# Phase 4 — Structural change, total monthly series (corpus v2, month-precision records, 2018-02 → 2026-06)

Source: `results/econometrics/its_results.json`, `fit_total.csv`, `placebo_total.csv` (script `econometrics/run_structural_change.py`). Interpretation pending PI checkpoint H-4. The 2022-11-30 date is a reference point; nothing below is a causal estimate.

## Interrupted time series at 2022-12 (log counts, month-of-year dummies, Newey–West 6 lags)
| Quantity | Estimate |
|---|---|
| Pre-period monthly growth (slope) | +0.0031 log points/month (SE 0.0005) ≈ 0.31 %/month |
| Level change at 2022-12 | -0.009 (SE 0.024, t = -0.40, p = 0.69) |
| Slope change after 2022-12 | +0.0108/month (SE 0.0009, t = 12.1, p = 6.4e-34) → post-period growth ≈ 1.39 %/month |
| Placebo (all 65 candidate months) | level: rank 63/65, p = 0.97; slope: rank 1/65, p = 0.015 |
| Placebo (baseline months only, n = 34) | level p = 1.0; slope p = 0.0 |
| Donut (±3 months excluded) | level +0.049 (p = 0.056); slope +0.0110 (p = 6.7e-24) |
| R² | 0.966 (n = 101) |

## Alternative event dates (same model)
| Event | Level change | Slope change |
|---|---|---|
| chatgpt (2022-12-01) | -0.009 (p = 0.690) | +0.0108 (p = 6.42e-34) |
| gpt3 (2020-06-01) | -0.091 (p = 0.122) | +0.0059 (p = 8.44e-05) |
| dalle2 (2022-05-01) | -0.075 (p = 0.006) | +0.0100 (p = 1.22e-28) |
| gpt4 (2023-04-01) | +0.025 (p = 0.305) | +0.0109 (p = 6.68e-30) |
| o1 (2024-10-01) | +0.103 (p = 0.008) | +0.0115 (p = 2.91e-10) |

## Unknown-break methods (seasonally adjusted log series)
- Bai–Perron (BIC and LWZ agree): break date(s) ['2022-04-01'] ; BIC by m = {'0': -463.8, '1': -555.25, '2': -551.53, '3': -545.89}
- PELT: ['2023-05', '2024-12']
- BOCPD MAP run-length resets: []
- Sensitivity to the last complete month: {k: (round(v['its_chatgpt']['level_change'],3), round(v['its_chatgpt']['slope_change'],4), v['bai_perron']) for k, v in r['sensitivity_last_complete_total'].items()}

## Reading (preliminary, H-4 pending)
The total output series shows no discrete level shift at the ChatGPT release month but a marked and robust acceleration of the growth rate (from ≈ 0.3 % to ≈ 1.4 % per month), which the unknown-break estimators date to spring 2022, i.e. before November 2022. The acceleration therefore predates ChatGPT in the aggregate series; category-level series (after classification) will show where it is concentrated. Caveats: right-truncation (series cut at 2026-06), OpenAlex coverage growth, and the exclusion of year-only dated records.