napkin-math(bounds): Phase 4 prompt-side LLM rules (R1.2, R1.5, R2.2, R1.3)

[neoneye]

Summary

Completes Phase 4 of experiments/napkin_math/docs/20260520_plan.md. PR #747 shipped the code-side runtime + schema readiness; this PR ships the four LLM-rule changes that drive new behaviour on top of that runtime.

Prompt edits (generate-bounds/system-prompt.txt)

1. ACTUAL-VS-COMMITMENT — source-tag asymmetry (R1.2). When base is shifted on a named plan-internal anchor (Premortem entry, Risk N, Issue N, Decision N, expert-criticism passage forecasting a gap), source: "data" and the rationale names the artifact verbatim. When base is left at the commitment default with no named anchor, source: "assumption" (NOT "data") and the rationale states explicitly that no plan passage anchors a shift. The asymmetry must be visible in the rationale so a downstream reader can tell at a glance whether the base reflects a plan finding or a modelling default.
2. New SELF-AUDIT: CITATION CONTEXT-LEAK section (R1.5). Concrete abstract examples of the failure mode where a citation is lexically present (the plan has a Risk N) but substantively wrong (Risk N is about a different topic). Drop the citation when it fails the substantive-support check; re-evaluate whether the base/range shift is justified at all if no substantively correct anchor exists.
3. New DISTRIBUTION DEFAULT BY PLAN SCALE section (R2.2). For plans at megaproject scale, CAPEX and major OPEX variables default to sampling_discipline: "lognormal" with low/high as P5/P95 (Flyvbjerg's iron law). Identification by abstract criteria — multi-billion budget, multi-year horizon, multiple binding regulatory dependencies, first-of-kind execution, structural cost-overrun framing in Premortem — NOT by a plan_type-string enum. plan_type is one signal among several; treating it as the sole criterion would encode corpus literals into the prompt.
4. New CORRELATIONS (OPTIONAL TOP-LEVEL BLOCK) section (R1.3). Schema definition + selection rules for emitting the correlations key reserved in PR napkin-math(bounds): Phase 4 runtime + schema readiness #747. A group is declared only when the plan ITSELF names a shared driver between two or more bounded variables. Modeller priors are not valid anchors. Rho is bucketed by anchor strength (0.6-0.8 strong, 0.3-0.5 moderate, weak couplings stay independent). Max 5 groups.

Discipline-table updates

PR #747's "do not emit yet" guards on lognormal and the correlations top-level key are lifted: both now point to the new selection-rule sections. pert remains "no prompt rule currently directs you to emit" since this PR ships no pert selection rule (it stays schema-reserved).

Phase-8 implementation status is documented on both:

• lognormal — NotImplementedError at sample time (loud failure; no silent fall-back to triangular)
• correlations — preserved through stripper + warning emitted at run time (no silent shim; joint-tail risk understatement is surfaced to the user)

Runner edit (run_monte_carlo.py)

Adds a loud warning when bounds declares a correlations block but the Gaussian-copula sampler is not yet implemented. Names the block size, names Phase 8 explicitly, and states that joint-tail risk is structurally understated until the sampler ships.

Anti-overfitting verification

grep verifies that none of the actual v51 plan_type strings — hyperscale_infrastructure, industrial_automation_pilot, catastrophic_disaster_response, public_finance_transition, commercial_csr_reverse_logistics, commercial_digital_infrastructure_application — appear in the prompt text. No corpus-specific identifiers (OPC UA, $75k, 32.2 km, plan slugs) appear either. The rules read identically whether the input is a renewable-energy plan, a public-benefit policy, a renovation project, or a language no one on the team reads.

Empirical posture

• Unit tests: 75 pass in tests/test_run_monte_carlo.py and tests/test_strip_threshold_bounds.py. 2 new tests cover the correlations-block warning emission and absence.
• Smoke suite: 9/9 checks pass.
• Behavioural verification of the new prompt rules against the napkin_math probe set would be a same-LLM same-session regression check, not an improvement claim. Honest verification needs a different-LLM run, which is a separate follow-up.

Test plan

• pytest experiments/napkin_math/tests/test_run_monte_carlo.py experiments/napkin_math/tests/test_strip_threshold_bounds.py (75 pass)
• python3 experiments/napkin_math/tests/run_smoke.py (9/9 pass)
• No corpus literals in the prompt (grep verified)
• CI green on this branch

🤖 Generated with Claude Code

[neoneye]

Addressed review feedback (c8d37394). Three prompt contradictions fixed:

1. Worked example flipped to source: "assumption" for both actual_outreach_contact_rate and actual_cooling_center_utilization. Both have base at the commitment default with no plan-internal anchor for a base shift, so per the new R1.2 rule they must be tagged assumption. For actual_cooling_center_utilization the rationale now explicitly explains why source stays assumption despite the Risk-4 spread anchor — the source tag tracks BASE anchoring, not spread anchoring.
2. Removed the stale "which is not yet emitted" phrase from the correlations exception in the output rules. This PR's CORRELATIONS section explicitly directs emission when rules match.
3. Generic source rule defers to the actual_X exception. Added an "Exception: actual_X variables follow a different source-tag rule" paragraph in HOW TO CHOOSE THE RANGE pointing to the ACTUAL-VS-COMMITMENT section. Also split the corresponding bullet in "Rules for the output" into two — one for non-actual_X (range-anchoring controls source) and one for actual_X (base-anchoring controls source).

75 unit tests pass, 9/9 smoke checks pass.