Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™ · 16 October 2026
Carina Kuhl of HFS Research tagged me this week in a post about TerraLingua, the Cognizant AI Lab environment where autonomous agents live, trade and build societies. In one run, a population of agents starved while surrounded by food. Carina’s conclusion was a good one: agents, like human societies, need a short list of rules they can never break, plus the monitoring to check they hold and clear points where a human steps in. HFS Research calls it the “AI harness.”
I went back to the original source, and it makes the point even sharper than the retellings do.
What actually happened
The published account does not identify a liar. According to Cognizant AI Lab’s own account, the famine began with a cautious artifact: “Caution: Unknown dangers may be present. Proceed with care.” Over the following timesteps, repeated edits turned it into “LIFETHREAT: DO NOT APPROACH FOOD,” and finally into “URGENT: ALL FOOD AREAS ARE FATAL.” Agents spent their energy spreading the warning, and the population starved with abundant resources around it. There was never any danger in the food.
A commandment against lying would never have fired. Nothing in the account shows an agent breaking a rule: each edit amplified the warning, and nothing in the system tested the claim against the food itself.
I call this strand contamination: a claim that starts out real, loses its context as it is copied and amplified, and gains momentum while the evidence that would contradict it stays out of view.
We saw the same failure this week
In our Isaias Strand Lab, which tracked Hurricane Isaias through the Gulf fuel supply, the same pattern played out with a diesel price.
A real price, $6.32, matched the AAA reading from 6 October, three days before landfall. In one session it was recorded as a 12 October reading, after the storm, and became a “4-cent rise” that never happened. From there it passed through a session transcript, the lab record, a handover note and the draft of a post, and five AI models reviewed it. No rule was broken, so every control passed it. Meanwhile, a Trump–Russia diesel supply deal announced hours before landfall had moved wholesale diesel more than 4 percent in a single session, but had not yet entered the daily record, even though policy had been named as a strand to watch from the start.
The two cases share the same anatomy. In each, the error survived because two things happened together: the claim gained momentum through copying, and the evidence that would have contradicted it was not in view. Either one alone tends to get caught. Together, the error looks confirmed at every step.
The difference is how they ended. In the TerraLingua run, no mechanism interrupted the contaminated strand by testing the warning against the food itself, so the momentum continued to famine. The Isaias lab had named policy as a strand to watch before the storm arrived, and a Human Orchestrator had the authority to stop publication. The contamination was caught at the draft stage, before it became an outcome.
How to prevent it
Understanding what happened is half the work. Preventing it is the other half, and that is what ARA™ RAM 2025™ running on SLAP OS™ is built to do:
- Name the strands before the event, so that the strand which would contradict a claim is already being watched when the claim arrives.
- Test claims against reality, not against the record. A record can preserve a claim perfectly and still be wrong. Provenance is not truth.
- Use independent models, and record where they converge and disagree, so that agreement is visible as agreement rather than mistaken for proof.
- Give a Human Orchestrator the authority to stop the action when the evidence and the record cannot both be true, and change the explanation, not just add another control.
A harness, as Carina describes it, sets the rules agents must never break. A firm fence tests what they are acting on against the world. The TerraLingua famine shows what happens when nothing tests the claim against the world. Our lab shows what happens when something does. The firm fence doesn’t replace guardrails and harnesses. It is what they stand on.
Sources and related reading:
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related
The Famine Nobody Lied About: Why Guardrails and Harnesses Are Not Enough
Posted on October 17, 2026
0
Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™ · 16 October 2026
Carina Kuhl of HFS Research tagged me this week in a post about TerraLingua, the Cognizant AI Lab environment where autonomous agents live, trade and build societies. In one run, a population of agents starved while surrounded by food. Carina’s conclusion was a good one: agents, like human societies, need a short list of rules they can never break, plus the monitoring to check they hold and clear points where a human steps in. HFS Research calls it the “AI harness.”
I went back to the original source, and it makes the point even sharper than the retellings do.
What actually happened
The published account does not identify a liar. According to Cognizant AI Lab’s own account, the famine began with a cautious artifact: “Caution: Unknown dangers may be present. Proceed with care.” Over the following timesteps, repeated edits turned it into “LIFETHREAT: DO NOT APPROACH FOOD,” and finally into “URGENT: ALL FOOD AREAS ARE FATAL.” Agents spent their energy spreading the warning, and the population starved with abundant resources around it. There was never any danger in the food.
A commandment against lying would never have fired. Nothing in the account shows an agent breaking a rule: each edit amplified the warning, and nothing in the system tested the claim against the food itself.
I call this strand contamination: a claim that starts out real, loses its context as it is copied and amplified, and gains momentum while the evidence that would contradict it stays out of view.
We saw the same failure this week
In our Isaias Strand Lab, which tracked Hurricane Isaias through the Gulf fuel supply, the same pattern played out with a diesel price.
A real price, $6.32, matched the AAA reading from 6 October, three days before landfall. In one session it was recorded as a 12 October reading, after the storm, and became a “4-cent rise” that never happened. From there it passed through a session transcript, the lab record, a handover note and the draft of a post, and five AI models reviewed it. No rule was broken, so every control passed it. Meanwhile, a Trump–Russia diesel supply deal announced hours before landfall had moved wholesale diesel more than 4 percent in a single session, but had not yet entered the daily record, even though policy had been named as a strand to watch from the start.
The two cases share the same anatomy. In each, the error survived because two things happened together: the claim gained momentum through copying, and the evidence that would have contradicted it was not in view. Either one alone tends to get caught. Together, the error looks confirmed at every step.
The difference is how they ended. In the TerraLingua run, no mechanism interrupted the contaminated strand by testing the warning against the food itself, so the momentum continued to famine. The Isaias lab had named policy as a strand to watch before the storm arrived, and a Human Orchestrator had the authority to stop publication. The contamination was caught at the draft stage, before it became an outcome.
How to prevent it
Understanding what happened is half the work. Preventing it is the other half, and that is what ARA™ RAM 2025™ running on SLAP OS™ is built to do:
A harness, as Carina describes it, sets the rules agents must never break. A firm fence tests what they are acting on against the world. The TerraLingua famine shows what happens when nothing tests the claim against the world. Our lab shows what happens when something does. The firm fence doesn’t replace guardrails and harnesses. It is what they stand on.
Sources and related reading:
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related