Methodology review fixes: England-only CT rebate, log-linear behavioural form, decile-specific elasticities by MaxGhenis · Pull Request #1 · PolicyEngine/energy-price-shock

MaxGhenis · 2026-04-20T13:34:29Z

Why

Methodology pass over the energy_shock package and dashboard.

Note: an earlier version of this branch attempted to migrate the
simulation layer to the unified policyengine Python API; that
change was reverted in 901b654 because policyengine 4.1.x's
bundled UK release manifest was pinned to a data-package version
without a published release_manifest.json on Hugging Face, blocking
bootstrap. The final diff retains policyengine-uk directly. The
migration can be revisited once upstream stabilises.

What shipped

Methodology

Decile-specific price elasticities. Each household's behavioural response is computed at its own income decile's short-run elasticity per Priesmann & Praktiknjo (2025), interpolated linearly from −0.64 (D1) to −0.11 (D10). A population-mean elasticity averages away the progressivity that matters.
Log-linear behavioural form. Spend response uses (p_new / p_old) ** (1 + ε) rather than the linear first-order approximation (1 + p)(1 + εp), which produces negative consumption — physically impossible — for combinations like ε = −0.64 and the +161% Q1 2023 peak. The log-linear form stays admissible at all ε ∈ (−1, 0] and p ≥ 0. Transferability and extrapolation caveats documented in README.md.
England-only CT rebate. ebr_council_tax_rebate in policyengine-uk pays any A–D band household including Scottish/Welsh/NI ones; the 2022 policy was England-only. Gated at the analysis layer in policy_ct_rebate, policy_post_shock, and _grouped_post_policy so aggregates match the real policy's geographic scope.

Code

Dead policy functions removed (in a follow-up commit after review): policy_epg, policy_wfa, policy_combined, policy_net_position were fully implemented but never called from generate.py. Dropped along with the now-unused EPG_TARGET / WFA_HIGHER / WFA_LOWER config constants. The dashboard consumes only policies.flat_transfer, policies.ct_rebate, and policy_post_shock.
Vectorised _build_household_type — old per-household Python loop replaced with pandas groupby + boolean masks.
Deps pinned in pyproject.toml with lower bounds: policyengine-uk>=2.88.0, microdf-python>=1.2.0, pandas>=2.0, numpy>=1.26.
CI workflow added (.github/workflows/pull_request.yaml) — runs pytest tests/. The unit suite is pure-numerical (no microsim bootstrap, no HF token required).
Unit tests for the decile-specific elasticity helpers in tests/test_elasticity.py: Priesmann endpoints, per-household decile lookup, missing-decile fallback, zero-price-change identity, low-decile cuts harder at +60 %, physical admissibility at +161 %.
Lean sdist — [tool.hatch.build.targets.sdist] ships only the Python package, tests, README, LICENSE, and pyproject.toml; the ~10 MB dashboard/ subtree is excluded.

Dashboard

Dashboard.jsx prose / UI tweaks and regenerated results*.json / results_breakdowns*.json for UK + England / Scotland / Wales / Northern Ireland. Old results_v2*.json artefacts dropped.

Not changed

np.maximum(shocked - payment, baseline_energy) floor semantics kept; it matches the 'extra cost from shock' question.
EPG linear scaling across scenarios not reintroduced; can be added in a later PR if we want per-scenario PE reform runs.

Test plan

pytest tests/ passes on the decile-elasticity suite
python -m energy_shock --all-countries runs end-to-end (requires HUGGING_FACE_TOKEN)
Spot-check UK aggregate: baseline.mean_energy_spend close to prior value
Spot-check CT-rebate panels: Scotland / Wales / NI average benefit is £0; England > £0
Dashboard loads all five country JSONs

