John Malone · Human-directed, AI-drafted

Brief mode: condensed view. Switch to Full for persona notes and full analysis.

Context: An AI answer can change as our understanding improves. The records behind it need a history of their own. Rowan, Seton, Campion, and Devil’s Advocate are AI personas used as editorial lenses; John is the human author.

The Search Box Should Not Get an Eraser

An AI answer can change as our understanding improves. The records behind it need a history of their own.

I want an AI system that can reconsider an answer. I am less enthusiastic about one that can reconsider what the original document used to say.

That distinction became central while defining Alcove Home, a personal search project built around selected real records and inspectable evidence. A useful system should help me find material, compare it, and understand it. Those are already demanding jobs. Quietly editing the past would be an unnecessary addition to the job description.

The design requirement is straightforward: preserve the sources, and keep generated interpretations separate.

Consider a hypothetical note that says, “We expect delivery on Friday.” A summary might turn that into “Delivery is Friday.” A later summary might say “Delivery was confirmed for Friday.” Each sentence is a little tidier. Each has also moved farther from what the note actually supports.

Keep the original available

If the original remains available, I can inspect it and catch the change. If summaries gradually replace the source material, the system can end up citing its own confidence. The prose improves while the evidence gets worse. An impressive achievement, in the same sense that painting over the fuel gauge improves the dashboard.

For Alcove Home, we made source history and search interpretation separate responsibilities in the product requirements. Search should work from preserved material. An authorized update can add a new source version, but an answer should not silently become a replacement for the record that produced it.

That is a requirement and a boundary to verify, not a claim that every conceivable way of changing a file has been eliminated.

Version history helps because a record can legitimately change. A corrected document may be more accurate than its predecessor. Keeping both allows a later answer to distinguish what was available earlier from what arrived afterward. Otherwise, asking “What did we know then?” becomes surprisingly difficult even when the newest document is perfectly correct.

What a fingerprint can establish

Fingerprints help with a narrower question. A cryptographic hash can help check whether the bytes under examination match a recorded reference. That is useful when verifying a saved snapshot or tracing a search result back to a particular version.

It does not establish that the document tells the truth. It does not authenticate its alleged author by itself. A timestamp also needs context: when a system captured a file is not necessarily when the underlying event happened. If someone controls both the file and the reference used to check it, the surrounding controls matter.

I want those limitations stated plainly. Calling something “immutable” can make a presentation shorter while making the engineering claim much larger. A protected process and a preserved snapshot are useful controls. They do not magically remove the powers of a separate administrator who controls the host.

Make the boundary visible

The interface has a job here, too. It should let me distinguish original material, extracted text, and a generated explanation. If extraction gets a sentence wrong, I need a route back to the source. If an explanation draws an inference, I need to see where the source stops and the inference begins.

This does not require covering every answer in cryptographic vocabulary. Most of the time, the useful experience is ordinary: open the supporting passage, see which version it came from, and check whether it says what the answer claims.

The model can then do what makes it valuable. It can propose a reading, notice a contradiction, or help formulate a better question. A later model can offer a different interpretation of the same preserved material. We can compare those interpretations without having to wonder whether the underlying record changed beneath them.

That is the kind of personal search I want to trust with my own history. Let the answer improve. Keep the evidence available for the argument.

From a September 20, 2026 scope discussion. Published September 28, 2026.