Repository navigation
Conversation
…peline constant The engine records how a run treats the population in fixed_inputs.population (weights, ages, pension types), read from what it pins and the model's ages, and the pipeline words the item from it, with #14 §3's vocabulary for ONS reweighting and cohort types; a missing record, a value without wording or a missing figure fails the build. The dashboard computes the strip only for a file with no block; a damaged block shows its valid items or nothing. The paths item's calibration, floor and shock-model wording come from the results. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…est reads the figure, not a literal With the backtests on the fiscal 0.1-point rounding, the realised 2012-start gap sits at the 55.8th percentile of the untilted model (was 55.2), so the method text and docs/METHOD.md say 56th; the tilted model stays 93rd. The test now checks the text quotes whatever percentile the file records, so it fails until the rebuild, like test_not_stale, and can't drift again. Also notes that a descendant starting its own session (setsid) is not reached by run_child. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n and data/scenarios pipeline.build runs every registered scenario after the paths (the same cached engine jobs as --scenario) inside the build's start/end check and returns them with the results; triple-lock-build writes each to data/scenarios/NAME.json. --scenario NAME still reruns one alone. A test checks that a scenario run holds the central run's fixed inputs (a key only later engines record, such as population, is compared once both files have it); stubbed tests check the build writes every scenario and what the scenario jobs and records hold. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…el run; figures unchanged) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… by the dataset load The dataset load (_managed) declares what it did to the survey's weights and ages (engine.LOADED_POPULATION), since a change made inside the dataset leaves nothing to compare against; population_treatment fails without a declaration and checks it against pinned weights and the model's ages. Pension types are recorded from the rule pinned_inputs follows (PENSION_TYPE_RULE), replacing a comparison that could not fail. The block fails unless the central run and every trajectory record the same treatment. Sensitivities outside the reweightings to past dynamics, and models without a short name, are worded generically; scripts/check_assumptions.py checks a pilot's results in seconds. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…what a finished job left behind Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…the OBR-wedge scenario Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ions block from the results Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # docs/RESULTS_SCHEMA.md # tests/test_results.py
… review How jobs are started, isolated and stopped never changes what they compute, so it now lives in triple_lock/jobs.py, outside engine.ENGINE_FILES: a runner fix no longer forces a full rebuild. engine.py keeps the jobs themselves, the cache key and the cache. The job process is `python -m triple_lock.jobs --job`, which watches its parent and runs engine._job. Tests: no engine file imports jobs.py, and the cache stores each job's output exactly as the job wrote it. Fixes from the independent review of b4607a2: - run_jobs keeps its signal handlers while it stops the running jobs; kill_children's SIGKILL pass runs even if a second signal cuts the grace period short; run_child kills a child started just before an interruption (no gap before its try). - A failed job is named in the log as soon as run_jobs sees it, not only when the running jobs finish; failures seen before a Ctrl-C are logged before the stop propagates. - rules.floor_binds counts a tie with 2.5% as the floor, as triple_lock_source does (unused in the pipeline). Changes engine.py's and rules.py's code, so it reruns every job: it is meant to ride on the next full rebuild. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 4b15274)
- A failing worker sets `stop` itself, so the job queued behind it never starts (on one worker it sometimes did: 16 of 100 trials). - run_child kills the rest of a job's process group when the job exits, so nothing it left running outlives it or holds its output pipes open. - The cache key parses source bytes, so it decodes a file as Python does (a coding cookie moves it, a byte-order mark does not); a u prefix no longer moves it. - pyproject.toml counts only its dependencies, optional dependencies, Python version and build system: rewording the description or bumping the version reruns nothing. - __init__.py, which runs in every job, is in the key. The docstring guard also catches "__doc__" as a string and getdoc imports; the import walk sees absolute imports. - held_pension_types reports shapes, not sizes; model_triple_lock_rate takes float()s, since round() of a numpy float rounds as numpy does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 07c7bb2)
…zon to policyengine-uk 2.102.5 policyengine-uk 2.102.5 (PolicyEngine/policyengine-uk#1901) lets families whose adults are all over State Pension age make new Housing Benefit claims (SI 2014/1230 reg 6A(4)). The example renters therefore no longer need a reported claim (housing_benefit_reported), the claims_all_entitled_benefits override, or would_claim_uc False. No other example input is one of the seven reported benefits that switch claims_all_entitled_benefits off, so its formula gives True and the override goes too. model_horizon copies policyengine-uk's parameter builders and hash-checks them. Under 2.102.5 three of them changed: - lag_cpi and lag_average_earnings now build through lagged_series.add_lagged_parameter and derive their end from the source; - convert_to_fiscal_year_parameters skips parameters flagged preserve_calendar_dates (the alcohol and tobacco duties). The copies now mirror those builders, still extended to 2042. Verified on all four published paths, both rules, 2027-28 to 2039-40 (2,912 cells). policyengine-uk 2.90.2 with the three inputs and 2.102.5 without them give bit-identical State Pension, income tax, Pension Credit, Housing Benefit, council tax reduction, Winter Fuel Payment and net income for every example. The rerun at 2.90.2 matches data/results.json exactly. New tests: - no example sets a take-up input, and every take-up flag computes True; - Hypothesis differential: the old inputs change none of those outputs for a wholly pension-age family (single or couple, any tenure). It fails on 2.90.2 and passes on 2.102.5; - differential: every processed parameter under model_horizon equals upstream's on every date to 2034. It fails if the calendar-dates skip is removed. Not mergeable yet: the pin stays at policyengine[uk]==5.3.0 (policyengine-uk 2.90.2), and under that pin check_upstream refuses these hashes. It needs a policyengine.py release bundling policyengine-uk >= 2.102.5, then a full rebuild. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit ccf4996)