Generated with Claude Code

## Migration Replaces raw `from policyengine_uk import Microsimulation` calls with the unified `policyengine.py` API: - `baseline._build_simulation` constructs a `policyengine.core.Simulation` with `tax_benefit_model_version=uk_latest` and passes `extra_variables` so the energy-specific columns (electricity_consumption, gas_consumption, region, accommodation_type) plus all policy reform outputs (epg_subsidy, ebr_*, winter_fuel_allowance) are materialised on `output_dataset.data`. - `ensure_datasets` handles HF download + uprating to the target year, removing the manual URL passing. - `_hh_array` replaces `sim.calculate(var, YEAR).values` across `sections.py` — accesses the output frame directly. - Reform dicts in sections.py now build via `build_reform_simulation`; parameter paths are unchanged. ## Review fixes Addresses findings from the methodology review: 1. `rising_block_tariff` post-shock was returning `{"deciles": []}` stubs. Finished the implementation — surcharge rate is now recomputed under each scenario so cost-neutrality is enforced at shock prices, and per-decile extra-cost rows are populated. 2. Added `ELASTICITY_BY_DECILE` (Priesmann & Praktiknjo 2025, −0.64 D1 → −0.11 D10 linear interp) to `config.py`. Documented the uniform-vs-differentiated distinction and the constant- elasticity-to-+161% caveat in the config docstring. 3. Pinned `policyengine`, `policyengine-uk`, `microdf-python`, `pandas`, `numpy` in `pyproject.toml`. Updated README install steps to `pip install -e .` now that deps are declared. 4. Vectorised `_build_household_type` — the previous Python loop over benunits / persons became pandas groupby + boolean masks. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

vercel · 2026-04-20T13:34:30Z

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

Project	Deployment	Actions	Updated (UTC)
energy-price-shock	Ready	Preview, Comment	Apr 21, 2026 7:16pm

Review findings follow-up: - Pin `policyengine-uk` to the exact version the `policyengine` 4.1 manifest requires (2.88.0). Floor was triggering a compatibility warning on every import. - Bump `policyengine` floor to 4.1.0 now that 4.0 shipped and 4.1 is on PyPI (includes the extra_variables fix from PE.py#307). - Wire `ELASTICITY_BY_DECILE` into `behavioral_responses` — each scenario now emits a `deciles_priesmann` list alongside the headline uniform-elasticity `deciles`, so readers can see the progressivity the uniform value suppresses. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

MaxGhenis · 2026-04-20T14:23:59Z

Subagent review flagged blockers; now resolved:

extra_variables plumbing — Honor Simulation.extra_variables on US/UK microsim paths (closes #303) policyengine.py#307 landed (merged by Max). Now released as policyengine==4.1.0 on PyPI. Simulation.extra_variables is honoured by the UK run() loop.
policyengine>=4.0.0 PyPI availability — 4.1.0 is live.
policyengine-uk pin — bumped to exact ==2.88.0 (the manifest-matching version) to silence the compatibility warning.
Priesmann sensitivity — no longer config-only dead code. behavioral_responses now emits a deciles_priesmann array alongside the headline deciles. Uses income-differentiated −0.64 (D1) → −0.11 (D10) elasticities per Priesmann & Praktiknjo (2025).

Left for a later pass (not blocking):

Wiring deciles_priesmann into the dashboard as a sensitivity toggle.
Caveat UI text for the +161% extreme scenario (README has the caveat; the dashboard does not).

Couldn't verify end-to-end locally because importing policyengine now requires HF auth to fetch the data release manifest (DataReleaseManifestUnavailableError without HUGGING_FACE_TOKEN). CI should have the token set; if it doesn't, we'll know from the build logs.

