[MaxGhenis]

Max said go (d2146, 28 Oct): merge on the gates below. This changes a published page. The explainer post went live with the matching note at 20:52Z (PolicyEngine/policyengine-app-v2#1473, d2147), so until this merges the dashboard still shows the average the post has withdrawn.

The binding C2 score on post-Budget data (#45, docs/C2_BINDING.md) failed: test A's floor-frequency bias is 0.267 against a limit of 0.25. The floor frequency is the share of years with earnings growth strictly below 2.5% (ts_backtest.py, METHOD); the model predicted 44.6% against 17.9% realised. Test A gained a seventh forecast (March 2022) once September's CPI completed its window; the dry run had six. Under Max's d955(c), the expected value then falls back automatically to (a): scenarios only, no expected value. The rebuild on the Budget's inputs, booked from 29 October, writes that file itself, but takes about a day of Microcosm compute (REBUILD.md's runtime paragraph). Until it lands, the live dashboard is the provisional pre-Budget (c) release. That release leads with the expected saving and its middle-80% range.

The explainer post is getting the matching note in PolicyEngine/policyengine-app-v2#1473. This PR's notice uses the same facts, order and wording, except two deliberate differences: the notice quotes no withdrawn figure, and the model's paths stay as labelled illustrations, which d955(a) lists (the post drops them).

What changes on the page

• A dated notice at the top: "Updated 28 October 2026". It says what the pre-Budget release showed, that the model failed a test we had registered before the Budget (linking Binding C2 re-score on October 2026 Budget inputs #45), and what was withdrawn. It adds what the pre-Budget release uses (forecast, PolicyEngine UK version and survey, read from the file) and what the rebuild uses: PolicyEngine UK 2.151.0, the Microcosm UK dataset (engine.py's binding bundle, d2011), and the OBR's October 2026 forecast to 2030 with its March 2026 long-term assumptions after (d2124).

• No expected value anywhere. withBindingOutcome() (dataHelpers.js) takes the provisional (c) file and:

  • drops expected_value;
  • records the (c)→(a) fallback the way the binding build does (uncertainty_ruling, expected_value_omission.reason);
  • leaves every other block identical (tested by reference).

  The page then uses its existing, tested scenarios-only rendering:

  • the scenario envelope is the headline: central forecast, OBR long-term path, the ±0.5pt earnings mean-path scenarios and the last-decade replay, all labelled "not a probability interval";
  • the expected saving, its middle-80% range, the share-losing average and the histograms are gone.

• Path names. Without an expected value, the model's paths are scenarios (d955(a): "model paths labelled as scenarios"). They become:

  • "Model path: a random draw";
  • "Model path: a middle-sized gap";
  • "Model path: a large gap".

  "Typical", "high-saving" and the per-path percentile are gone, and the paths tab says these are illustrations, not a typical or likely path. The names follow the expected value alone (pathsAreScenarios), whatever the ruling, so they also apply to the rebuild's scenarios-only file.

• No typical path or average anywhere. The Distributional impact tab's poverty card no longer gives "the typical path" as context, and its footer no longer points to an average over all paths. The "What each path is" list no longer places the scenario paths at the median, the 90th percentile or a percentile band.

• Methodology tab:

  • it shows the omission reason;
  • it labels the C2 scores it shows (from the file) as the dry run's, adding that, rescored on 28 October with a seventh past forecast, the model failed.

Any other file, a sample included, passes through unchanged. Remove BINDING_C2_OUTCOME, withBindingOutcome and BindingOutcomeNotice with the rebuild, whose file records the outcome itself.

Checks

• New BindingOutcome.test.jsx (13 tests):
  • passthrough for passing and failing binding files, a file with no expected value, a sample and a provisional sample;
  • path names that follow the expected value alone, even under ruling (c);
  • every tab's whole text, collapsed <details> footers included, scanned for typical, percentile, median and average-over-paths wording. Only four phrases are allowed: the envelope's earnings scenarios, the past-years check score, and the two lines saying no path is typical and no average is shown;
  • the transform's exact changes, with every other block kept by reference;
  • idempotence and no mutation of the committed object;
  • scenario path names;
  • on every tab: the notice once, no withdrawn figure, no "provisional until", no typical or percentile path names, and nothing broken or "unavailable";
  • the "What each path is" list on every tab uses the scenario names;
  • the reason states the statistic as earnings growth below 2.5%, never "the floor setting the rise", and its numbers are consistent (bias = predicted − realised, above the limit);
  • the notice's inputs and rebuild sentences, and no "Budget's figures";
  • the method tab's omission, dry-run and rescore labels;
  • no notice on the committed file itself or on a sample.
• Four existing tests in Steps.test.jsx now follow the file's path names, and whether it has an expected value, instead of hard-coding "90th percentile" or the typical-path context. One new test checks that a file without an expected value names no typical path.
• Test runs:
  • bun run test: 188 passed, 4 skipped.
  • TRIPLE_LOCK_RESULT_FIXTURE=c2-pass: 174 passed.
  • TRIPLE_LOCK_RESULT_FIXTURE=c2-fail: 1 failure plus the ruling-(b) suite. Both already fail on main in that mode and are untouched here; they belong to the rebuild.
• Mutation check. Each of these makes the new tests fail: the old floor wording, an inconsistent bias, "Budget's figures" in the notice, a hard-coded random path name, a dropped model version, the old method-tab line, a transformed sample, the typical-path context, the footer's average, the percentile footer, and path names that require ruling (a).
• Review. Sol at 588b696 requested changes on five points. 101f3ba fixed the floor wording; 563db6f fixed the other four. A re-review at 563db6f is running (20261028-172859-pr48-review-r2).
• Build and lint. bun run lint shows the same 2 warnings as main. bun run build passes.
• Browser check. I checked the page in a local next start of this branch: the notice, the Budget-impact scenario envelope, the paths tab names and the methodology text. There were no console errors.

Merge gates

1. Binding C2 re-score on October 2026 Budget inputs #45 merged: the notice states its result.
2. An independent review approves at the head.
3. CI green. Max's go is given (d2146).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
uk-triple-lock	Ready	Preview	Oct 28, 2026 9:29pm UTC

[MaxGhenis]

Merge audit for head 563db6f99c941d94079b8d6c195c6d2e32c773f5 (base main at b4c56c2, after #45 merged).

• Decision: d2146, decided by Max on 28 October (~21:00Z): merge once its Sol reviews approve, CI is green and Binding C2 re-score on October 2026 Budget inputs #45 has merged. Binding C2 re-score on October 2026 Budget inputs #45 merged at 22:22Z (b4c56c2).
• Reviews:
  • GPT-6.1 Sol review at 588b696 asked for changes (five blockers); 563db6f fixes them.
  • GPT-6.1 Sol r2 at 563db6f: APPROVE, job 20261028-172859-pr48-review-r2, verdict at ~/reviews/uk-triple-lock-2026-09-29/release-night/pr48/review-sol-r2.md. Its one non-blocking note is the allowlist granularity in BindingOutcome.test.jsx. That code goes away when Binding rebuild on the October 2026 Budget inputs (scenarios only) #49 deletes the withdrawal notice after the binding rebuild, so it isn't changed here.
• CI: on the merge with the new main, gh pr checks 48 exits 0. Pipeline run 40780494920 passed (31 min) and the Dashboard run 40780494918 passed.
• mergeable is MERGEABLE; there are no CHANGES_REQUESTED reviews; the PR isn't a draft.

Merged with --match-head-commit 563db6f.