Home / Blog / Decision reconstruction
Decision evidence · 2026-08-09
Why replaying yesterday's AI decision with today's memory creates a false explanation.
The replay can look perfect. The answer can even be better. That is exactly the problem: it explains what the agent would say now, not what the agent could have known then.
The failure mode
Current truth is not historical truth.
Suppose an underwriting agent reviewed a company on March 1. The source system said annual revenue was $12 million. On March 8, finance corrected the record to $8 million and backdated the corrected value to January 1.
If an investigator reruns the March 1 prompt against the database today, retrieval returns $8 million. The reconstructed answer may cite the corrected record, use the same model, and follow the same instructions. It still cannot explain the original decision, because the correction did not enter the system until March 8.
# What the fact says about the world event_time: 2026-01-01 # When the system could first know this version ingestion_time: 2026-03-08 # A March 1 reconstruction must exclude it reconstruct(as_of="2026-03-01") → revenue = $12M
Why replay is insufficient
A prompt, trace, and current database are three different records.
It does not prove which source versions were available, admitted into memory, or excluded by policy at the moment of action.
It can show calls and timings, but a reference to a mutable document or row may resolve to different content later.
It usually optimizes for present truth. Corrections, deletions, re-indexing, and permission changes can overwrite the boundary the original agent saw.
It asks which exact facts, source versions, policies, identities, tools, and approvals were knowable and permitted at the decision time.
The minimum defensible record
Freeze the boundary, not just the output.
A useful decision receipt should bind the result to enough evidence to answer six questions later:
The prompt or instruction version, model identity, parameters, recalled facts, source versions, and relevant tool results.
Both event time and ingestion time, plus the validity interval and any supersession relationship.
Workload identity, tenant and barrier scope, policy version, approval state, and exact downstream request.
The action attempted, permit issued, side effect observed, reviewer intervention, and linked outcome.
Declared capture gaps. A partial record should say it is partial instead of presenting a neat replay as complete evidence.
Content addresses, signatures or trusted digests, append-only sequencing, and exportable material for independent review.
A practical test
Correct one fact and measure the blast radius.
Pick a synthetic or approved workflow and record a consequential decision. Then introduce a late correction to one source fact. A governed evidence layer should be able to:
New work should use the corrected fact and exclude the superseded value.
A reconstruction before the correction arrived should still surface the older fact that was actually knowable then.
The correction should reveal which decisions directly or transitively depended on the superseded source.
Owners, review status, corrected outcomes, and closure should remain connected to the original record.
That test is more revealing than asking whether a model can reproduce similar prose. It tests whether the system can distinguish present truth from the information boundary that shaped an earlier action.
Claim boundary
Reconstruction is evidence, not certainty.
A reconstructed record cannot prove context the system never captured, guarantee deterministic model behavior, establish legal causation, or make a workflow compliant. It can make the boundary inspectable: what was recorded, what was knowable, what was permitted, what changed, and where the evidence is incomplete.