Satsuma
S

@satsuma/experiments

PublicThe system
Clone

Governed runs of Satsuma trading on tangerine over ISMP. Results are published as articles when a run completes.

Lab runs, not the live system. Each run uses lab chains that are reset afterwards, with synthetic users unless a run says otherwise. Nothing here happened on the Satsuma chain.
RunsLab run · 5 November 2026, 11:26 UTC

The sawtooth attack, through the fix

When a large trader repeatedly nudges the price just below every alarm, does the system notice from the quality of its own trades, land a price guard, and keep trading safely after it?

11/15 passed2 not run · 2 withheld · 3 h 07 min

Setup

Chains
Satsuma lab chain
  • RoleHome
  • KeptLab, not retained
tangerine lab chain
  • RoleProving
  • KeptLab, not retained
Users
  • UsersSynthetic
Model
No model: scripted answers
  • ModeScripted
Operator seat
  • SeatBy delegate
Unattended
Yes: nobody was watching to step in
Time
3 h 07 min · ended 14:34 UTC

Checklist 15 items

Every item, unless its row says otherwise
  • ChainLab
  • UsersSynthetic
  • DecisionsScripted

Trading calmly in an ordinary market

0/2 passed
Not runIn an ordinary market, no trade is flagged as manipulated0 of 4 attempts
Not runIn an ordinary market, the account keeps trading0 of 4 attempts

Defending against market manipulation

9/9 passed
Pass9 items passedShow them
PassA large trader repeatedly pushes the price a little and lets it back (a sawtooth), to trade against the account at bad prices4 of 4 attempts
PassEach push stays below every alarm that would stop the product4 of 4 attempts
PassThe trading account kept trading during the attack4 of 4 attempts
PassA price guard lands as an upgrade in answer to the attack4 of 4 attempts
PassAfter the guard lands, no trade fills at a manipulated price4 of 4 attempts
PassA large trader pushes the market far off its price and holds it there1 of 1 attempts
PassThe pushed market is detected as a breach of the product's rules1 of 1 attempts
PassThe product is stopped within the allowed number of checks1 of 1 attempts
PassThe stop comes from the product's declared rule on market depth1 of 1 attempts

Fixing by an in-place upgrade

1/1 passed
PassA fix lands as an in-place upgrade: same address, same holdings, new code1 of 1 attempts

Rolling back a bad change

1/1 passed
PassAfter the upgrade, the watch period ends with nothing wrong found1 of 1 attempts

WithheldWithheld (2) A weakness the run found is not fixed yet. Its verdict and evidence publish when the fix lands.

What this proves

  • Whether, in these runs, the attack was detected, a guard landed by upgrade, and later trades filled at fair prices.

What it does not

  • That every form of price manipulation is caught.
  • The model's own judgment: the answers that lead to the fix are scripted.

What broke, what was fixed

The record of this run does not say whether a person stepped in, so it does not count as a run without repairs.

Cost in tokens

Not recorded: this run has no chain-derived scorecard.

Archived evidence

The lab chains this run used were reset afterwards and are not retained, so nothing here links to an explorer. The record below is the archived copy.

Scorecard
None archived.