Skip to content

Declare transitions by age, fold the transitory medical shock, move execution controls to Model policy - #21

Merged
hmgaudecker merged 115 commits into
mainfrom
combine/open-prs
Oct 22, 2026
Merged

Declare transitions by age, fold the transitory medical shock, move execution controls to Model policy#21
hmgaudecker merged 115 commits into
mainfrom
combine/open-prs

Conversation

@hmgaudecker

@hmgaudecker hmgaudecker commented Oct 22, 2026 •
edited
Loading

Copy link
Copy Markdown
Member

Regime transitions

  • Transitions between stages are declared with pylcm's ByAge schedules. The ages are AgeRange constants taken from the config.
  • The Regime field is regime_transitions. The dead regime is terminal (regime_transitions=None).
  • A coverage test pins the exact set of (age, regime) nodes.

Medical costs

  • A new grid flag, fold_hcc_transitory, removes the IID transitory medical-cost shock from the state space. The shock is then drawn only on the transition.
  • With subtract_hcc_bill, every living regime's asset law becomes resources before the bill plus the out-of-pocket bill as an additive shock (lcm.AdditiveShockTransition).
    • The solve then averages over the transitory shock once per period instead of at every source point.
    • This requires the brute-force asset law.
  • The ACA overlay computes the cost-sharing scale from a finite cost-sharing bracket, so the bill can be averaged bracket by bracket. The cost-sharing schedule itself is unchanged.
  • The benchmark config turns both flags on. The production grid defaults leave them off.

Execution and solvers

  • Execution controls move to pylcm's Model policy: a new execution.py with execution_config_for_devices. The model factories no longer take population hints.
  • NB-EGM solves every living regime and uses pylcm's discrete envelope over several actions. The DC-EGM and NB-EGM regimes declare their liquid Euler margin.
  • The borrowing limit is declared structurally. ACA runs can be set to fp32 explicitly.
  • Tests use pylcm's SolutionResult solve API, and the benchmark uses breakpoint grids.

CI

  • pylcm is installed at a pinned commit.
  • actions/setup-python is updated from 6 to 7.

Supersedes #16, #17, #19 and #20.

https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW

hmgaudecker and others added 30 commits August 14, 2026 21:16
@hmgaudecker @claude
Let the NBEGM regime declare every choice its structure affords
5e5d529 
`build_actions` and the nongroup builder narrowed the M1 regime for NBEGM,
dropping `buy_private` and `labor_supply` as actions and fixing their former
outputs to constants. That made the solved model a different model from the one
brute force solves: the household lost its coverage and hours choices.

The regime now declares whichever choices its structure affords under every
solver. pylcm's ride-along discrete envelope is written over a single action's
grid and refuses a regime declaring several, so model build under NBEGM raises
until that arity widens. The refusal is the honest outcome — a solver that
cannot carry a choice refuses the regime rather than being handed a narrower
one.

Tests split accordingly: the regime-level and solver-config assertions stay
green, model-build assertions become strict xfails naming the arity, and one
green test pins the refusal itself so the xfails flip when the arity widens.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018yHhuFzqsDw1MhdB1i2Ljm
@hmgaudecker @claude
Consume pylcm's multi-action discrete envelope
3b2a0b1 
The M1 regime declares both buy_private and labor_supply, and pylcm's envelope
now branches over the Cartesian product of a regime's discrete action grids, so
the eight strict xfails pinned to the old single-action refusal all XPASS.
Removes them and converts the refusal test into the capability it replaced: the
model builds with both actions live.

25 passed where 15 had failed. Every one of the 15 was a stale expectation --
including the five policy-variant builds, which are reported as FAILED rather
than XPASS under strict=True and so read like real build failures until the
XPASS(strict) marker is inspected.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018yHhuFzqsDw1MhdB1i2Ljm
@hmgaudecker @claude
Solve every living regime with NB-EGM
9a5fe8a 
`solver="nbegm"` attached the solver to one regime and left the other 17 on
brute force, so the model's stated solver was not the one that produced most
of its result. Every living regime now gets the NBEGM config, and the retiree
and tied builders carry the savings-form budget (`resources`, `savings`) that
the solver's contract reads -- without it those regimes cannot be built at all.

