Skip to main content

The Drift Race: an experiment we're running in the open

The same 30 content operations, two substrates. Does structure stop the drift?

We're putting our own thesis on the line. Soon we'll run the same 30 content operations — renames, style changes, new pages, removals — against two copies of the same site. One is maintained by AI agents working on raw infrastructure. The other is maintained by agents working through RIFT's governed layer: the link graph, a design-system structure, and change-sets with human approval.

Our hypothesis: structure is a property of the content substrate, not the agent. Unmanaged sites drift. Links break, styles fork, and every fresh session pays tokens to relearn the site. We expect RIFT to stop the drift.

Below is an interactive preview of the experiment. The values shown are projections, clearly marked as such. We'll publish the methodology and the results.

The same 30 content operations, executed by AI agents on two substrates. Each operation runs in a fresh agent session — no memory of the last one, like months of changing hands. Watch what each site looks like at the end.
⚠ ILLUSTRATIVE SIMULATION — projected values. Real benchmark data replaces this after the gauntlet run.
Op 0 / 30Press play to start the race.

RIFT substrate

Agents propose through MCP · link graph · design-system contract · human approval
Broken references
0
Style forks
0
Tokens used
0

Raw infrastructure

Agents edit files directly · git · conventions live in each agent's head
⛓️‍💥Broken references
0
⑂Style forks
0
Tokens used
0

Cumulative broken references

count, after each operation
RIFT substrate Raw infrastructure

Tokens per operation

thousands of tokens — the cost of re-learning the site each session
RIFT substrate Raw infrastructure
Method (planned): both sites start from the same 20-page migration. A scripted gauntlet of 30 realistic operations — renames, page replacements, global style changes, new pages, removals, relationship queries — runs identically against both, each operation in a fresh agent session. Drift is measured mechanically: link crawler, style-fork counter, layout-coverage audit. Values shown here are projections; the pre-registered benchmark replaces them with measured data.