…t policyengine.py; record the pair as uncertified policyengine.py 6.2.1 pins policyengine-uk 2.102.3 and raises on import beside 2.118.0 (its data certification finds no release certified for the runtime model), and no bundle yet carries the pensioner fixes (d778). So: - pyproject.toml pins policyengine-uk==2.118.0 (with huggingface-hub and openpyxl, which policyengine.py used to bring), with a comment saying why; requirements-lock.txt is a clean install of .[uk,dev] constrained to the previous versions, so only policyengine.py's own packages leave and policyengine-uk/core move (core 3.32.16). - triple_lock/datasets.py: the Enhanced FRS 1.56.16 (the build 5.3.0 and 6.2.1 certified), 1.57.4 and Microcosm populace_uk_2023, each pinned to a revision and a SHA-256, fetched once into .cache/datasets and hashed before every use, as the managed loader did. - engine._managed loads the file by path (each simulation reads it afresh and extends it with its own parameters, as before) and attaches the provenance; run_path reads the unreformed parameters without loading a dataset (one load fewer per path). - Every run records model: the installed policyengine-uk and core, the dataset's pin, and certified: false with the reason. The results and scenario provenance carry it in place of release_bundle; the dashboard's replication line reads either. - datasets.py is an engine file (in the job key); TRACKED_PACKAGES drops policyengine and adds tables. Tests: the pins agree across pyproject and the lock and with the environment; the recorded model version equals the installed policyengine-uk (and __version__ where a release has one; 2.118.0 has none); materialize fetches once, refuses a file off its pin. Both Enhanced FRS releases load and compute on 2.118.0 with no warnings and no NaN (pilot load check). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The bridge and rehearsal runners refused whenever one-minute load exceeded twice the CPU count (36 on the 18-CPU host). On a shared host load measures contention, not danger; Fleet ops measured 150-350 for much of 11 October, which stalled Part J with 57-65 GiB free. One shared helper, triple_lock.host, now decides admission for both: refuse iff available memory is below the runner's floor or load exceeds TRIPLE_LOCK_MAX_LOAD_PER_CPU x CPUs (default 6, Subfleet's background guard; --max-load-per-cpu overrides). Non-finite readings refuse; a malformed limit stops the runner at startup. Memory floors are unchanged (bridge 40 GiB; rehearsal 40, 40 x workers for Microcosm, 8 during a job). Every host check records minimum_available_gib, max_load_per_cpu, load_limit and admitted; the configured limit and its source go in provenance.host_admission (bridge) and host_admission (rehearsal). host.py is outside engine.ENGINE_FILES: no job key, recipe hash, C2 handoff fingerprint or committed receipt changes. REBUILD documents the cutoff and the Fleet ops reservation triple-lock-rebuild-2026-10 (ready/release steps), and corrects step 2: the binding C2 re-score is the uk-triple-lock-c2 step of the loaded launchd job com.maxghenis.triple-lock-2027, not a separate label. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Moves the engine to the released certified
policyengine[uk]==6.2.6bundle: UK 2.125.1, core 3.33.0 and Enhanced FRS 1.58.0. The released wheel, immutable HF manifest and downloaded dataset SHA-256 are independently verified. The upstream audit covers fiscal lists, GB masks, resolved household heads, pension types, additional pension and the remaining horizon extensions. Publication and validation entry points enforce certification before model jobs.The historical 2.120.0 / 1.56.16 pilot remains uncertified. Frozen source/formula archives and independently pinned privacy proofs authenticate retained fiscal content. Whole single-record diagnostic families are omitted while surviving values and original provenance stay unchanged. New publishers and dashboard readers omit unsupported concentration diagnostics.
OPEN draft; pre-Budget numerical execution is incomplete. Host load exceeded the documented admission limit of 36 on 18 CPUs; a fresh vm_stat/load check at 11:51 ET found load 148.79. The certified central/40-path bridge, GB-versus-DWP coverage, Microcosm load check and timed 20 Enhanced FRS + 2 Microcosm sample did not complete. No certified savings or current measured runtime estimate is available to quote.
The ruling-(c) rehearsal uses an unauthorized dry handoff and private output directories. It cannot authorize a binding build or write results.json. Receipts retain safe timing/invariant metadata; completed cold timings exclude cancellation races. Runtime planning accounts for serial stages and leaves estimates unavailable without measurements. Bridge publication checks linked gross/net/offset, UK/GB and HB complements under the ten-record rule. Provenance is frozen across stages and checkpoints, and final receipt replacement rolls back on drift. Ageing reports stay private pending a separate joint contrast support audit.
Independent HARD review: APPROVE at
590ddac292bc9ec54d8f513bb99e288ad9dfb612; review record. This verdict covers implementation and explicitly excludes the missing numerical closeout.Validation at final head
b31137ea055ab7e2b2ef7dd8b63df1302f5affdf: Dashboard passed lint, tests and build. Pipeline is running the full suite; exact-head Python validation remains pending. Both workflows passed at the preceding documentation head 89a1309: 1,544 Python tests passed, one skipped and exactly three permitted stale-results failures (615.61 seconds; completion gate confirmed 1,548 collected tests). Local dashboard: 95 passed, lint zero errors/two existing warnings, production build passed. No unexpected test failures are waived.Execution report, certification, upstream audit and REBUILD record the evidence and remaining work. After committed 28 October inputs: binding C2/from 29 October
uk-triple-lock-c2, certified rebuild, review, then paper/dashboard update. The launchd host schedule still needs verification. Rehearsal and historical pilot outputs supply no binding replacement figures.