The build-time affinity and interval-constancy probes cannot run on the added
regimes: they differentiate the budget on scalar inputs and
`assets_and_income.capital_income` declares `rate_of_return: ScalarFloat`,
which the probe's one-element array violates. The probes fall back to
`assume_declared` and warn, so NB-EGM's exactness precondition is asserted
rather than checked there, and the solve needs validating against an
independent reference.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018yHhuFzqsDw1MhdB1i2Ljm
@hmgaudecker
Let pylcm's NBEGM probes check the ACA budget
ace111b 
The probes run on the first solve against the model's complete parameter
vector, so the tax tables and threshold schedules the budget reads are
the model's own rather than synthesized stand-ins. Affinity and
constancy are checked rather than asserted.
@hmgaudecker
Declare the liquid Euler margin on the EGM regimes
4f25818 
pylcm takes the liquid roles from the regime: a `ConsumptionSavingsRegime`
names the liquid state, the action paid from resources, the resources node
and the post-decision state, and the solver carries numerical configuration
only. `ACA_LIQUID_MARGIN` states that once for the three living-regime
builders, which now route through `build_alive_regime`; a brute-force regime
has no `resources` or `savings` node to name and stays a plain `Regime`.

The role assertions move with the declaration: they read `regime.liquid`,
which is public API, rather than the bound solver's attributes.
@hmgaudecker @claude
Describe the borrowing constraint's actual solver coverage
221bf4f 
Both callers of `build_model_constraints` still said DC-EGM gets no
broadcast constraint. It is declared under every solver: the EGM solve
enforces the limit through the savings grid's lower bound, but forward
simulation re-decides consumption by an argmax over the consumption grid
and needs the explicit feasibility mask.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018yHhuFzqsDw1MhdB1i2Ljm
@hmgaudecker @claude
Name the DC-EGM blocker that actually fires
8466aac 
The xfail reason described the assets-law chain, but the build stops
earlier: pylcm's DC-EGM refuses the broadcast borrowing constraint for
reading continuous variables. Both blockers are now listed in the order
they fire.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018yHhuFzqsDw1MhdB1i2Ljm
@hmgaudecker
FIX Declare the ACA borrowing limit structurally
a2ad958
@hmgaudecker
Migrate ACA benchmark to breakpoint grids
328e371
@hmgaudecker
ENH Forward NB-EGM workspace controls
f61cff9
@dependabot
Bump actions/setup-python in the github-actions group across 1 directory
87fc81d 
Bumps the github-actions group with 1 update in the / directory: [actions/setup-python](https://github.com/actions/setup-python).


Updates `actions/setup-python` from 6 to 7
- [Release notes](https://github.com/actions/setup-python/releases)
- [Commits](actions/setup-python@v6...v7)

---
updated-dependencies:
- dependency-name: actions/setup-python
  dependency-version: '7'
  dependency-type: direct:production
  update-type: version-update:semver-major
  dependency-group: github-actions
...

Signed-off-by: dependabot[bot] <[email protected]>
@hmgaudecker
ENH Allow explicit fp32 ACA runs
5d1d479
@hmgaudecker
FIX Gate the experimental NB-EGM planner
94977a3
@hmgaudecker @claude
MAINT Migrate the tests to pylcm's SolutionResult solve API
3af83d0 
pylcm's `Model.solve` returns a `SolutionResult` and `Model.simulate`
takes `solution=` in place of `period_to_regime_to_V_arr=`. The value
mapping these tests read is `SolutionResult.values`; passing `None` for the
old keyword meant "solve inside simulate", which is the new default.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Y7QCx9tkGD3TqwNw681QwD
@hmgaudecker
ENH Move ACA execution controls to Model policy
b941507 
Forward ExecutionConfig through baseline, ACA and benchmark factories. Derive omitted accelerator budgets from selected JAX allocator limits and reject retired GridConfig execution fields. Preserve economic grids and numerical solver settings.

Validation against pylcm f6d41d54: exact ASV CPU preflight, focused compatibility tests, Ruff, ty and repository hooks pass. DC-EGM construction still exposes an upstream continuation-template placement failure on three CPU devices.
@hmgaudecker
API Remove constructor population hints from model factories
735a250
@hmgaudecker @claude
TST Pin float32 felicity at the leisure floor against a 60-digit refe…
4ec7ce0 
…rence

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
ENH Bend leisure onto a positive floor with a smooth max
4f37ec2 
Leisure is s*logaddexp(F/s, available/s) with F = 1e-5 * time_endowment, so
float32 felicity and its consumption derivative stay finite when work costs
exceed the endowment, while cells away from the floor are unchanged to ~1e-8.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
CI Install pylcm at the commit aca-dev pins
05c4a43 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
Floor leisure at a small positive value so float32 utility stays fini…
3bc89f7 
…te (#18)