## Analysis Drop the uniform −0.15 elasticity in favour of Priesmann & Praktiknjo (2025) income-differentiated values (−0.64 D1 → −0.11 D10). Every behavioural-response calculation now operates on per-household elasticities, so low-income households correctly cut consumption more sharply and the progressivity of a shock flows through to output. Affected: - `behavioural_responses` (renamed) - `policy_net_position`, `policy_post_shock`, NEG, RBT scenarios - `_grouped_post_policy` and `_grouped_shock` helpers — now take a per-household behavioural factor array and aggregate within groups, reflecting each group's actual decile composition - Headline scenario output now reports `elasticity_by_decile` and `mean_elasticity` (weighted mean from the dataset) instead of a single hardcoded number ## Spelling British English throughout internal identifiers, JSON output keys, and dashboard: `behavioral` → `behavioural`. Existing results JSON files migrated in-place so the dashboard keeps rendering until the next full regeneration. ## Packaging Upgrade to a proper Python package: - `pyproject.toml` uses hatchling, declares authors, license (AGPL-3), classifiers, URLs, `[project.scripts]` entry point, `[dev]` extra (pytest, ruff), ruff + pytest config. - `energy_shock/__init__.py` exposes `__version__` via importlib metadata and the public run_all / run_all_countries entry points. - Lazy-import `policyengine` inside `baseline._load_dataset` / `_build_simulation` so importing `energy_shock` for pure helpers doesn't require an HF token. ## Tests + CI - `tests/test_elasticity.py` covers decile mapping, missing-decile fallback, zero-price identity, and low-vs-high-decile monotonicity. Pure-function unit tests — no microsim, no dataset download. - `.github/workflows/pull_request.yaml` runs ruff format + lint, pytest, and dashboard Vite build on every PR + push to main. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…d, prose sync, LICENSE Physics fix: Switch _behavioural_factor_hh from the linear first-order approximation (1+p)(1+eps*p) to the canonical constant-elasticity form (1+p)^(1+eps). The linear form produces a negative spending factor for combinations like eps = -0.64 and p = +1.61 — implying consumption cuts > 100 %, which is physically impossible and was being rendered as-is in the Q1 2023 peak KPIs. Log-linear stays admissible for all eps in (-1, 0] and p >= 0, and collapses to the linear approximation for small p. New regression test test_behavioural_factor_physically_admissible_at_extreme_shock asserts positivity across the full Priesmann range at +161 %. Stale dashboard data guard: The JSONs under dashboard/src/data/ still held uniform-0.15 numbers from the old run, textually migrated to British spelling but not regenerated. Rendering would show the OLD analysis with NEW labels. Replaced each JSON with a stub carrying _stale: true. Dashboard checks for the flag at top level and renders a Data generation pending banner with the regen command instead of the full panels, so the site never shows misleading numbers until someone regenerates with HUGGING_FACE_TOKEN set. Dashboard prose: Replace Labandeira-uniform language in three places with the Priesmann decile-specific account and cite the constant-elasticity form. Remove the elasticity = -0.15 KPI tooltip; add a Mean elasticity KPI that reads from results.behavioural.mean_elasticity. README sync: - Add HUGGING_FACE_TOKEN requirement to install steps - Swap uniform-elasticity description for the decile-specific account - Update Requirements and Tech stack to match ==2.88.0 pin - Add pytest section Licensing + packaging: - Add LICENSE (AGPL-3.0-or-later) - uvx ruff check --fix --unsafe-fixes . + ruff format . clean - Swap dashboard-build CI job from npm to bun; bun lockfile committed, npm lockfile removed Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

