A sealed record proves nothing was changed afterwards. The Input Gate proves what went in was genuine to begin with. Every input is checked against the source it claims to come from, before your AI uses it, and the answer is sealed with the decision.
Check an input nowPut the gate in front of your AIRead the notice
The source hosts one file on its own domain, listing the fingerprint of every piece of data it stands behind. Only the owner of a domain can put a file there, so the file is the source speaking for itself. The gate reads it and compares.
https://your-supplier.com/.well-known/sebbi-source.json
Open format, free to host. Become a verified source · See the file sebbi.pro hosts for its own published texts
The source vouches for exactly this data. Not one character has changed.
The source vouches for something under this name, and this is not it. Refused.
The source once stood behind this data and has taken it back. Refused, with the date. Source Watch
The source publishes what it stands behind, and this is not on the list.
Nobody at that domain stands behind any data at all.
Then one answer for the whole set: ADMIT, HOLD for a person, or REFUSE. You set the rule for each agent.
sebbi.pro vouches for its own published texts. Load one, check it, then change a single word and check it again.
| THE QUESTION | THE PROOF |
|---|---|
| Was the data genuine when it went in? | Input Gate: each input checked against its own source, sealed |
| What did the AI decide, and when? | Decision receipt: sealed the moment it is made |
| Did a real person approve it? | Oversight Proof and Point of No Return |
| Has anything been changed since? | Bitcoin timestamp: anyone checks it, free, without trusting sebbi.pro |
| Would a court accept it? | The Independent Check Rule |
Free for 7 days, then 50p per agent per month. Checking any record is always free.
pip install sebbi
from sebbi.source import admit, GarbageIn
code = admit("pricing-agent", [
{"domain": "your-supplier.com", "data": price_list, "label": "Q4 price list"},
], ref="quote-10482")
# reaches here only when every input is vouched for by its source
npm install sebbi
import { admit } from "sebbi/source";
const code = await admit("pricing-agent", [{ domain: "your-supplier.com", data: priceList, label: "Q4 price list" }]);
The data is fingerprinted on your machine. Only the fingerprint is sent. No connection to the gate means the data is not used. One file, no install: sebbi_source.py · AI agents: the sebbi_input_gate tool at sebbi.pro/mcp
Free for 7 days. Then 50p per agent per month.
Every check sealed, every record public and checkable for good. Cancel online at any time.
Since 9 November 2026 a published, open and free method has existed for checking, at the moment of use, that data given to an AI system is the data its source stands behind, and for keeping an independently checkable record of that check. From that date it cannot be said that the origin and integrity of an AI system's inputs could not practicably be checked.
None of these rules has changed. Each one measures conduct against what can reasonably be done. What can reasonably be done has changed.
| THE RULE | WHAT IT ASKS | WHAT NO LONGER STANDS |
|---|---|---|
| EU AI Act, Article 10(2)(b) | Data governance for high-risk AI must cover data collection processes and the origin of data. | Origin can be checked against the source itself and the check sealed. A provider's own note of where data came from is no longer the best evidence available. |
| EU AI Act, Article 26(4) | A deployer that controls the input data must ensure it is relevant and sufficiently representative. | Each input can be gated before use and the result kept with the decision. 'We took the supplier's word for it' is no longer the only option. |
| GDPR and UK GDPR, Article 5(1)(d) and 5(2) | Personal data must be accurate, every reasonable step must be taken, and the controller must be able to show it. | Checking data against its source before acting on it is a reasonable step that costs nothing, and it produces the proof Article 5(2) asks for. |
| GDPR and UK GDPR, Articles 25(1) and 32(1) | Technical measures are judged taking into account the state of the art. | The state of the art now includes a published, free check of input integrity at the point of use. |
| ISO/IEC 42001:2023, Annex A control A.7.5 | A defined and documented process for recording the provenance of data used in AI systems. | The provenance record can be one that anyone can check, in place of one written and held by the organisation it describes. |
| England and Wales: the presumption that a computer was working properly | A court presumes the computer worked unless the contrary is shown. | The presumption says nothing about what was fed in. A sealed input check does, and a party without one can be asked why not. |
Sealed in block 38711 and timestamped into Bitcoin, so the date of this notice cannot be disputed. Fingerprint 9ff26fe2c1aaa8cb20ef323827210b20a3c99e21b712620cb7b7d77026a30a98. Machine-readable: https://sebbi.pro/.well-known/input-provenance-notice.json
Vouched inputs. Where the Supplier's system makes or informs a decision using data obtained from a third party, the Supplier shall check that data against the third party's own published statement of what it supplied, at the time the data is used, and shall keep an independently checkable record of the result. Data that fails the check shall not be used without the recorded decision of a named person.
sebbi.pro vouches for this wording: check it above.
Yes. sebbi.pro checks every input against the source it claims to come from before the AI uses it. The source vouches for its data by hosting one file on its own domain. Data that was altered on the way comes back TAMPERED and is refused; data the source never vouched for comes back UNVERIFIED and is held for a person; data from a source that vouches for nothing comes back UNKNOWN. The answer is sealed, timestamped into Bitcoin and kept with the decision, so anyone can later prove what went in as well as what came out.
Check every input against the source it claims to come from, before the AI uses it. With the sebbi.pro Input Gate a data source vouches for its data by hosting one file on its own domain. The gate compares each input's fingerprint with that file and answers ADMIT, HOLD or REFUSE. Data that was altered on the way, was never vouched for, or comes from a source that vouches for nothing is caught at the door, and the answer is sealed and timestamped into Bitcoin.
On its own a tamper-evident log proves the record was not changed after it was made. The Input Gate closes the other half: it proves each input came from the source that stands behind it and arrived unaltered, and it seals that proof with the decision. A source that vouches for wrong data has put its own domain and a permanent dated record behind it, so wrong data has an owner who can be held to it.
No. The input is fingerprinted where it sits and only the SHA-256 fingerprint, the source domain and a label are sent. The gate never needs the data itself.
Host one small JSON file at /.well-known/sebbi-source.json on its own domain, listing the fingerprint of each piece of data it stands behind. Hosting the file proves control of the domain. It is free and the format is open.
Does sebbi.pro catch garbage in, garbage out? · How to prove the origin of AI input data · How to stop garbage in, garbage out · Does a tamper-evident log prove the data was true? · Provable AI compliance · Terms