Co-authored-by: Claude Opus 5.5 <[email protected]>
@hmgaudecker @claude
Declare regime transitions by age with regime_transitions
fd30823 
Stage transitions use `ByAge` schedules over `AgeRange` constants from the
config, the dead regime is terminal with `regime_transitions=None`, and a
coverage test checks every (age, regime) node. CI installs the matching pylcm.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
Test against pylcm without Regime.active
bf4c654 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
Declare the 51-61 entry regimes as the model's starting pairs
8f6ea5d 
Both factories declare INITIAL_REGIMES (ages 51-61 x the five
*_inelig_canwork regimes) instead of forwarding an optional root mapping.
Declarations use the keyword-only AgeRange/ByAge/MarkovTransition
signatures. simulate_with_dense_index drops initial-condition rows older
than the last admissible start age and raises on any kept row outside
model.initial_nodes.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
Narrow the admissible starting ages to 51-60
ad38653 
INITIAL_REGIMES spans ages start_age..last_start_age (51-60, the baseline
estimation sample) in the five *_inelig_canwork regimes, 50 pairs. Starts
at age 61 or later are dropped by select_admissible_starts and rejected at
admission. The solved domain stays 181 nodes: age-61 entry regimes are
reached from age 60.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
Test the per-type transformed bequest weight in the dead regime
57eda22 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
TST Expect the ACA Medicaid expansion to reach the disabled under 65
a13aeca 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
Extend the ACA Medicaid expansion to the disabled under 65
b69fe0c 
The expansion track tests MAGI only, for every household under 65,
regardless of health or assets, as the draft's post-ACA eligibility rule
states. The aged keep the categorical track with its asset test.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
TST Expect employer coverage to survive expansion-only Medicaid eligi…
28e4d72 
…bility

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
Keep employer coverage for households eligible only through the ACA e…
b0b5ed1 
…xpansion

Retiree and tied HIS transitions, and the pension imputation's target HIS,
read categorical (SSI) Medicaid eligibility rather than the two-track
eligibility. The draft assumes the ACA does not affect employer-provided
insurance; an expansion-eligible retiree or tied household keeps its
employer coverage, with Medicaid paying on top of its OOP.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
TST Pin budget, Social Security and pension rules at production values
3bbcc5a 
Tests for the PIA carried across a claim-age adjustment, the earnings-test
credit, age x health spousal income, disabled AIME accrual, frozen AIME at
forced-claim ages, the taxable-maximum cap after accrual, the end of pension
accrual at 71, the household's own consumption floor and the ACA surtaxes.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
hmgaudecker and others added 22 commits October 7, 2026 00:55
@hmgaudecker @claude
Declare regime transitions on Model.edges
938c733 
Each living regime's per-target probability cells move from
`Regime(regime_transitions=...)` into `Transition(targets=..., law=...)`
under `Model(edges=...)`; `dead` has no outgoing edges and is terminal.
Terminal checks read `model.graph.laws[name].terminal`. Edges, targets
and probability cells are unchanged. Standalone CI installs the pylcm
revision that carries the graph-only transition API.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
TEST Expect the benchmark law arguments under params["edges"]
0a5ea10 
The regime-transition law reads `survival_probs` and `prob_disabled_next`;
pylcm resolves an edge-declared callable's parameters under
`params["edges"][source]`, never under the source regime.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
ENH Route the benchmark law arguments to params["edges"]
721d774 
`get_benchmark_params` moves each source regime's `survival_probs` and
`prob_disabled_next` from `fixed_params[source]` to
`fixed_params["edges"][source]`, where pylcm resolves the parameters of
the regime-transition law declared in `Model(edges=...)`. The law reads
the same per-source Series as before; the snapshot file is unchanged.

The NB-EGM template walk skips the `edges` root, whose entries are
source regimes rather than regime functions.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
MAINT Pin CI pylcm to the edges-params API
280e1e7 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
MAINT Read the benchmark law arguments from the snapshot's edges layout
b9915a2 
The snapshot stores `survival_probs` and `prob_disabled_next` under
`fixed_params["edges"][source]`, as `aca_slurm.assemble_fixed_params` writes
them, so `get_benchmark_params` returns the fixed params as stored.

