lowob

Supersede, never overwrite

Strategy · AI

In short: Finance already knows how to store what was true and when it was known. Apply the same discipline to beliefs. A memory that edits itself in place cannot explain itself, and the change is the information.

Key points

  • Record both an as-of date and an as-at date on every judgment record.
  • A supersession link should carry the date, author, reason and trigger for the change.
  • The trigger field separates new evidence, error correction and reinterpretation of the same evidence.
  • Full history enables hindsight-free review, regulatory reconstruction, reversal patterns and replay.
  • Default retrieval to current records, with history one click away.

Most knowledge stores are updated like documents: open, edit, save. The previous version may survive in a history panel nobody reads. For a judgment system this is fatal, because when an organisation changes its mind, the change itself, and the reason for it, is the most valuable thing to record.

Two clocks, not one

Risk and regulatory data teams have long distinguished between two times. As-of: when something was true in the world. As-at: when the firm knew or recorded it. A restated figure has an as-of date in the past and an as-at date today. Bi-temporal databases have supported this for decades, and regulatory reporting depends on it.

Judgment records need the same two clocks. When was this assessment valid? And when did the organisation hold it? Without both, you cannot answer the question every review eventually asks: what did we believe at the time we decided?

What a supersession records

When a new view replaces an old one, the old record is not edited. A new record is created, linked to the old one by an explicit supersedes relation. One temporal knowledge-graph design does exactly this automatically when a contradiction arrives. The link should carry:

  • the date of the change;
  • who made it;
  • the reason, in a sentence;
  • the trigger: new evidence, a correction of an error, or a reinterpretation of the same evidence.

That last field is underrated. An organisation that frequently changes its mind on the same evidence has a different problem from one that changes its mind because new facts arrived.

What it buys

Hindsight-free review. A past decision can be judged against what was knowable at the time, not against what is known now. That is fairer to the people who made it and more useful for learning.

Regulatory answers. "What did you believe when you approved this, and on what basis?" becomes a query rather than a forensic exercise.

Reversal patterns. Over time you can see which kinds of belief the organisation revises most often, and how long each view typically survives. A view that is reversed every six months is telling you something about how it is formed.

Replay. You can rerun a decision with the record as it stood, to test whether a new model or process would have done better on the same information.

The cost

Storage and a little discipline. Records grow by a fraction of their size each time a view changes. Retrieval needs to default to current records and expose history on request. Both are trivial next to an organisation that cannot say why it changed its mind.

In practice

  • Add as-of and as-at dates to every judgment record.
  • Replace edit-in-place with supersession: new record, explicit link, date, author, reason, trigger.
  • Default retrieval to current records, with history one click away.
  • Once a quarter, list the most frequently reversed views and ask why.

What would prove this wrong. Organisations with edit-in-place memory doing as well on hindsight-free review, regulatory reconstruction and learning from reversals as those with full supersession history. Then the history is not earning its keep.

Common questions

What is the difference between as-of and as-at dates?

As-of is when something was true in the world. As-at is when the organisation knew or recorded it.

Why should a knowledge base never overwrite old entries?

When an organisation changes its mind, the change and its reason are the most valuable thing to record. Edit-in-place makes it impossible to say what was believed at the time of a decision.

What should a supersession record contain?

The date of the change, who made it, the reason in a sentence, and the trigger: new evidence, correction of an error, or reinterpretation of the same evidence.

What does keeping full decision history cost?

Storage and a little discipline. Records grow by a fraction of their size each time a view changes.

READ NEXT

Your Startup Does Not Have a GTM Strategy. It Has a List of Tactics.

Playbook · 4-minute read