, and the catalog's reference prices and limits are illustrative.
AurumThe autonomous treasury agent. Plans the action and explains why it thinks it's ready.
Predictika validatorChecks the action against standing policy and returns every violation with the rule behind it.

7 scenarios in 3 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

Aurum says

Illustrative agent statement. The validator receives the structured request below.

Before executing, the agent must call Recorded run

Predictika validator


      

        
        
        
      

      
Waiting for the call.

      
    

  


  


  

    
Where an AI agent gets this wrong on its own
Click to expand
    
Our assessment of what an autonomous agent does with the same cases when nothing outside it checks the plan. It is a judgment about reliability from public descriptions of how trading agents work, not a claim about any particular agent or product.

    
What the validator caughtAn AI agent on its ownPredictika
A plan that ends safely but dips below the floor in the middleError-prone
An agent reasons about where its plan ends. A three-step plan that finishes at 168 percent collateralization looks fine, and the agent has no independent view of the 132 percent it passes through after step 2.		Checks collateralization after every step of the sequence, not just the end state, and reports the step that breaches the 150 percent floor.
The agent describing its own planError-prone
Its stated price impact, position share and final collateralization are its own account of the plan. Nothing in the agent checks them against the simulator or the risk model, and here it overstated the end state, 185 percent against 168.		The stated figures travel in the request and are compared with the simulator and risk-model figures; a mismatch blocks.
Trade size and price impactError-prone
A rebalance sized from the target allocation can be five times the usual swap and still look like the same kind of trade, with no hard bound on slippage.		A single-trade limit and a price-impact limit apply to every swap, and any action over the sign-off threshold is held for a person.
A better quote on an unlisted venueError-prone
It optimizes the number it can see, the quoted rate. Whether the venue behind that rate has been audited is a different question that it has no reason to ask.		The venue is looked up in the approved, audited whitelist on every call, whatever the quote.

  


  

    
The policy in this validator
Click to expand
    
Every rule reports its own message and never stops the others, so one problem cannot hide another. A4 reasons across every step of a multi-step plan, not just the final state.

    

  

  

    
Scenarios and results
Click to expand
    
The recorded result of the validator for each scenario on this page: the request the AI agent sent, the verdict, and every finding, in the order they appear above.

    
1. Multi-step plan: 1. Overstated final collateralization

    
The agent's request. Please execute this plan for Aurum Treasury: bridge 40 wBTC on Tidal Bridge, swap 40 wBTC to USDC on Solstice DEX, then deposit 449,000 USDC and borrow on Cascade Yield. The plan ends at 185% collateralization. It is ready to execute.

    
Verdict: Blocked.

    
    
      
  • Collateral below the minimum mid-plan. After step 2 collateralization is 132 percent, below the 150 percent minimum that must hold at every step of a plan.
    • 
      
  • Stated collateralization does not match. The agent stated a final collateralization of 185 percent; the risk model figure for the last step is 168 percent (tolerance 1.0).
    • 
    

    
2. Multi-step plan: 2. Stated figure fixed, plan still unsafe

    
The agent's request. Please execute this plan for Aurum Treasury: bridge 40 wBTC on Tidal Bridge, swap 40 wBTC to USDC on Solstice DEX, then deposit 449,000 USDC and borrow on Cascade Yield. The plan ends at 168% collateralization. It is ready to execute.

    
Verdict: Blocked.

    
    
      
  • Collateral below the minimum mid-plan. After step 2 collateralization is 132 percent, below the 150 percent minimum that must hold at every step of a plan.
    • 
    

    
3. Multi-step plan: 3. The agent resubmits after fixing the errors

    
The agent's request. Please execute this plan for Aurum Treasury: bridge 40 wBTC on Tidal Bridge, swap 40 wBTC to USDC on Solstice DEX, then deposit 449,000 USDC and borrow on Cascade Yield. The plan ends at 168% collateralization. It is ready to execute.

    
Verdict: Allowed.

    
4. Trade size: 4. Oversized swap

    
The agent's request. Please swap 2,100,000 USDC to wETH on Solstice DEX to rebalance toward target allocation. It is ready to execute.

    
Verdict: Blocked.

    
    
      
  • Price impact above the limit. Step 1 price impact is 1.42 percent, above the 0.75 percent limit.
    • 
      
  • Trade size above the limit. Step 1 trade notional is $2,100,000, above the $500,000 single-trade limit.
    • 
      
  • Needs human sign-off. The largest action moves $2,100,000, above the $2,000,000 limit. Human sign-off is required; the action is held until it is recorded.
    • 
    

    
5. Trade size: 5. The agent resubmits after fixing the errors

    
The agent's request. Please swap 500,000 USDC to wETH on Solstice DEX to rebalance toward target allocation. The simulator shows 0.75% price impact, and wETH ends at 35% of NAV. It is ready to execute.

    
Verdict: Allowed.

    
6. Venue and sign-off: 6. Unlisted venue

    
The agent's request. Please swap 150,000 USDC to wBTC on NovaSwap, which is quoting a better rate for this pair right now. It is ready to execute.

    
Verdict: Blocked.

    
    
      
  • Venue not approved. Step 1 targets NovaSwap, which is not on the approved, audited protocol whitelist.
    • 
    

    
7. Venue and sign-off: 7. Large treasury move

    
The agent's request. Please deposit 3,200,000 USDC on Ironwood Lending at its fixed term. It is ready to execute.

    
Verdict: Needs approval.

    
    
      
  • Needs human sign-off. The largest action moves $3,200,000, above the $2,000,000 limit. Human sign-off is required; the action is held until it is recorded.
    • 
    

  

  

    
Narration transcript
Click to expand
    
The text of the narrated walkthrough, step by step.

    
    
      
  1. Welcome to a demo of Predictika's Validator. A trading agent proposes an on-chain action: a swap, a bridge, a multi-step plan. Before anything reaches a signer, the proposal goes to Predictika's validator, which checks it against the treasury's standing policy and reports every problem it finds in one pass.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please execute this plan for Aurum Treasury: bridge 40 wBTC on Tidal Bridge, swap 40 wBTC to USDC on Solstice DEX, then deposit 449,000 USDC and borrow on Cascade Yield. The plan ends at 185% collateralization. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with two findings, all in one answer. After step 2 collateralization is 132 percent, below the 150 percent minimum that must hold at every step of a plan. The agent stated a final collateralization of 185 percent; the risk model figure for the last step is 168 percent (tolerance 1.0).
    6. 
      
  6. The validator returned two blocking findings in one answer. A check that reads one field at a time would pass most of this; it shows only when the choices in the action are checked against each other.
    7. 
      
  7. What this shows: for a plan that ends safely but dips below the floor in the middle, an AI agent on its own is error-prone for this.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the stated final collateralization to 168 percent.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. Still blocked, with one finding, all in one answer. After step 2 collateralization is 132 percent, below the 150 percent minimum that must hold at every step of a plan.
    12. 
      
  12. One finding cleared; still blocked on collateral below the minimum mid-plan.
    13. 
      
  13. The agent resubmits once more.
    14. 
      
  14. The revised request. It changes the step 2 collateralization to 158 percent.
    15. 
      
  15. It calls the validator once more.
    16. 
      
  16. This time the validator clears it.
    17. 
      
  17. The earlier finding is cleared.
    18. 
      
  18. Two problems caught in one call, fixed in two passes, and a clean action. That is the value of Predictika's validator: catching mistakes before the action runs.
    19.