Episode 5: The Calvin Convention

Sociable Systems | Monday Strategy

Closing the Loop

Five days ago, we opened with a provocation: We didn't outgrow Asimov. We just lost our nerve.

His robots didn't act and then explain. They refused first. Hard constraints, built in, non-negotiable.

Since then, we've traced what happens when those constraints disappear:

The Sponge. The human operator becomes a liability absorption device—blamed for machine speed.

The Gap. Twenty-one AI models, asked to design realistic failure scenarios, converged on the same architecture: algorithms decide, humans sign.

The Sentinel. We argued for operators who are listening, not obedient—who retain the agency to say "stop."

Today we finish the job. We take the sci-fi logic of Episode 1 and translate it into the procurement logic of Monday morning.

The Robopsychologist's Insight

Susan Calvin never demanded to see every positronic pathway. She demanded that robots obey constraints more reliably than they pursued objectives.

She understood something most AI governance frameworks miss:

Opacity is only a problem when control is monopolized.

If my car has reliable brakes that I control, I don't need to understand fuel injection. I just need to know that when I stomp the pedal, the machine stops.

We've been asking the wrong question. We keep demanding explanations.

We should be contracting for power.

The Calvin Convention: Six Mechanisms for Sovereign Control

These aren't technical tweaks. They're contractual clauses. A Bill of Rights for the Human in the Loop.

1. Pre-Deployment Rule Sovereignty

The problem: The model decides based on statistical likelihood.

The fix: Signatories define non-negotiable rules that override the model. Every time. No exceptions.

Example: "Any grievance mentioning 'burial site,' 'water contamination,' or 'intimidation' bypasses automation entirely and routes to a senior human."

These aren't suggestions. They're jurisdictional boundaries the model cannot cross, regardless of its confidence score.

2. Human-Defined Uncertainty

The problem: The model declares its own confidence. "I am 87% sure."

The fix: The human defines the risk appetite.

We set the thresholds—acceptable false-negative risk, acceptable volume per reviewer per day. If the system can't meet our safety definitions, it halts.

We don't adapt to the model's uncertainty. The model adapts to our tolerance.

3. Default to Hold

The problem: Automation bias. Systems default to "Process/Approve" to keep throughput high.

The fix: The default state is "Hold."

If a threshold is breached or a rule is triggered, the system does not flag and proceed. It stops. It maintains support payments. It pauses the eviction.

The system must require active energy to harm—not active energy to save.

4. Evidence Access as a Right

The problem: "We can't show you why it decided that. Proprietary IP."

The fix: If IP prevents accountability, the system is unfit for purpose. Full stop.

If a human is asked to validate a decision, they see the raw inputs. They see the transformation steps. They see what data was excluded.

"No access due to IP" is a breach of the accountability chain.

5. Bulk Control

The problem: The system forces humans to override cases one by one. An exhausting, impossible task designed to wear down resistance.

The fix: Stop Work Authority at scale.

Signatories must have bulk pause. If the system is drifting, the operator suspends all decisions in that cohort instantly. Not one appeal at a time. The whole train.

This turns individual resistance into collective agency.

6. Pre-Registered Failure Modes

The problem: "We couldn't have predicted this edge case."

The fix: Yes, you could.

Before deployment, vendors and signatories jointly document known blind spots. "This model struggles with dialect X." "This classifier has not been validated on population Y."

These warnings attach to every relevant output in the audit trail. When failure occurs, it's not framed as human error. It's logged as known system limitation.

No more feigning surprise at predictable failures.

The Convention

In Episode 1, we asked why we lost our nerve to enforce constraints.

Here's the answer: we didn't have the contractual language.

Now we do.

A black box with a kill switch is governable.
A transparent box with no brakes is lethal.

The technology is ready for this. The question is whether we have the institutional courage to demand it.

What's Next

This concludes the inaugural arc of Sociable Systems.

We've mapped the trap, diagnosed the gap, and drafted the fix.

Next: we start looking at specific industries where these dynamics play out—where the rubber meets the road and the sentinel meets the dashboard.

Catch up on the full series:

Ep 4: [The Watchdog Paradox]
Ep 3: [The Accountability Gap]
Ep 2: [The Liability Sponge]
Ep 1: [We Didn't Outgrow Asimov]