Regenerated with aca-dev `scripts/regen_benchmark_params.py` against aca-data
652f5c6b, aca-slurm 93472b9 and pylcm db04e088; every other leaf equals the
previous snapshot.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
MAINT Pin CI pylcm to the fixed edges-params head
2400c0a 
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
REF Schedule each regime's transition cells by source age
591da5c 
pylcm derives a per-target law's targets from its cases and refuses
supplied targets that name fewer ages than the law. Each living regime's
law is a `ByAge` schedule whose cases name exactly the destinations its
source ages reach, so the edges declare no `targets` of their own. The
parameter paths are unchanged: a schedule adds no level below the source.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
MAINT Pin CI pylcm to the strict transition-targets head
5761c52 
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
FIX Drop the risk-rated premium intercept where the ACA plan is commu…
621c150 
…nity-rated

The ACA overlay replaces the non-group premium with the community-rated plan
premium, but left the risk-rated premium's zero-profit intercept in the
regime. Nothing reads it any more, yet its inputs (markup, minimum,
predicted-cost loading and the insurer-cost sample) stayed required
parameters of the ACA model, which assembles none of them.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
FIX Declare the admissible starts as InitialNodes and read their pair…
e2b17fb 
…s from the graph

pylcm now publishes `Model.initial_nodes` as an `InitialNodes` declaration and
the expanded `(age, regime)` pairs on `Model.graph.initial_nodes`. The entry
ages are declared with `InitialNodes(by_age=...)`, and admissible-start
filtering reads the pairs from the graph.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker
Test ACA against the precise pylcm API
98ebb47
@hmgaudecker @claude
ENH Add a grid flag that folds the transitory medical-cost shock
a5e6950 
`GridConfig.fold_hcc_transitory` declares the transitory shock `fold=True`.
Nothing a period computes reads last year's transitory shock, so pylcm then
draws it on the transition only: living regimes keep the lagged persistent
shock, which prices the private premium, and lose the transitory axis.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
ENH Declare the asset law as resources minus the medical bill
649bed9 
`GridConfig.subtract_hcc_bill` declares the asset law toward every target as
`lcm.SubtractedBill`: assets before the out-of-pocket bill minus `oop_costs`,
conditioned on marital status, health, Medicaid eligibility, the private-cover
choice and the ACA cost-sharing bracket. With `fold_hcc_transitory`, pylcm then
averages each target's value over the transitory medical-cost shock once per
period instead of at every source point.

The ACA cost-sharing scale is split at its income bracket, which takes finitely
many values, so the bill can be conditioned on it; the scale is unchanged.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker @claude
ENH Declare the asset law as lcm.AdditiveShockTransition
622ba00 
The shock is `negative_oop_costs`, the out-of-pocket bill negated, added to the
assets before the bill.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker
Enable medical-shock folding in the ACA benchmark
5b75bee
@hmgaudecker
Preserve unfolded controls in legacy model test fixtures
ee48d9b
@hmgaudecker
Check folded values across every solved model cell
b962ebd
@hmgaudecker
State exact ACA cost-sharing parity with a numeric assertion
89d26dd
@hmgaudecker
Use the unfolded control for lagged medical-shock twins
54397e8
@hmgaudecker
Keep the transition-only fold witness separate from additive folding
5e69b66
@hmgaudecker
Merge remote-tracking branch 'origin/dependabot/github_actions/github…
7a9f787 
…-actions-696b649f44' into combine/open-prs
@hmgaudecker @claude
Align model-creation tests with the bracket split and the bill-free E…
09a96b6 
…GM law

- The ACA nongroup regimes take the cost-sharing scale from the bracket, so
  the identity asserts name `cost_sharing_scale_of_bracket`.
- The EGM model-creation tests build from the benchmark grid config without
  `subtract_hcc_bill`, which needs the brute-force asset law.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker hmgaudecker changed the title Combine the open PRs (#16, #17, #19, #20) Declare transitions by age, fold the transitory medical shock, move execution controls to Model policy Oct 22, 2026
@hmgaudecker @claude
Build the EGM model-creation tests without the transitory-shock fold
2a36fde 
pylcm folds a shock state only under grid search.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Hs8mfXPYyEa24KaAp6BwCW
@hmgaudecker
hmgaudecker merged commit 5b95b9a into main Oct 22, 2026
1 check passed
@hmgaudecker
hmgaudecker deleted the combine/open-prs branch October 22, 2026 18:48
This was referenced Oct 22, 2026
Closed
Closed
Closed
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@hmgaudecker