viktor/fire-planner
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix(fire-target): Family/home targets monotonic (kills Family==Household)
Some checks failed
Build and Push / lint-and-test (push) Has been cancelled
Details
Build and Push / build (push) Has been cancelled
Details
Build and Push / deploy (push) Has been cancelled
Details
Build and Push / notify-failure (push) Has been cancelled
Details

Browse source 
The recompute solved each Case's FIRE number with an independent binary search,
so Monte-Carlo path noise + the coarse £15k tolerance made Family (Household + 2
kids) tie or even UNDERCUT Household (6 hard inversions + 5 exact ties across the
22 countries) — the ~£20k kids cost quantised to ~£0.

Now solve the Cases in increasing-cost order and lower-bound each by the previous
Case's target on the SAME return paths: a heavier Case (more spend / +kids /
+home) can never need less net worth than a lighter one, so
Solo <= Household <= Family <= Family+home holds by construction. tol tightened
15k -> 1k so the genuine but small kids/home increment resolves instead of
snapping to the previous grid step. Kids/home were already modelled correctly
(verified) — this is purely a solver-resolution + monotonicity fix.

Found + verified via the fire-countdown flaw-hunt workflow. 346 tests pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Viktor Barzin 2026-07-01 22:42:34 +00:00
parent 6efe3b0c31
commit 1b8809a01b
2 changed files with 33 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

19
fire_planner/__main__.py
View file

@ -494,6 +494,16 @@ async def _recompute_fire_targets(
)
jur = jurisdiction_for_city(slug)
kids_cf = kids_annual_spend(ratios, kids_base=kids_base)
# Solve the Cases in increasing-cost order and lower-bound each
# by the previous Case's target on the SAME return paths. A
# heavier Case (more spend / +kids / +home) can never need LESS
# net worth than a lighter one, so this guarantees
# Solo <= Household <= Family <= Family+home by construction —
# killing the Monte-Carlo-noise inversions where Family looked
# cheaper than Household. tol is tight (£1k) so the genuine but
# small kids/home increment (~£20k at the 99% bar) resolves
# instead of being quantised to the previous grid step.
prev_target = 0.0
for case in (Case.SOLO, Case.HOUSEHOLD, Case.FAMILY):
spend = case_base_spend(case, ratios)
home_variants = (False, True) if case is Case.FAMILY else (False,)
@ -509,10 +519,13 @@ async def _recompute_fire_targets(
kids_start_year=KIDS_START_YEAR, kids_end_year=KIDS_END_YEAR,
with_home=with_home, home_amount_gbp=home_amount, home_year=home_year,
)
# Bound the search to a sane SWR band (spend × 60 ≈
# 1.67% floor) so the binary search converges fast.
# hi = sane SWR band (spend × 60 ≈ 1.67% floor); lo =
# previous (lighter) Case's target -> monotone chain +
# a small band for the nested solves (fast).
hi = min(5_000_000.0, spend * 60.0)
res = solve_target_nw(
paths, inp, hi=min(5_000_000.0, spend * 60.0), tol=15_000.0)
paths, inp, lo=min(prev_target, hi), hi=hi, tol=1_000.0)
prev_target = float(res.target_nw_gbp)
await upsert_fire_target(sess, inp, res, n_paths)
written += 1
tag = "+home" if with_home else ""