replication package

Dataset DOI 10.5522/04/21308124,<br>Apache License 2.0. 458,232 issues, 39 projects, 12 public Jira repositories.<br><br>The dataset ships as a MySQL dump. `extract.py` stream-parses it directly and requires no<br>database server.<br><br>### Pipeline<br><br>```<br>python3 extract.py # TAWOS.sql -> issues.csv (458,232 rows)<br>python3 p2\_analysis.py primary # total-effort outcome, revised estimates excluded<br>python3 p2\_analysis.py sens\_resolution # resolution-time sensitivity<br>python3 p2\_conformal.py # conformal vs quantile interval calibration<br>python3 delegation\_rule.py # regenerates Table 3, Table 4 and Table S6<br>python3 figs.py # regenerates Figures 1, 2, 3 and S1<br>```<br><br>### Analysed sample<br><br>|Specification|Projects|Predictions|<br>|-|-|-|<br>|primary (Total\_Effort\_Minutes)|25|31,356|<br>|sensitivity (Resolution\_Time\_Minutes)|32|43,536|<br><br>Inclusion: story point > 0; outcome > 0; story point not revised after estimation;<br>at least 150 prior issues before the first prediction; at least 60 evaluable predictions<br>per project.<br><br>### Method<br><br>Error is taken in log space, e = ln(actual) − ln(predicted), giving the exact decomposition<br>MSLE = mean(e)² + var(e) into a bias and a dispersion term.

Three forecasters are evaluated<br>strictly out of sample under a time-ordered walk-forward within each project:<br><br>* **H** human — story point × historical minutes-per-point rate (velocity planning)<br>* **R** reference class — historical median duration of prior same-type issues<br>* **M** model — gradient-boosted quantile regression at the 0.10 / 0.50 / 0.90 quantiles<br><br>Inference across projects uses Wilcoxon signed-rank tests, 10,000-sample bootstrap<br>confidence intervals, and matched-pairs rank-biserial effect sizes.

The unit of analysis is<br>the project (n = 25), not the issue.<br><br>### Known limitations<br><br>* The outcome is a **duration proxy** derived from issue state transitions, not logged<br>effort. Only 900 issues in the corpus carry logged work, too few to analyse.<br>* The corpus contains **none of the Layer 1 telemetry** the framework specifies — no CI<br>outcomes, work-in-progress levels, dependency depth, review latency or team composition<br>history.

The P2b test is therefore a joint test of P1 and P2b, and a null is what P1<br>predicts under thin telemetry.<br>* The model arm is one model class with one feature set.<br><br>### Environment<br><br>Python 3.13 · numpy 2.4 · pandas 3.0 · scipy 1.17 · scikit-learn 1.8 · matplotlib.<br>Random seed 20260921 for all bootstrap resampling.<br><br>### Figure palette<br><br>The two-colour categorical palette (#2E6FA8, #C2681A) was validated for colour-vision<br>deficiency separation, lightness band, chroma floor and surface contrast before use.<br><br><br><br></pre><p></p>