Research note 00 · Traceability, time, and append-only trust
Intelligence without a protocol is a voice with no record. It can be brilliant, and still leave nothing behind that anyone else can check.
So this station starts at the smallest honest question: can we prove that something existed, at a time, and that the record of it was never quietly rewritten? Two documents already answer that for the internet. This is an invitation to carry their answer into AI — and a reading list for anyone who wants to dig with me.
◆ proof of existence the minimum requirement for “traceable” root · 9f2c7ab4e1d0The two anchors you named
Both are public, free to read, and battle-tested at internet scale. Everything after this section is scaffolding around them.
Time-Stamp Protocol (TSP) · IETF Standards Track
Proof that something existed at a time — the minimum requirement for “traceable”. A trusted authority signs a hash plus a moment; the artifact itself never has to be revealed.
Read RFC 3161 at the RFC Editor →Certificate Transparency · IETF Experimental
The reference design for append-only public logs, proven at internet scale. Merkle-tree proofs let anyone verify that a log added an entry and never removed or reordered one.
Read RFC 6962 at the RFC Editor →“I don’t worry about AI. I’m working on AI with protocol — AI for AI, AI for good.”
Position of this station · open for global collaboration
Reference library · everything below is a live, primary source
Time first, then order, then receipts, then the rules we will be measured against, then the paper trail researchers already keep by hand.
Hashes, time-stamps, and how evidence survives decades.
The core request/response for “this digest existed by this instant”.
Modern hash agility for time-stamp tokens — read it beside 3161, not after.
The governance half: what an authority must promise to be believable.
Long-term preservation: re-time-stamping before algorithms age out.
The same durability idea in an XML archive world.
The origin text. Linking and witnessing, before anyone said “blockchain”.
Where the tree in every transparency log comes from.
The primitive all of this rests on; pair with FIPS 202 (SHA-3).
From certificates to any claim: logs anyone can audit.
The original design: SCTs, inclusion proofs, consistency proofs.
A decade of operational lessons folded back into the spec.
How the logs are actually run, monitored and trusted today.
The clearest short explanation of why public logging changes incentives.
A general-purpose transparency log you can point at non-certificate data.
Logs cheap enough to run and verify from plain static files.
The best intuition-builder for proofs, tiles and witnesses.
A transparency log shipped to millions of developers, quietly.
Live standards work on auditing what a service claims about identities.
The layer that would carry models, datasets, prompts and outputs.
The closest thing to “CT generalised”. If AI needs a log, it starts here.
Registries, statements, receipts — the vocabulary to reuse.
Compact, machine-checkable proof formats for constrained clients.
How you sign small structured objects at scale.
Unsung and essential: same meaning must hash to the same bytes.
Provenance for media, the public-facing edge of this whole argument.
Portable, verifiable claims about agents, models and people.
Who signed? An identifier that isn’t owned by one platform.
Governance texts that already demand traceability, in plain words.
“Accountable and transparent” as a measurable function, not a slogan.
Where provenance and content authentication are named as controls.
Logging, record-keeping and marking of synthetic content, in law.
The shared vocabulary most national policies were built on top of.
The most globally adopted ethical baseline — useful outside the West-only frame.
How an organisation proves it does what it says, repeatedly.
Research practice that a protocol could make verifiable instead of voluntary.
The artifact that most wants a signature and a time-stamp.
Provenance of data, written by hand — ready to be made checkable.
A large-scale audit of where training data actually comes from.
Evidence of how little is currently disclosed — the gap this work fills.
Open questions · where I would like help
If any of these is your field, this is the honest place to start a conversation.
When an AI output is time-stamped, what exactly is proven — the bytes, the prompt, the model weights, or only the moment?
Can a Merkle log absorb inference-scale volume, or does it need batched commitments per epoch, per model, per tenant?
What is the AI analogue of a certificate — a model card, a weights digest, a training-data manifest, or all three bound together?
Who operates the logs, who monitors them, and who audits the auditors across jurisdictions that do not trust each other?
How long must this evidence survive, and what does hash agility plus post-quantum re-time-stamping cost in practice?
Which harms does traceability genuinely reduce — and which does it merely relocate, onto users or onto the least powerful?
Where the work happens in the open
Mailing lists are free, public and archived. No membership, no gate.
Follow every draft, working group and revision as it moves.
The canonical archive — read the primary text, never a summary of it.
My feeling has a shape now: not fear of intelligence, but care for its record. Build the protocol, and AI can be answerable — to people, and to other AI. That is what “AI for good” looks like when you can verify it.
Open for collaboration · AI for AI, AI for good