You kept everything.Nothing connects any of it.
Four artifacts in four systems with nothing tying them to the same event are four artifacts, not a pack. This grades the joins, and the joins are what nobody grades.
Everything was logged. Nothing was joined.
Somebody asks what happened on the fourteenth. The decision is in the agent's output store. The inputs it saw are in an observability platform that rolls off after ninety days. The policy that was live is in a repo with no deployment record. The approval is in a ticket. Every one of those exists, and there is no identifier that runs between any two of them, so the only person who can say which rows belong together is the engineer who built it — and the account you produce is their memory with attachments.
From the decision to the context it saw, to the policy in force, to whoever authorised it. A chain is only as good as its joins, so the pack takes its weakest.
Two classes in the worked example are identical except for whether one person is still available. One reads NEEDS A WITNESS. The other reads NO RECORD.
The class in the worked example keeping seven years of artifacts is the one that cannot be assembled at all. The one keeping thirty days is fully answerable.
Remove one person and watch evidence become testimony.
Nothing ties the decision to the policy version in force at the time.
The class in the example keeping seven years of artifacts is the one that cannot be assembled at all. The one keeping thirty days is fully answerable.
This grades whether an account can be assembled, never a person, and says nothing about whether the decision was right.
The full output on the worked example, verbatim.
Seven decision classes, one command, no dependencies. This is the complete run, pasted as printed.
AUDIT EVIDENCE PACK ASSAY - could somebody who was not there assemble the account spec_version 1.0.0 | 7 decision classes read ID DECISION CLASS CONTEXT POLICY AUTH VERDICT DECIDED BY ------------------------------------------------------------------------------------------------------------ A-01 Refund auto-approval id id id ANSWERABLE - A-02 Discount auto-approval id person id NEEDS A WITNESS decision_to_policy A-03 Support ticket auto-close id - id NO RECORD contradiction_cap A-04 Content publish release id - id NO RECORD witness_gone A-05 Vendor payment release person id id NEEDS A WITNESS decision_to_context A-06 Inventory reorder id id id ANSWERABLE - A-07 Account closure id id id NO RECORD no_hub A LINK TO SOMETHING THAT IS NOT THERE A-03: decision_to_policy was recorded as carried, but the artifact at the other end does not exist. THE WITNESS HAS GONE A-04: decision_to_policy depended on somebody who is no longer available. BY QUESTION (read across every decision, worst first) QUESTION VERDICT ANSWERABLE WEAKEST ---------------------------------------------------------------------- decision_to_context NEEDS A WITNESS 5 of 6 A-05 decision_to_policy NO RECORD 3 of 6 A-03 decision_to_authorization ANSWERABLE 6 of 6 A-01 ESTATE 2 of 7 decision classes are ANSWERABLE end to end VERDICT: SOMETHING HERE HAS NO RECORD FIX FIRST: decision_to_policy (NO RECORD) - weakest at A-03 REPORTED, NOT GRADED Longest retention in the file: 2555 days. This figure is compared against nothing. 1 decision class(es) have no decision record at all: A-07 This grades whether an account can be assembled, never a person, and says nothing about whether the decision was right or whether it was in fact consistent with the policy.
Read the BY QUESTION block rather than the rows. This organisation can always say who authorised a decision and can almost never say which policy was in force — one problem with one fix, which reading it class by class would have shown as five.
Three rules, and the reason each one is not negotiable.
An identifier connecting the decision to a policy version that does not exist connects to nothing, whatever your register says. This is the one place the assay overrules an answer you gave it, and it only ever lowers a reading.
A pack that assembles because somebody remembers how the systems fit together stops existing when they leave. It did not degrade at that moment — it was never evidence. NEEDS A WITNESS names that state precisely rather than politely.
Which decisions are weak is rarely actionable. Which of the four questions you can never answer is one problem with one fix, and it is usually the same fix everywhere: one identifier that travels.
A reconstruction check, not a retention policy.
- A read-off of what carries each join between the four artifacts an account needs.
- A runnable engine, a workbook that reproduces it, and a demo that agrees with both.
- A per-question reading that names which of the four you can never answer.
- An emitter: one flag and the finding comes out as a verdict envelope.
- A retention policy. Whether you keep records, and for how long, is a different instrument.
- A judge of whether the decision was right, or whether it really was consistent with policy.
- A regulator-facing defensibility opinion. It asks whether an account can be assembled, nothing more.
- A scanner. Every input is an answer you supply about your own wiring.
Scope: this is an engineering instrument. It grades whether an account can be assembled, never a person, and asserts no regulatory position. It does not tell you whether a decision was right or whether it was consistent with policy, and it is not legal advice.
For the estate somebody will one day ask about.
- Automated decisions in your business affect customers, money or access.
- You could not say today whether the inputs a decision saw can be tied back to the decision itself.
- An account of any past decision would currently depend on one engineer being available.
- Somebody is going to ask, and you would rather find out now than then.
- One identifier already travels through every system you own. There is nothing here to find.
- You need a retention schedule. Different instrument.
- You want an opinion on whether a decision was defensible. This asks whether it is reconstructable.
The three it sits between.
Whether you keep records, and for how long. This assumes you do and asks whether they can be assembled into an account of one event.
OpenWhether one automated action left a usable record of itself. This asks whether four such records can be shown to describe the same event.
OpenWhat you do after an incident. This decides whether the reconstruction that postmortem depends on is possible at all.
OpenDirect answers, before you buy.
Four artifacts with nothing between them
are four artifacts.
One purchase, lifetime access, 12 months of updates. $119, once.
Scope: an engineering instrument. It grades whether an account can be assembled, never a person, asserts no regulatory position, and is not legal advice.
Sold by RedHub AI LLC · Secured by Stripe · redhub.ai