A software project estimation methodology for the AI-assisted era. It replaces independence of experts with independence of techniques.
Wideband Delphi works because experts are independent. An LLM has no such independence: a single model asked several times gives not independent opinions but well-anchored guesses (rephrasing the question shifts its estimate more than it would a human's). So the source of diversity is not different "roles" of one model, but different estimation methods.
The same project is estimated with several independent techniques (WBS decomposition + PERT, analogy / reference class, throughput Monte Carlo, parametric models). Each technique's blind spots are guaranteed by construction, not random. When they disagree, the gap is not averaged away but diagnosed through each method's known blind spot — for example, a WBS leaf-sum below the analogy => the charge for the seams between tasks was left out. We calibrate one method with another. The output is a range with an explained residual, not a single number.
METHODOLOGY.md— the framework: methods, their blind spots, pipeline, boundaries.findings.md— working log: why the methodology is built the way it is (sources, experiments).examples/— worked cases:BMS,FaxRxTx(blind estimation).mars_model/— hypothesis checks on open datasets (PROMISE);results.md+ fitting code.pitch_en.md— a short project pitch.
Early stage. The methodology is worked out on two cases (n=2); the empirical part
runs on open data and is reproducible (data + code in the repository). The claims are
deliberately modest: this is a skeleton with checkable results.
- Validation on more real projects with known actuals.
- Agent wrapper: automating the pipeline as isolated agents, one per method — deliberately not one model in several "roles".