Someone asked us recently whether DarkMatter is solving a problem that doesn't actually exist. The question was direct: nobody demands cryptographic proof of a text message, so why would anyone demand it for an AI agent's decision?
It's a fair challenge. And the answer changed how we think about what we're building.
Part one
Why text messages "work" as evidence without verification
Text messages, emails, and Slack logs are accepted in disputes all the time without anyone verifying them cryptographically. But here's what's actually happening when they're accepted: they're not being proven. They're being trusted until someone has a reason to question them.
That trust rests on three things that most people never think about:
An authoritative system of record. Your texts come from Apple or your carrier. Your emails come from Google Workspace or Microsoft Exchange. Those are institutional sources with their own audit infrastructure. Nobody asks you to prove a Gmail is real because Google is implicitly vouching for it.
A clean chain of custody. An auditor doesn't verify the cryptography on a Slack export. They verify the process. Who exported it? From which account? With what permissions? If that process is clean, they accept the data.
Low economic incentive to dispute. Disputing logs is expensive and slow. Unless the stakes are high enough to justify that cost, nobody goes there. Most of the time the stakes aren't high enough. So the logs just... work.
Text messages aren't proven. They are trusted until challenged. That's a very different thing.
When that trust does break, when someone does have a reason to challenge the logs, things get complicated fast. Forensics experts appear. Chain-of-custody questions surface. Screenshots get dismissed as weak evidence. The process that made everything feel solid starts to crack.
Part two
Why AI logs are fundamentally weaker
Here's where the comparison falls apart. AI logs are missing all three of those trust layers.
There's no authoritative system of record. OpenAI and Anthropic are not treated like legal systems of record the way carriers and email providers are. You cannot subpoena "the canonical version" of an AI conversation in a reliable, institutionally recognised way. What you have is an export, and exports can be selective.
There's no stable chain of custody. AI outputs are non-deterministic. The same prompt, run twice, produces different outputs. That means a transcript can be reconstructed to look like the original without being the original. Prompts can be modified. Outputs can be re-run. None of this is hypothetical. It's just how the technology works.
| Property | Carrier / Email logs | AI agent logs |
|---|---|---|
| Authoritative source | ✓ Institutional | ✗ No equivalent |
| Tamper-evident by default | ✓ Yes | ✗ Outputs can be replayed |
| Signed at creation | ✓ Yes | ✗ No |
| Reproducible without being provable | ✓ Hard to fake | ✗ Easy to reconstruct |
There's actually high incentive to dispute. This is the most important difference. With AI decisions, people want to push back. The model said the wrong thing. The agent made a bad call. The output was different to what was reported. These challenges are already happening and will become more common as AI takes on more consequential work.
With AI, the question isn't just "did this happen?" It's "can you prove what the AI actually decided, and that nobody changed the record after the fact?" Those are different questions. And today, you mostly can't answer the second one.
Part three
When this actually matters
Most AI decisions today don't need this. Someone using an AI to help draft a support reply doesn't need a tamper-evident record of what the model suggested. The stakes are low. Disputes are unlikely. Logs are fine.
But that changes the moment AI starts making decisions that carry real consequences.
An AI agent that flags a customer account for fraud. One that approves or rejects a procurement request. One that classifies a support ticket as high-priority and escalates it. Or doesn't. One that recommends a medical summary for a clinician. In those situations, "the AI said so" stops being an acceptable answer the moment someone is affected by the decision and wants to understand or challenge it.
Regulated industries using AI. Financial decisions. Healthcare summaries. Content moderation at scale. Multi-agent pipelines where decisions chain together.
Enterprise AI agents with budget authority. Autonomous customer-facing decisions. Any AI system where "why did it decide that?" is a question someone will ask.
Most everyday AI assistance. Drafting help, research summaries, internal tools where the human reviews everything and the stakes are low.
Part four
What we're actually building
DarkMatter is not a log verification tool. That framing makes it sound like we're adding a stamp of approval to something that was already fine.
What we're building is closer to what DocuSign did for contracts, or what financial ledgers do for transactions: it creates a record that's difficult to dispute because it was signed at the moment it was created, not exported after the fact.
When someone challenges what your AI did, this is the record you show them.
At L3, the agent signs its decision before the record leaves the agent. At every level, the payload is hashed client-side before transmission. The signature or hash is stored outside your system. If anything changes after that, whether the output, the reasoning, or the context, the chain breaks. Anyone can verify this independently, using standard cryptography tools, without trusting DarkMatter to tell them the answer.
That's not log verification. That's non-repudiation. The record says what it says, and there's no plausible way to claim otherwise.
Part five
Are we too early?
Probably, for most companies. Right now, most AI agents are still advisory. A human reviews the output, and the human's judgment is what actually matters. In that world, DarkMatter is nice-to-have at best.
But the direction is clear. AI agents are moving from advisory to autonomous. From suggesting to deciding. From assisting humans to acting on behalf of humans. Every step in that direction increases the moment when "we have logs" stops being enough and "we can prove it" becomes necessary.
Not: "Do people verify AI logs today?"
But: "Where do AI logs stop being trusted and start being disputed?" That boundary is exactly where DarkMatter lives. And it's moving toward more companies, not fewer.
The infrastructure that makes AI decisions provable needs to exist before the disputes arrive, not after. That's always been true of trust infrastructure. The time to build a ledger is before the first transaction is questioned, not after.
We're building it now.
DarkMatter records AI agent decisions as tamper-evident records stored outside your system. Every record carries a proof level that shows exactly how much you can rely on it, and anyone can verify it independently, without a DarkMatter account.