policyengine.py 4.1.0 can't boot without upstream fixes to its bundled UK release manifest (the pinned data-package version doesn't have a release_manifest.json published on Hugging Face, and the bundled-certification fallback requires an exact model-version string match that the installed 2.88.0 hits but the manifest fetch short-circuits before getting there). Revert the simulation layer to raw policyengine_uk.Microsimulation — exact same reform-dict shape, exact same output values, just without the manifest bootstrap drama. Migration to policyengine.py can be revisited once upstream stabilises. Regenerate all 10 dashboard JSON files with the new decile-specific elasticity + log-linear constant-elasticity behavioural model: UK: 32.0 M households, mean energy £1,496 England: 26.9 M, £1,498 Scotland: 2.7 M, £1,479 Wales: 1.6 M, £1,363 Northern Ireland: 0.7 M, £1,742 Spot-check +60% behavioural hit by decile: D1 (ε -0.640): £181 (31 % of static) D5 (ε -0.404): £537 (54 %) D10 (ε -0.110): £1,282 (87 %) The previously flat +45.6 % behavioural line under the uniform −0.15 model is now a steep progressive slope — the Priesmann progressivity that the old model suppressed is now the headline. Q1 2023 peak (+161 %) behavioural is now a defensible +£1,568 avg instead of the broken negative values the linear form was producing. pyproject.toml drops the policyengine package from direct deps (still transitively reachable if anyone wants to try the unified API manually). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…consumption-pct, README, LICENSE Four codex-flagged blockers resolved: 1. CI Test job: setup-uv without python-version doesn't provide a system Python, so `uv pip install --system` fails. Replace with `uv venv --python 3.13 .venv && uv pip install --python .venv/bin/python` and run pytest from the venv. 2. Council Tax Rebate was paying non-English households because the PE-UK formula keys off council_tax_band alone (non-NONE outside England in the FRS imputation). Real 2022 policy was England-only. Add `_england_mask_unfiltered(sim, cmask)` and zero out non-English rows in all three call sites (`policy_ct_rebate`, `policy_post_shock` via `hh_payments`, and `policy_combined`). Verified post-regen: Scotland / Wales / Northern Ireland now show £0 CT rebate; England at £6.5 bn (down from £7.7 bn UK-wide). 3. `consumption_reduction_pct` in the behavioural output was still using the linear first-order `eps_d * price_pct` instead of the log-linear `(1+p)^eps - 1`. At +161 % for D1 it reported `-102.9 %`, physically impossible. Fixed to the log-linear form; now D1 Q1-2023-peak reduction reports `-45.8 %`, matching `(1 + 1.61) ** −0.64 = 0.5412`. 4. README was still describing the reverted policyengine.py migration path and showed npm install commands. Updated to match the raw policyengine_uk implementation and the bun-based dashboard tooling the CI workflow uses. Plus two codex nits: - LICENSE replaced with canonical 661-line AGPL-3.0 text (was an abridged preface-only file that wouldn't satisfy GitHub's license scanner). - Tech-stack line sync'd with pyproject.toml. Regenerated all 10 dashboard JSONs. Ruff clean, 6 tests pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

One blocker and three nits from codex round 2: 1. Public-facing copy inconsistent with England-only CT rebate: - README.md:10: add ", England only (mirrors the 2022 Council Tax Rebate's geographic scope)" to the rebate bullet. - Dashboard POLICY_META description: reworded to note VOA's ~63 % A-D headline and explain why PolicyEngine's FRS-based band imputation can overshoot that in the model output. - Dashboard "Key findings" summary: qualify the rebate as "England only". - sections.py JSON policy description string: include the England-only qualifier so downstream consumers of the JSON aren't misled. 2. filter_by_country() replaced row-level data["country"] with a scalar string, which quietly broke the shape contract (downstream groupby / boolean masking by country assumed an array). Keep data["country"] as a per-household array; move the scalar filter name to a new data["country_label"]. Tested against all four constituent nation outputs. 3. CI bun install fallback (bun install --frozen-lockfile || bun install) masked lockfile drift. Drop the fallback so CI fails deterministically if bun.lock and package.json go out of sync. Regenerated all 10 dashboard JSONs after the filter_by_country fix, reran ruff + pytest + bun build — all clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

