Controlled Boot Is Not a Deterministic Universe

Scope: Engineering or methodology note based on a September 17, 2026 snapshot. Source-level behavior is not a public availability or universal compatibility claim. Any scenarios are illustrative, not customer outcomes.

“Deterministic replay” is an attractive phrase. It can also suggest more control than a real browser workflow provides. KodeBack’s controlled-boot profile is more specific: it fixes selected execution conditions while leaving other browser behavior native.

That specificity matters because the profile is part of the comparison. A passing result is meaningful only in relation to the conditions under which the donor and candidate were observed.

Control selected inputs, not the whole application

In the September source snapshot, controlled boot fixes the browser’s Date value, seeds Math.random, and exposes buffered same-origin HTTP fetch responses in request-invocation order. Requests still begin normally. The profile also has explicit buffering and pending-request bounds.

Timers, animation frames, monotonic performance clocks, and Web Crypto remain native. A page whose behavior depends on those surfaces can still vary. The preset is therefore a bounded attempt to stabilize boot, not a universal simulator of browser execution.

Save the profile with the plan

The selected profile is bound into the saved plan and observation identities. A later check reuses it and does not accept a profile override. Updating the CLI does not silently change an existing workspace’s baseline.

This prevents a subtle form of moving the goalposts. An operator should not be able to switch to a more forgiving execution profile during a check while continuing to describe the result as the original comparison. Changing the profile requires a new preparation and a deliberate transfer of candidate edits.

Preserve raw disagreement

The controlled preset does not normalize an unstable result into passing. Raw observations, ordered network comparisons, console assertions, and seal validity retain their own requirements. Exhausting a required bound or failing to install the selected conditions cannot qualify the baseline.

The source also supports an explicit native-boot preparation profile for unmodified browser execution. That can be useful, but native timing and randomness may produce differences without candidate edits. Neither profile establishes every action or visual property of the application.

When evaluating a recovery report, ask which inputs were controlled and which remained native. “The same page passed” is incomplete. “This candidate satisfied this boot contract under this saved profile” is a claim another engineer can interpret.