Hindsight in the postmortem
You read the account as it is now and conclude the decision was obvious. It was not, and the record from that week proves it.
Read a record as it stood that day, without today leaking in.
Every save writes a revision. Time Machine takes a record and two days, resolves the revision that was in effect at the end of each, and shows both panels with a field by field difference table. Where it cannot answer, it says which reason applies and renders no values at all rather than quietly falling back to the current draft.
In the workspace
From an empty record to something you can act on, in the order you do it.
Pick the record
Choose the record type, then the record by its title. Changing the type clears the selection rather than carrying it over.
Pick two days
Neither later than today. The earlier one becomes the earlier panel on its own.
Read the difference table first
Fields in stable catalog order, then work back up into the panels for the full context around what moved.
In the app
Captures of Growth-OS itself, rendered with sample data rather than our own records.


What goes wrong when this work lives in spreadsheets, documents and one person's memory.
You read the account as it is now and conclude the decision was obvious. It was not, and the record from that week proves it.
A fallback to the newest version is worse than no answer, because it looks like an answer.
Creation is decided by the lowest revision number in the window, never by the earliest timestamp, so a clock that ran backwards does not become evidence.
Highlights
Record missing, no history, not yet created, history truncated, or a draft it cannot read. Each says which, and none of them shows the current version instead.
Conversion is always UTC, so a colleague in another country reads the same day you do.
A stored draft is read through a deliberately looser schema, so a revision saved under since-changed rules is still readable.
Honest limits
This is recorded time, not valid time. A record edited on the eighteenth about something that changed on the first looks the same as one that changed on the eighteenth, unless effective dates were authored on it.
It is not a full decision replay. External source content, historical model behaviour and linked metric definitions are not pinned, and following a link opens that record as it stands today.
These pages share records with this one, so what you write here shows up there.
The page opens on this project's records. The in-app guide describes every field.