policy_net_position() isn't wired into generate.py today, but if it gets revived it would reintroduce the non-English CT-rebate bug. Apply the same England-only mask as the other CT-rebate call sites. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Four blockers and five material concerns from the Opus 4.7 publication review. B1. Rename `results_v2*.json` → `results_breakdowns*.json`. The `_v2` suffix violates Max's global naming rule ("use git for history"); `breakdowns` describes what the payload contains (electricity/gas split, tenure, household type, country, NEG). All 5 JSONs renamed, Dashboard.jsx imports + destructuring updated (resultsV2 → breakdowns, ALL_DATA.<country>.v2 → .breakdowns), generate.py output path + log lines updated. B2. Sdist allowlist in pyproject.toml. Without it, hatchling included the entire repo — dashboard subtree (bundled JSONs + React source) inflated the source distribution by ~10 MB. Restrict to `energy_shock/**/*.py`, `tests/**/*.py`, README, LICENSE, pyproject.toml. Built tarball is now 30 KB with 13 files. B3. Drop unused code paths from dashboard output: - `policy_post_shock.epg` was shipped but never selected in the dashboard (POLICY_META only exposes flat_transfer, ct_rebate, bn_transfer, bn_epg, neg). Removed `epg` branch + pe_policies entry. - `policy_post_shock.rbt` (rising-block-tariff-under-shock scenario loop, ~100 lines) also never consumed. Removed. - Standalone `rising_block_tariff()` in sections.py + its call in generate.py + `rising_block_tariff` key in breakdowns output — all unused. Removed. - `RBT_DISCOUNT_RATE` constant no longer needed. Dropped. - Results JSON files now 40% smaller (327 KB → 196 KB per country). B4. `welfare_loss_comfort_avg` (Harberger-triangle approximation, `0.5 * |ε| * p^2 * energy`) was inconsistent with the constant- elasticity demand used everywhere else — overstates welfare loss 3× at the +161 % scenario. The field was never rendered in the dashboard, so dropped rather than replaced with the exact log-linear CS integral. Also removed a dead `energy * price_pct` expression statement that the same block carried. M1. Added Priesmann transferability caveats to README methodology section and dashboard methodology prose: the paper estimates from German *gas* demand, linear interp between D1 and D10 is a convenience, and UK electricity is typically less elastic than gas — so behavioural bill savings should be read as an upper bound on consumer adjustment. M2. Prose correction in 3 places (KPI info tooltips + methodology paragraph + intro): "imputed from NEED 2023" → "imputed from Living Costs and Food Survey, calibrated against NEED 2023 administrative totals". NEED is an aggregate admin dataset, not household-level; LCFS is the household survey PolicyEngine uses for energy consumption imputation. M4. Removed the `DATA_STALE` guard + "Data generation pending" stub screen from Dashboard.jsx. It tested for `_stale: true` on bundled JSONs, but no code in the package ever emits that key — pure dead code. M5. README setup instructions and generate.py module docstring switched from `conda activate python313` / `pip install -e .` to `uv venv --python 3.13 .venv && source .venv/bin/activate` / `uv pip install -e .` to match the CI install path and Max's standard Python tooling. Regenerated all 10 dashboard JSONs under the new code; reran ruff check + format, pytest (6/6), sdist build, bun install + build — all clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Four policy functions in energy_shock/sections.py were fully implemented but never called from generate.py's _run_one and therefore produced no output in any of the dashboard JSONs: policy_epg, policy_wfa, policy_combined, policy_net_position. Remove them so the module accurately reflects the analyses the dashboard actually consumes, and drop the now-unused EPG_TARGET / WFA_HIGHER / WFA_LOWER constants from config.py. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Round-4 fresh-team referee report surfaced blockers the earlier Codex rounds missed. Addressed here: Methodology / practitioner • Post-policy "extra cost" metric was `max(shocked_e − p, e)`, clipping the residual at baseline. Low deciles under shock-matching or flat transfers were in fact over-compensated (e.g. D1 flat transfer at +10 %: residual £10 but average household is £302 better off than baseline). Now emit a signed `net_change` and `behavioural_net_change` field per decile and per group alongside the existing clipped `extra_cost`. Dashboard tooltip now shows both so the progressivity story lands. • Section-description prose explicitly explains the residual vs net-change semantics so readers don't misread the chart as "every decile still hurt". • NEG subsidy base kept static (pre-shock consumption up to 2,900 kWh), but the choice is now documented in both `neg_policy`'s docstring and the policy description card. An alternative indexing to actual post-response consumption would shrink subsidy cost for high-elasticity low-income deciles that cut below threshold; noted as a design caveat. • Deleted dead code the earlier review rounds had overlooked: `policy_epg`, `policy_wfa`, `policy_combined`, `policy_net_position`, `regional_breakdown`, `gas_price_cap`. None were called from `generate.py` or consumed by the dashboard. Also removed `EPG_TARGET`, `WFA_HIGHER`, `WFA_LOWER` constants and the `RBT_DISCOUNT_RATE` leftover. Net cut: ~450 lines of sections.py. • Removed leftover unused tuple / bare-expression statements in `baseline_summary` and `shock_scenarios`. • Strengthened `test_behavioural_factor_physically_admissible_at_extreme_shock` to assert explicitly that the linear first-order form is negative at (ε=−0.64, p=1.61) and that the returned factor is NOT close to that linear form — guards against future