[nick-developer]

Requested by Cod3r · project thread

What and why

Roadmap T07b (wave D plan, section 5). T07's DP mode was nearly flat in ε because stage 1 only released one-way histograms and one Gaussian correlation matrix per table. This adds synthesis.privacy.dp.measure: mst | aim (relsynth run --dp-measure). The default stays copula, so nothing changes for existing specs.

• Selection under the same zCDP accountant: MST (maximum spanning tree on the L1 error of the independent model, exponential mechanism edge by edge) and AIM-style adaptive rounds with annealing, kept to a forest so inference is exact. Selection is OpenDP's report noisy max (Gumbel) with its zCDP map (mechanisms.noisy_max).
• Nodes of a table: its attributes, its truncated degrees and, for a child table, the parent row's cells (the cross-table link). Domains are compressed to pair_bins cells (default 8) from released marginals only.
• Fit: Private-PGM for a tree (inverse-variance node estimates, IPF edge joints); seed rows are drawn from the forest, child rows given parent cells by belief propagation. Stage 2 is unchanged.
• Proofs: Props. 1300–1306 (docs/proofs/sec_v26_dpselect.tex); guide docs/dp.md section 7; CHANGELOG and README updated.
• Fixes found by the sweeps: an empty seed table kept no columns in check enforcement; the scorecard's text naive Bayes failed on mixed str/null labels and allocated 32 GB for a 66k-level condition (now an inverted index above 50 M cells; smaller tables unchanged, unit-tested).
• Includes c011193 (the types fix from ci(service): run the service end to end on kind and docker compose; fix two Helm chart bugs (T09 follow-up) #44, commit 4e3026f) so the mypy job passes before ci(service): run the service end to end on kind and docker compose; fix two Helm chart bugs (T09 follow-up) #44 lands.
• dp-eval workflow: smartnoise-synth 1.0.8 MST/AIM comparison on ACS, ε sweeps on Chinook, Contoso and AdventureWorks, attack audit (tools/dp_audit.py). It now runs on PR open/ready, weekly and on demand, not on every push.

Results (dp-eval run 42141805823, three repetitions, 95 % t intervals)

• ACS, mean two-way TVD: copula 0.078 → 0.070, aim 0.062 → 0.053, mst 0.063 → 0.055 (ε 0.5 → 8). smartnoise MST 0.108 → 0.090; smartnoise AIM 0.074 (ε 0.5), 0.054 (ε 1), and out of memory on the runner at ε 4 and 8.
• Scorecard share: AdventureWorks aim 0.913 → 0.923 (copula 0.903 → 0.915); Chinook aim 0.883 → 0.912; Contoso aim 0.805 → 0.814, but the copula is better at ε 8 (0.839). Intervals overlap on all three.
• Attack audit, Contoso ε 1: membership TPR at 1 % FPR 0.011 (mst) and 0.012 (aim), bound 0.027; no check fails.

How it was tested

• ruff check src tests tools passes
• C1300–C1306 (pytest -m science tests/science/test_dp_select_checks.py): 9 passed. C1301 replays the noisy outputs and choices into neighbouring databases: every score vector and two-way histogram moves within its charged sensitivity.
• tests/unit/test_text_nb.py: the sparse naive Bayes path gives the dense path's predictions
• Full CI gate (ci on ad69d78 green)
• dp-eval results (above; report roadmap/reports/T07b.md in the project files)

Statistics and privacy

• Each new claim has a proposition, proof and check
• Default output unchanged (measure: copula is the default)
• Claims register entries DP-10 to DP-14 for docs/dp.md section 7 are drafted; they go into docs/assurance/claims.yaml once feat(assurance): claim register, assurance case and regulation mapping (T17) #41 is on main

Checklist

• Commits signed off
• CHANGELOG, README, docs
• No data committed
• No new runtime dependency (smartnoise-synth only in a CI job's own venv)

---

Generated by Claude Code

🤖 Generated with Claude Code

https://claude.ai/code/session_01PSVqwDw4YbuLAH7eDJyBvk