Training that uses the weight budget as a capacity-achieving code can make calibration machinery unnecessary; this lab measures when that statement holds.
Start with the glossary, curated findings, reproduction walkthrough, and living verdict ledger. The measured-history appendix keeps supported earlier mathematics and physics results without making this page chronological.
Citation policy: name the exact repository commit SHA and the exact verdict
entry in docs/RESULTS.md that supports the claim. The
ledger is living, so an unpinned citation is not reproducible. Repository
metadata is in CITATION.cff.
llmopt is a mathematics and physics research lab organized around
executable instruments, declared comparisons, and oracle-checked readouts.
The main paths are:
llmopt/search/— symbolic derivation search with explicit rewrite rules, structural and learned evaluators, proposal policies, transposition memory, and verification at the boundary.llmopt/mathgen/— seeded generators for calculus, linear algebra, ordinary differential equations, mechanics, and proofs, with symbolic checks built into generation or evaluation.llmopt/quantum/— model-Hamiltonian ground-state instruments and a verified ZX-graph search path for circuit reduction.llmopt/train/andllmopt/intmath.py— closed-system model births, exact integer primitives, controlled diets, and training interventions that can be compared trajectory by trajectory.llmopt/quantize/— weight diagnostics, sensitivity probes, closed-form allocation, packed crystal artifacts, and the capacity meter that selects which allocation regime deserves testing.llmopt/eval/— equivalence, calibration, latency, and statistical instruments used as supporting readouts rather than substitutes for capability gates.
Experiments are recorded as named arms and cells. A gate accepts or rejects a
declared result; a pin fixes an artifact or trajectory contract. Those words,
the maturity labels, and the controlled scope vocabulary are defined in
GLOSSARY.md.
Install the editable package and development dependencies with:
python -m pip install -e ".[dev]"The status board, theory map, riff ledger, handoffs, and machine-readable
index remain living authority surfaces: docs/BOARD.md,
docs/THEORY.md,
docs/RIFF-LEDGER.md,
docs/handoffs/, and
docs/results-index.jsonl.
-
[SINGLE-SEED] [REGIME-SCOPED: at-capacity house crystals]The house crystals tested at capacity could be packed from weight sigma alone, with no calibration data or allocator search, while remaining inside their declared capability gate. This is a Mac MPS,n=1-per-crystal result, not a universal quantization rule: PACKED CRYSTAL C0+C1 VERDICT indocs/RESULTS.md. -
[SINGLE-SEED] [DEVICE-SCOPED] [FORMAT-BOUND] [REGIME-SCOPED: Qwen2.5-0.5B]THE SIGMA-PACK CLAIM DOES NOT TRANSPORT TO Qwen2.5-0.5B. On the registered RTX 3080 fake-quant arm, max-anchored and calibrated grids exploited weight-tail structure that the house crystals did not have. The boundary is one model, one device, andn=1; it is not a law of web-trained dense models: PACKED CRYSTAL C6 VERDICT and PACKED CRYSTAL C6c VERDICT indocs/RESULTS.md. -
[REPLICATED] [DEVICE-SCOPED] [FORMAT-BOUND] [REGIME-SCOPED: measured deployment artifacts]The packed integer GEMM path reproduced its digest across the registered Mac MPS and RTX 3080 CUDA devices. This independent-device route covers one house crystal's integer GEMM path, not a full integer end-to-end decoder: PACKED CRYSTAL C4 VERDICT and R-PASS VERDICT indocs/RESULTS.md. -
[NULL] [DEVICE-SCOPED] [FREE-RUN-GATED] [REGIME-SCOPED: specified diet and recipe]Removing scaffold-token loss did not repair the registered gravmoe capability gate; the arm degraded while its format diagnostics remained intact. This is a Mac-localn=1null: VERDICT SOL-ADOPTION-1 indocs/RESULTS.md.
The broader closed-system record, including positive results, nulls,
retractions, and amendments, is curated by evidence maturity in
docs/FINDINGS.md. Historical mathematics and physics
measurements that are still supported but are not part of that main arc are in
docs/MEASURED-HISTORY.md.
RJOB_LOCAL=1 python -m llmopt.reproduce gravmoe-rb1On the current Mac-local contract, PASS means the final training-trajectory
digest exactly matches the committed gravmoe-rb1 pin. The adopted command is
booked by VERDICT SOL-ADOPTION-1 in docs/RESULTS.md;
repository commit 4ef9cd511369023d69db7332aebf36517de62951 made the path self-contained from committed windows.
Trajectory agreement is not oracle correctness. It certifies the pinned weight path and teacher-forced training readouts. Free-run symbolic solve scoring additionally needs the uncommitted diet row text and its oracle; the artifact-backed gate arms therefore run in an explicit trajectory-only mode.
The walkthrough in docs/REPRODUCE.md explains setup,
expected output, the full pin registry, and where oracle-backed correctness
claims live. Use python -m llmopt.reproduce --list to inspect the registry;
do not interpret digest equality as a capability score.
The calibration-free packing law is established only for the measured
at-capacity house crystals. The negative PACKED CRYSTAL C6 VERDICT is
equally narrow: Qwen2.5-0.5B, one registered RTX 3080 device, fake quantization,
and n=1. More models, independent implementations, and capability-gated
deployment artifacts are needed before either boundary can move.
Many training comparisons remain single-seed and device-scoped. The README
keeps those fences visible; docs/FINDINGS.md carries the
current maturity labels, and docs/BOARD.md separates live
work from closed results.
Trajectory reproduction still stops short of self-contained free-run oracle scoring because the row text is not committed. Until that input can be shared lawfully, the public artifact proves the trajectory and teacher-forced readouts only.
The mathematics and physics charter is intentionally narrow. Candidate work
belongs in docs/RIFF-LEDGER.md; uncertainty becomes a
claim only after a declared arm, a named verdict, and the applicable regime,
device, format, and replication fences are booked.
Licensed under Apache-2.0.