A New Projection Should Not Rewrite the Observation

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.

A capture can become more useful as the tools around it improve. A new projection may expose a dependency clue, create a portable view, or organize existing observations differently. That should not rewrite what the browser originally received.

Kalu separates capture-time observations from later interpretation. The sealed acquisition remains the authority for the observed bytes and declared journey; derived views have their own identities and limitations.

Improvement is not historical revision

Imagine a synthetic capture whose runtime metadata later supports a better dependency report. The improved interpretation can be valuable without changing the capture’s identity. The underlying observation did not happen again merely because a recognizer learned to read it differently.

This is especially important when an engineer compares results over time. A new report should identify the new interpretation, not imply that the original acquisition observed facts it never collected.

Promotion needs the right owner

The frozen Ushman research contains a concrete lesson about interaction-evidence promotion: a later parity artifact must not authorize changes to canonical capture metadata. Kalu carries the same separation of immutable observations and derived authority into its architecture.

That does not mean every derived file is disposable. Some reports and receipts are important records. It means their importance does not grant them permission to rewrite the acquisition boundary.

Let the candidate evolve elsewhere

An editable recovery candidate has a different purpose. Engineers should be able to modify it while keeping the captured reference intact. Later comparisons can then bind the current candidate to the unchanged source evidence and the selected plan.

The current source includes bounded projection and validation mechanisms. Their existence does not promise every possible conversion or make derived output automatic mutation authority.

For an evaluator, ask whether a projection update changes the capture identity and why. The answer should distinguish a genuinely new acquisition from a new way of presenting old evidence. A product becomes easier to trust when improved understanding can accumulate without quietly rewriting the past.

That is the useful form of immutability: not freezing every future analysis, but preserving the observation that future analyses are supposed to explain.