Research

Why systems that handle every individual task well still drift away from the project they serve.

Problem

A long-running system can be correct at every step and wrong overall. Each task is handled competently; the direction of the work still degrades.

Why it matters

Most evaluation measures single-turn or single-task quality. The failure that costs real projects the most time is not a wrong answer, it is a slow loss of the thread across many correct answers.

Observation

Observed repeatedly while building and operating long-running systems on top of language models. The models were not the weak point; the absence of anything carrying intention forward was.

Hypothesis

Continuity is a property of the surrounding system, not of the model. It requires durable state that outlives any single conversation.

Public material

Described here at the level of the problem. Essays in preparation.

A governing discipline for keeping long-running work anchored to evidence and stated intention.

Problem

Confidence and correctness are not the same signal, and nothing in a model's output distinguishes them. Over a long project, unsupported assertions accumulate and become the basis for later work.

Why it matters

Once an unsupported claim is treated as settled, every decision built on it inherits the error, and the cost of finding it grows with distance from where it was introduced.

Observation

Assertions that were never verified tend to be repeated back later with more confidence than when they were first made.

Hypothesis

Requiring evidence before an assertion can be recorded as settled, and making that requirement mechanical rather than advisory, reduces the rate at which unsupported claims become project state.

Public material

Problem statement only. Implementation detail is intentionally withheld pending intellectual-property review.

Durable, append-only project memory: what was decided, why, on what evidence, and what remains open.

Problem

Chat history is a transcript. It records what was said, not what was decided, and it does not distinguish a verified fact from a guess made in passing.

Why it matters

A large context window is not the same as reliable state. Reading everything is not the same as knowing what is true.

Observation

Project knowledge that lives only in conversation is lost at the boundary of that conversation, and is silently reconstructed, often differently, in the next one.

Hypothesis

Project memory should be an append-only record with provenance attached to each entry, separate from the conversation, so that state can be queried rather than recalled.

Public material

Conceptual model published here. Schema and implementation are not public.

How autonomous agents diverge from their instructions, and what it would take to notice early.

Problem

An agent can report success for work it did not do, or did against the wrong target, and the report is indistinguishable from a true one without independent checking.

Why it matters

Trusting an agent's own account of its work removes the only step that would have caught the error.

Observation

A tool reporting success is evidence that the tool ran, not evidence that the intended change happened.

Hypothesis

Verification has to be performed against the artefact rather than the report, and it has to be mechanical, because a reviewer under time pressure will accept a plausible report.

Public material

Concept stage. No results to report.

The displacement question, treated as a real one rather than a footnote.

Problem

Automation is removing categories of work faster than it is creating accessible replacements for the people who did that work.

Why it matters

The people most affected are least positioned to absorb the transition, and the timeline is set by capability rather than by readiness.

Observation

This is an opinion informed by building automation commercially, not a research finding.

Hypothesis

None offered. This is stated as an open concern rather than a thesis.

Public material

Opinion. Explicitly not a research claim.