Observable · Decomposable · Measurable · Controllable — an application of "holistic performance analysis" carried over from performance coaching
A descriptive practitioner note by Andreas Ehstand, independent researcher (ORCID 0009-0006-3773-7796). A way of thinking, not a validated instrument or a standard.
Scope note. This is a descriptive framework — a structured way to look at your own working process with AI. It is not professional, legal or medical advice; not a test, score, or instrument for assessing, ranking or screening other people; and not empirically validated — the predictions it makes are stated below as open, untested hypotheses. Use it as a lens, not as a guarantee.
A way to make the human side of working with AI improvable, by passing one's own process through four sequential questions: is it observable, can it be decomposed, can it be measured honestly, and which part is actually within my control to change?
The building blocks are old and well established. This framework does not claim to invent any of them:
The contribution is deliberately narrow — the transfer and its operationalisation, not the parts: it takes the human–AI collaboration itself as the unit of analysis; it forces a human-vs-tool factor split that counters the common misattribution of blaming the model for a human-input bottleneck; it builds in measurement honesty (every indicator notes what it does not measure, against Goodhart's law); and it keeps the lever on the controllable human side. Everything else is borrowed and acknowledged.
1 — Observable. What of the collaboration is made visible — and what evaporates? Make the interaction recordable instead of letting it vanish. Typical mistake: looking only at the result, never the path.
2 — Decomposable. Which separate factors make up the performance? Break the whole into 3–7 named sub-steps (framing × giving context × checking × correcting × integrating). Typical mistake: treating it as one indivisible block.
3 — Measurable (honestly). How does a factor become a number without distorting the goal? One honest indicator per sub-step, and a note of what it does not capture. Typical mistake: confusing the easily-counted with the important.
4 — Controllable (by me). Which lever can I myself change reproducibly — and which is outside my control? This lens is about your own agency over your own process — not about controlling other people. Typical mistake: spending energy on the non-controllable (waiting for "a better model") instead of the controllable (your own routines and checks).
A writer uses an AI assistant and concludes "this tool is unreliable." Running the four lenses, they log where corrections happen, split the work into stages, count corrections per stage, and find the count concentrates in one stage — context-giving — which is fully within their control. The upshot: the felt experience misattributed a human-input bottleneck to the tool. The lens relocates the improvement lever from the uncontrollable (waiting for a better model) to the controllable (a reusable context template). This is a single illustrative case, not evidence — on testing it might prove no better than ordinary reflection; §6 states how it could be refuted.
Observe → Decompose → Measure → Control → Loop. Did the indicator move? If not, it was the wrong factor (back to Decompose). The value is meant to be in repeating the loop, not in one pass — though whether it outperforms ordinary trial-and-error is an open question (§6), not an established result.
Stated so the framework can be wrong. None has been run; they are a research agenda, not evidence.
It does not apply to purely subjective, non-decomposable experiential quality.
A self-analysis lens for the person doing the work. Not an instrument to test, score, rank or screen other people; not safety-critical, legal, or medical guidance; descriptive, not a guarantee of outcomes.
A transfer from high-performance coaching, where one decomposes performance into separately trainable factors rather than judging it "good or bad." The contribution is this transfer, not the invention of factor analysis.
Cited descriptively; this framework credits these works and does not claim their authority.
Author: Andreas Ehstand, independent researcher (ORCID 0009-0006-3773-7796). Licence: CC BY-NC-ND 4.0. Descriptive, evolving document; not legal, medical or professional advice. Permanent identifier: doi:10.5281/zenodo.20613914