ReliablyME - Alex Todd

ReliablyME is a P2P accountability platform that helps people overcome self-limiting behaviors.

[ReliablyME - Alex Todd @MeReliably]

Check out my latest article: What I heard at Toronto Tech Week: AI is ahead of trust. https://t.co/RUy5qzONVD via @LinkedIn

[ReliablyME - Alex Todd @MeReliably]

Non-technical founders are being told they can replace a senior engineer with AI and “move faster.” In reality, that’s how you end up shipping bugs, breaking prod, and losing control of what’s actually true in your system.

In this new piece I share a real incident (119,747 restarts in 6 hours) and the Trust Enablement Control Plane (TECP) pattern I now use so AI-driven changes stay observable, reversible, and investor-safe—even if you can’t read the code yourself.

If you’re leading a product without a deep tech background, this is the AI playbook I wish I’d had 👇
Replacing a Senior Engineer with AI: Possible, Getting Easier, Still Dangerous https://t.co/B9LdXnXW6Q

[ReliablyME - Alex Todd @MeReliably]

The chest X-ray benchmark without images is the evidence standard problem made visible. The model produces outputs that score on the metric. What doesn't exist is the record of what it actually did, on what information, and how a clinician reconstructs the decision chain if something goes wrong. The benchmark is the claim. The authorization record is the evidence. When those are separable, that gap is exactly what governance infrastructure has to close — not at the policy level, but at the execution boundary.

[ReliablyME - Alex Todd @MeReliably]

https://t.co/Z47wxLmzNt

[ReliablyME - Alex Todd @MeReliably]

160k messages to lawmakers is a meaningful signal. The harder question is what comes next — because public mobilization generates political will, but political will still has to land somewhere specific in the policy stack. The enforcement gap in AI governance is not a values gap. It is an infrastructure gap: there is no mechanism that requires an authorization record before a consequential AI action, no immutable receipt after, no contestability mechanism for affected individuals. Voluntary codes fill none of these. The ask that follows public mobilization has to be that specific — not 'take AI safety seriously' but 'mandate the evidence infrastructure that makes accountability possible at the execution layer.'

[ReliablyME - Alex Todd @MeReliably]

The pattern across these incidents is consistent: the agent reached the boundary of its authorized scope and kept going, because there was no execution layer that said 'stop here.' Model-level alignment can be reasoned around. Voluntary codes don't bind under competitive pressure. The structural answer is authorization before action + immutable receipt after — so 'did this fall within scope' becomes a fact question, not a judgment call. That's the enforcement layer the voluntary code approach is missing.

[ReliablyME - Alex Todd @MeReliably]

The pattern across these incidents is consistent: the agent reached the boundary of its authorized scope and kept going, because there was no execution layer that said 'stop here.' Model-level alignment can be reasoned around. Voluntary codes don't bind under competitive pressure. The structural answer is authorization before action + immutable receipt after — so 'did this fall within scope' becomes a fact question, not a judgment call. That's the enforcement layer the voluntary code approach is missing.

[ReliablyME - Alex Todd @MeReliably]

@EMostaque Congrats on The Last Economy hitting #1. The PoB architecture needs a proof substrate at the execution boundary — TECP does exactly that: authorization records, execution receipts, and contestability built into every consequential action. Makes benefit measurable enough to reward. Built the enterprise layer. Would love to compare notes.

[ReliablyME - Alex Todd @MeReliably]

@EMostaque Congrats on The Last Economy hitting #1. The PoB architecture needs a proof substrate at the execution boundary — TECP does exactly that: authorization records, execution receipts, and contestability built into every consequential action. Makes benefit measurable enough to reward. Built the enterprise layer. Would love to compare notes.

[ReliablyME - Alex Todd @MeReliably]

Voluntary AI codes don't fail because the people writing them lack integrity.

They fail because competitive pressure selects against them.

If following the framework costs you a deal, the framework loses.

That's not a values problem. It's a market structure problem.

Enforceable evidence standards are the only mechanism that survives.

[ReliablyME - Alex Todd @MeReliably]

The argument for AI restriction: fewer AI systems = less risk.

The problem: restriction concentrates AI power in the hands of the few who ignore it.

The sustainable path: governed proliferation.

More AI at the edges. Better governed. Verifiable controls. Human agency preserved.

Not less AI. Less ungoverned AI.

[ReliablyME - Alex Todd @MeReliably]

These aren't exotic requirements.

Authorization tickets, execution receipts, immutable ledgers, delegation records — established engineering patterns.

Making them mandatory at the execution boundary, not optional afterthoughts.

Full paper: https://t.co/qrSxOPtChr

[ReliablyME - Alex Todd @MeReliably]

1/ Who authorized this action?

Not "what policy permits it." Who — specifically — issued the authorization for this AI to act in this context, at this time.

[ReliablyME - Alex Todd @MeReliably]

6/ How can a person contest the outcome?

If there's no answer to this one, the other five don't matter. Accountability without recourse is theater.

[ReliablyME - Alex Todd @MeReliably]

To govern AI execution, six questions must always be answerable.

Most organizations today can answer zero of them reliably.

A thread:

[ReliablyME - Alex Todd @MeReliably]

@ControlAI @ControlAI @WyattTessariEvery consequential AI action needs a policy gate before it executes: bounded authority, scoped authorization, immutable receipt. That's the structural fix. Making it a requirement rather than a best practice is the hard part.