Case study / Discovery and source recovery
Find what's actually there.
A useful answer about an old system begins with the things that really happened, not the neat story somebody can generate about them afterwards.
The problem
Old decisions rarely arrive in one convenient folder.
Work spreads across tickets, code, spreadsheets, documents and the people who remember why a strange decision once made sense. Over time, apparent duplication can turn out to be different versions of the same thing; the most confidently repeated explanation might only describe the latest visible branch.
When I was recovering the history of my own long-running creative work and conversations, that problem became surprisingly concrete. A good summary wasn't enough if it couldn't tell me exactly where a claim came from.
The approach
Recover the structure before asking AI what it means.
The method I designed starts with a source census: identify records, preserve relationships and establish which versions and branches exist. Only then should analysis group, interpret or propose which information deserves to become canonical.
That distinction led to Coin Cascade, a browser prototype that draws message nodes and branches from explicit parent-child relationships. It includes alternate routes, terminated paths, search, a movable canvas and evidence detail for a selected node.
Crucially, the prototype's graph is not generated by asking a model to invent the history. A model might eventually help annotate the recovered material, but the underlying relationships must come from the source export.
Result and boundary
A working visual question. Not yet a recovery engine.
The demonstrator proves a way to navigate and inspect a branching evidence graph. Its current data is mocked. It has not yet completed real-export parsing, source-level validation at scale or enterprise-system discovery.
The next useful test is to ingest an authorised real export, reconstruct its relationships deterministically and compare the result with a known source of truth. Only after that would I explore the more ambitious enterprise applications: system dependencies, architectural decision recovery and safe modernisation plans.
It's a small experiment with a serious underlying rule: don't promote an attractive interpretation into authoritative history until you can show the evidence.
Sources and limits
Something you can inspect.
Open Coin Cascade to inspect the mocked graph and click its nodes. The related Platform Archaeology workbench entry describes the wider proposed method.
Evidence classification: documented method and interactive prototype; no completed production ingestion or customer delivery claimed. The next validation milestone is explicitly unfinished.
Elsewhere on the workbench