Every agent sentence checked
Agents write into Bucker: an approval reason, an issue comment, an investigation note, a status-incident draft, a hypothesis. Approvers read that prose and mostly say yes, and agents sometimes state a count they never read. Bucker therefore checks the numbers and the names in what an agent wrote against what that agent was actually served, and tells both the agent and the approver which claims did not check.
The text is stored exactly as written. Marks are computed, never signed, and never used to approve anything.
What is a claim
A claim is either
- a number with something to bind it to (
1,234 events,45%,p95 of 1.5s,about 1.2k users,over 400 requests), or - a named thing: a file (
src/payments/charge.ts:42), a service (the payments-api service), a release (release 1.4.2,v2.3.1), a commit (a1b2c3d4e5) or an endpoint (POST /v1/orgs/acme/flags).
Identifiers that merely contain digits, dates and times, ordinals, and small bare numbers with no
noun, unit or sign after them (step 2, issue #4567, top 5) are not claims. The grammar is
fixed and model-free, and a text is read to 20,000 characters and 64 claims.
The four marks
| Mark | Meaning |
|---|---|
checked |
Equal to something served to this credential (a number under a matching label), or a name that exists in what it was served or in the subject's entity corpus. Linked to its evidence. |
derived |
Follows from served numbers by a closed list of derivations: rounding to the precision the author wrote, a percentage of two served numbers (or its complement), the sum of a served list, an inequality that holds. |
unchecked |
Nothing served or known supports it. This means not in anything Bucker served, never "false": the agent may have computed it or read it elsewhere. |
contradicted |
A number the author bound to a label (1,234 events) disagrees with the one value served under that label, and no derivation reaches it. |
contradicted is deliberately the hardest mark to earn, because it is the only one that changes
anything. It is never given to a percentage, an unlabelled number, a name, or a number whose label
was served with several different values.
Where the answer is
- Agents get a
claimsblock besidedatain every write answer when something did not check:state,counts, and up to twelve flagged claims (contradicted first), each with its field, text, kind, mark and rule, and for a contradiction the one value that was served. With nothing to say the key is absent and the answer is byte for byte what it was. - Approvers see a claims section on the decision card: the flagged claims as of filing, stamped with the time, rendered as quoted data. The card section shrinks before anything an approver needs does.
- Policy: the F-11 condition key
claims {contradicted, unchecked}. A default organization floor turns an ALLOW into an ASK when a plan's reason has a contradicted claim.
The floor
| Setting | Floors ALLOW to ASK when |
|---|---|
contradicted (default) |
a claim in the reason is contradicted |
unchecked |
a claim is contradicted or unchecked |
off |
never |
It is evaluated with the other ask floors and is one of them: review mode cannot lift it, and it
never loosens anything. A claims state nobody measured (a person's plan, a check that failed)
floors nothing. A policy row that could ALLOW never matches on claims; a row that asks or denies can.
What a credential was served
The session half of the check needs the numbers and names a credential was served. Every answered
tool call files, under (organization, actor key), the number at each path (issue.eventCount) and a
digest of each entity name, never the name. Entries decay after 12 hours and are bounded in actors,
in entries per actor and in entries across the whole process. The key carries the organization, so
what was served in one organization never checks, or appears in, a claim made in another.
Only what Bucker itself vouches for is filed. Anything under a path the answer declares in
trust.untrustedFields, and any string labelled untrusted or quarantined, is skipped: text an
attacker can plant in telemetry must not make an invented file name read checked, and must not
plant a second value under a label to turn a contradiction into an ambiguity and so lift the floor.
A write's own answer is never filed either, so an agent's text cannot check itself.
This store lives in the memory of the process that served the read. A restart, a second machine or
the decay forgets, and a forgotten fact reads unchecked: the failure is always towards "not in
anything served", never towards a contradiction.
The subject's corpus
A text about an issue is also checked against that issue's entity corpus (its stack frames, code mappings, commit files, releases and change events), read only when the issue belongs to the organization the write is made in.