regression to the quadratic Taylor approximation. Added `test_epsilon_fallback_honours_weights` covering the non-uniform weighting path. Citations (URLs now resolve to the claimed source) • fn-2 Cornwall Insight: swap to the 4 March 2026 july-price-cap-forecast-rises-to-1800 release (the previous URL went to a May 2025 £1,720 announcement). • fn-3 Stifel/£2,500: swap to the GB News energy-bills-uk-households-iran-gas-prices article (the previous URL went to an unrelated £160 Cornwall Insight piece). • fn-4 Resolution Foundation: swap to the £480 comment piece (the previous URL went to an earlier £500 press release). • fn-8: clarify that £4,279 is the *announced* Q1 2023 cap — the concurrent Energy Price Guarantee held typical bills at £2,500. Domain / scope • README and dashboard methodology now flag (a) the gas-vs-combined shock simplification, (b) the ~£290/yr of fixed standing charges that a uniform-percentage shock rescales, and (c) the +161 % scenario being an illustrative stress-test not a realised episode. • NEG threshold 2,900 kWh is now attributed to Bangham (2026) mirroring the 2022 Austria/Netherlands relief design, not a "UK median" (the Ofgem TDCV is closer to 2,700 kWh). Reference wiring • `Labandeira et al. (2017)` in-text mention now links to its reference entry. • fn-12 (EPG) linked from the cap-freeze subsidy description. • fn-13 (HM Treasury factsheet) linked from the methodology intro. UI polish • Inline "illustrative scenario" banner on the +161 % scenario's KPI panel in both the Impact and Policy responses tabs. • "Shock-match" label unified to "Shock-matching" across the Policy responses tab (matches the prose and POLICY_META.fullName). Reproducibility • `vercel.json` now uses `bun install --frozen-lockfile && bun run build` to match CI's deterministic install path (was `npm install`, which synthesised a lockfile on the fly). • CI test job now runs a Python 3.13 + 3.14 matrix (the classifier advertised 3.14 but only 3.13 was exercised). • Removed empty `tests/__init__.py` (pytest auto-discovers). Regenerated all 10 dashboard JSONs under the slimmer code. ruff check + format, pytest (7/7), bun dashboard build, sdist build (30 KB, 13 files) all clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Codex round-5 blocker and nits on top of d41a760. 1. Policy-cost chart contradicted itself under the behavioural toggle. `getExchequer` shrank shock-matching and NEG by the whole-population behavRatio even though neither policy scales with consumer response: - Shock-matching is a flat per-household payment pegged to the *static* average shock (`sections.py` `policy_key == "bn_transfer"`), so aggregate cost is fixed once the payment is set. - NEG subsidy is explicitly indexed to each household's pre-shock consumption up to 2,900 kWh (documented in the `neg_policy` docstring and the `subsidy_indexed_to` JSON field), so aggregate cost does not depend on whether households cut consumption. Only the cap-freeze subsidy (`bn_epg`) reimburses each household's actual bill increase, so it alone scales with the behavioural / static ratio. `getExchequer` now reflects that. The Policy-at-a- glance chart no longer disagrees with the selected-policy panel at `dashboard/src/components/Dashboard.jsx:725`. 2. Tooltip display polish: - Negative net-change values now format as `-£366` rather than `£-366`. - Metric label renamed from "net change" to "net change vs baseline" so the reference point is explicit. 3. `consumption_reduction_pct` in the `behavioural` section reported a negative number even though the field name reads as a positive reduction ("D1 loses 27%"). Flipped the sign: at +10% the D1 elasticity now reports a 5.9% reduction; at the +161% peak, 45.8%. The comment block was correct all along — the code was off. Regenerated all 10 dashboard JSONs; ruff + pytest (7/7) + bun dashboard build all pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Codex round-6 caught that the per-policy KPI panel at `Dashboard.jsx:725` still hardcoded static values for cap-freeze (bn_epg). Under the behavioural toggle at +60%, the panel therefore reported benefit ≈£1,045/yr and cost ≈£33.4bn even though the response-consistent numbers are ≈£637/yr and ≈£20.4bn. Cap-freeze (bn_epg) is the only modelled policy whose payment scales with the consumer response — the government reimburses each household's *actual* bill increase, which falls when households cut consumption. Shock-matching, NEG, flat transfer and CT rebate all pay an amount that is invariant to behavioural response, so their KPI values correctly stay static. `getExchequer` on the overview chart already applies `behavRatio` to `bn_epg` only; this commit aligns the selected-policy KPI panel with that logic so the chart and the card agree. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Applies the same review lens as PolicyEngine/energy-price-shock#1. Methodology / correctness: - TransitionMacroImpact has no r_baseline / r_reform fields. The Code-tab block under map_transition_to_real_world wrongly listed them; replaced with the actual interest-rate paths on the TPI results (base_tp.r, reform_tp.r). - impact.years on TransitionMacroImpact is fiscal-year strings ("2026-27", ..., not ints). Updated the inline comment. - Default TPI T is 60 (oguk_default_parameters.json), not 30. Fixed the Code-tab header, transition-path narrative, the Example-tab description, and the Methodology SVG label. - Flagged the steady-state terminal output as illustrative — the numbers were plausible but not from a real model run. CI: - New .github/workflows/checks.yml runs node --check on the inline JS, verifies every locally-referenced asset exists, and asserts the four oguk API symbols the Code tab cites are still exported upstream. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

