Illustrate prescriptive sensitivity and conflict analysis in templates

[chriscoey]

What

Extends three existing prescriptive-optimization templates to demonstrate the new sensitivity (LP marginals) and conflict / IIS (infeasibility diagnosis) analysis on solve():

• factory_production — a maximize-profit product-mix LP. One solve(sensitivity=True) reads each factory's capacity shadow price and each product's reduced cost + basis status, joined back to its entity by key. A binding factory vs. a slack one shows the zero/nonzero shadow-price contrast, and the maximize objective mirrors the sign conventions of the minimize example below.
• supplier_reliability — a minimize-cost sourcing LP. Reads supplier-capacity and product-demand shadow prices plus supply-lane reduced costs / basis status, then re-solves supplier-disruption scenarios.
• cicd_runner_allocation — a binary assignment MIP. When a maintenance outage makes the schedule infeasible, solve(conflict=True) returns the conflict set (IIS): the stranded jobs and the binding runner cap, read back by entity key.

Each marginal or conflict joins to its grounding entity by key through the constraint's declared back-pointer (keyed_by={...} on satisfy), mirroring the automatic variable back-pointers. Each README walks through the model, the marginals/conflict, and how to act on them. All three runners were validated end-to-end against the solver, including assertion-level checks on objectives, marginal signs, basis statuses, and IIS membership.

Merge gating

These examples exercise the sensitivity and conflict APIs added in the upcoming PyRel release (RelationalAI/PyRel#1617), so the runners only work against a build that includes that API. Before merge, once that release ships: raise each template's relationalai dependency floor to the first version that includes sensitivity/conflict (the pins still say 1.0.14, which predates the API), and re-run the three templates against the released build.

🤖 Generated with Claude Code

[github-actions]

The docs preview for this pull request has been deployed to Vercel!

✅ Preview:	https://relationalai-docs-tkoj4zdke-relationalai.vercel.app/build/templates
🔍 Inspect:	https://vercel.com/relationalai/relationalai-docs/6KTAcrgtwokoJNNWYKLmCCG3tMbK

[cafzal]

Review — sensitivity + conflict template examples

Re-reviewed against the upstream feature PyRel #1617, now merged to main (b120f95c3). All three scripts py_compile clean and ruff passes; frontmatter vocab is valid. API usage matches the merged source — solve(sensitivity=True) with an objective present, solve(conflict=True) on the MIP, keyed_by declared on satisfy(), attribute names correct, no conflict_message. The factory (maximize) / supplier (minimize) sign-convention contrast is correct, and the cicd outage is provably infeasible against the sample CSVs (its assert guards hold).

Required after release (not a merge blocker)

relationalai==1.0.14 predates the sensitivity/conflict API. Once the release that ships #1617 is out, bump the dependency floor at all six spots, then re-run all three templates against the released build:

• v1/factory_production/pyproject.toml:12, v1/supplier_reliability/pyproject.toml:12, v1/cicd_runner_allocation/pyproject.toml:12 (the relationalai==1.0.14 pins)
• README prerequisite lines: factory_production/README.md:68, supplier_reliability/README.md:62, cicd_runner_allocation/README.md:59 (the >= 1.0.14 lines)

Exact version is not knowable yet — #1617 is on the repo's 1.8.x line, the customer SDK line is 1.0.x, and the feature also gates on the rai-solver-service #246 producer deploy. Until pinned, the READMEs' expected-output blocks remain static rather than live-captured.

Minor (non-blocking)

• README "How it works" snippets show .inspect() where the scripts use .to_df() / print (.inspect() appears in no script). Make verbatim or mark illustrative.
• supplier_reliability industry: "Supply Chain" is off the dominant portfolio convention "Supply Chain & Logistics" (pre-existing; the PR didn't touch it).

[chriscoey]

All the post-release items are in. The dependency floor is bumped at all six spots: the three pyproject pins are relationalai==1.9.0 and the three README prerequisite lines say >= 1.9.0, with 1.9.0 being the release that ships the sensitivity/conflict API. The two minor items are also addressed: each README readback snippet now notes that .inspect() is the quick interactive form while the scripts materialize the same selects with .to_df(), and supplier_reliability's industry tag is aligned to "Supply Chain & Logistics".

I re-ran all three templates end-to-end against the released build: each exits 0 with all assertion guards holding. The README expected-output blocks are now live-captured from those runs. Statuses, objectives, shadow prices, reduced costs, and the conflict membership matched the previous blocks exactly; the rows the READMEs hedge as tied-optima-dependent (order splits, basis statuses, the equal-cost ubuntu job split) were updated to the captured solutions.