Pass 4's plausibility review only ever saw normalized scores, never the
raw physical estimate behind them -- confirmed via live testing this was
exactly what caused a real misfire (gemma2:27b cited a real cyclist's
correct 5 W/kg, log-normalized to "0.159" against a car's power scale, as
grounds for rejecting an ordinary bicycle). review_plausibility now takes
raw_metrics + normalized_scores + metric units, and the prompt explicitly
instructs reasoning from the raw value first. Verified live against phi4:
it now cites the actual raw number and correctly explains why a low
normalized score doesn't mean the estimate or concept is bad.
Repository write methods used in the pipeline's hot path now take an
optional commit=False, and Pipeline defers commits during the fast/
deterministic passes (1, 3, and 2 without an LLM), flushing every 200
combos and on any exit path (finally block covers normal completion,
cancellation, and any other exception). LLM-involving calls (pass 2 with
an LLM, all of pass 4) still commit immediately -- those are slow and
crash-prone and worth protecting per-write; the deterministic passes
aren't, and recomputing them is now measured at under a second for the
full domain rather than worth 8,000+ individual fsync'd commits. Full
2,970-combination domain run: multiple minutes -> 0.91s. Test suite:
~70s -> ~15s.
Also fixes two more issues found while auditing the estimator for a real
run: CARGO_KG_PER_STRUCTURAL_KG was 500 (no real vehicle carries 500x its
own structural mass in cargo -- a magnitude bug, not a modeling choice),
corrected to 2.5. And space/rocket platforms' range_fuel now reports the
domain's ceiling instead of an arbitrary placeholder constant -- vacuum
coast isn't resistance-limited, so "distance before running out of fuel"
isn't a meaningful question for these the way it is for ground/air/water
vehicles; the real constraint is delta-v budget, a different metric this
pass doesn't model.
Validated with a live full-domain run (phi4, real Ollama calls): 115
reviewed, 0 malformed/null reviews, 0 verdict-vs-status mismatches.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Constraint resolver: aggregate mass/footprint across a combo instead of
pairwise-only checks, treat medium/atmosphere as agreement not supply/demand,
reduce multi-provider checks by best/sum instead of AND-ing every provider,
fail closed on unrecognized mutex values, add a propulsion-viability
(thrust-to-weight) rule. Seed data updated to match (nuclear/solar-sail
footprint floors, water-medium exclusions, explicit ground/gravity providers).
Domain metric units were stored globally per metric name instead of
per-domain, silently corrupting cost_efficiency for every domain but the
first one seeded — fixed with a schema migration.
Stub estimator's cost_efficiency/safety/availability/reliability were a
backwards formula and flat constants; replaced with heuristics grounded in
each entity's thrust_profile/energy_form/infrastructure.
LLM estimate_physics() now receives each metric's unit and expected range
instead of a bare name, fixing wildly miscalibrated estimates traced back to
the prompt's own hardcoded example anchoring the model to the wrong order of
magnitude. Sharpened the safety-estimation and plausibility-review prompts.
Deduped provider parsing logic into llm/parsing.py.
Web pipeline form can now pick an LLM provider per run instead of only via
server env var.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- physcom core: CLI, 5-pass pipeline, SQLite repo, 37 tests
- physcom_web: Flask app with HTMX for entity/domain/pipeline/results CRUD
- Docker Compose: web + cli services sharing a named volume for the DB
- Clean up local settings to use wildcard permissions
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