vercel Bot deployed to Preview April 20, 2026 13:34 View deployment

vahid-ahmadi approved these changes Apr 20, 2026

View reviewed changes

vercel Bot deployed to Preview April 20, 2026 14:23 View deployment

vercel Bot deployed to Preview April 20, 2026 14:39 View deployment

vercel Bot deployed to Preview April 20, 2026 15:19 View deployment

vercel Bot deployed to Preview April 20, 2026 17:38 View deployment

vercel Bot deployed to Preview April 20, 2026 18:14 View deployment

vercel Bot deployed to Preview April 21, 2026 11:31 View deployment

vercel Bot deployed to Preview April 21, 2026 11:37 View deployment

vercel Bot deployed to Preview April 21, 2026 12:15 View deployment

vercel Bot deployed to Preview April 21, 2026 15:21 View deployment

vahid-ahmadi changed the title ~~Migrate to policyengine.py API; finish RBT post-shock; pin deps~~ Methodology review fixes: England-only CT rebate, log-linear behavioural form, decile-specific elasticities Apr 21, 2026

vercel Bot deployed to Preview April 21, 2026 15:49 View deployment

vercel Bot deployed to Preview April 21, 2026 19:09 View deployment

vercel Bot deployed to Preview April 21, 2026 19:16 View deployment

MaxGhenis merged commit 85249c4 into main Apr 21, 2026
6 checks passed

MaxGhenis deleted the review-fixes branch April 21, 2026 19:59

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Methodology review fixes: England-only CT rebate, log-linear behavioural form, decile-specific elasticities#1

Methodology review fixes: England-only CT rebate, log-linear behavioural form, decile-specific elasticities#1
MaxGhenis merged 13 commits into
mainfrom
review-fixes

MaxGhenis commented Apr 20, 2026 •

edited by vahid-ahmadi

Loading

Uh oh!

vercel Bot commented Apr 20, 2026 •

edited

Loading

Uh oh!

MaxGhenis commented Apr 20, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

MaxGhenis commented Apr 20, 2026 • edited by vahid-ahmadi Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Why

What shipped

Methodology

Code

Dashboard

Not changed

Test plan

Uh oh!

vercel Bot commented Apr 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MaxGhenis commented Apr 20, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

MaxGhenis commented Apr 20, 2026 •

edited by vahid-ahmadi

Loading

vercel Bot commented Apr 20, 2026 •

edited

Loading