Issues

A smooth narrow pulse (a Gaussian or sech² in a clock) is stepped over: Y = 0 for 2.212 at every tolerance, with no warning

lanl/bngsim#1286 · opened

on Oct 25, 2026

1

The clock-rate column of a narrow kinked pulse stops converging at 2e-5 to 9e-5 from rtol 1e-10 (w <= 0.005), where central differences of plain runs reach 1e-7

lanl/bngsim#1285 · opened

on Oct 25, 2026

A moving size whose rate bngsim could know is refused since #1204: rateOf of a constant rate, a floor that holds, a cycle of constant rules

lanl/bngsim#1281 · opened

on Oct 25, 2026

A rate-law equality the state sits on is judged at the run's start: an event that lands on it mid-run is still silent, and a stretch that outlasts the run is refused

lanl/bngsim#1279 · opened

on Oct 25, 2026

#1133's read-ahead still costs 1.15-1.34x at 20001 rows after #1200 (1.08-1.21x with sensitivities): read ahead only where a trigger can cross

lanl/bngsim#1278 · opened

on Oct 25, 2026

A fixed-time event and a state trigger a column moves, on one instant, fire as two batches: S jumps 3.75 to 4.25 across k = 1 and the run returns dS/dk = 1.25 with no refusal

lanl/bngsim#1277 · opened

on Oct 25, 2026

A sensitivity run stopped by another root within 1e-12 before a sample still records that row before the event (#1182's read-ahead needs t_now == t_out)

lanl/bngsim#1276 · opened

on Oct 25, 2026

.net writer: an assignment-rule species is written as a $-fixed species beside a same-named observable or function; bngsim reads the rule, run_network holds the initial value with no message

lanl/bngsim#1274 · opened

on Oct 25, 2026

amici_sens: judge each sensitivity parameter on its own scale (after #1167)

lanl/bngsim#1261 · opened

on Oct 24, 2026

Proposal: a differential harness that runs each model through every path (sensitivity vs plain, vs central differences, analytic vs FD Jacobian, KLU vs dense, compiled vs interpreted, C vs de_DE locale) and requires the same answer, as a nightly beside the parity checks

lanl/bngsim#1257 · opened

on Oct 24, 2026

1

The analytic Jacobian of a size read through a diamond of rules takes exponential time to set up: 3.1 s at 12 rules, 35.8 s at 16, 53.5 s at 18, where jacobian="fd" builds in under 10 ms

lanl/bngsim#1255 · opened

on Oct 24, 2026

An exact equality on a state, if(Bobs==0, kb, 0), is silently wrong on KLU with the FD Jacobian: Y(10) off by 1e-3 at rtol 1e-8 and 4.7e-3 at 1e-10, where dense and the analytic Jacobian are right

lanl/bngsim#1254 · opened

on Oct 24, 2026

S4 agreement check: Koo2013's final state at rtol 1e-6 swings with the output grid in every engine, so its 5% verdict is luck (bngsim 4.8% off RoadRunner today, 1.1% in July)

lanl/bngsim#1250 · opened

on Oct 24, 2026

Under a comma-decimal LC_NUMERIC, the .net loader refuses `0.5*kf` and reads a group factor of 0.5 as 0: its number parsing follows the C locale

lanl/bngsim#1235 · opened

on Oct 23, 2026

Make the full suite a clean signal under BNGSIM_CODEGEN_JIT=mir: 93 forced-MIR failures hide new ones (cc-artifact tests, engine-name asserts, #1212, #1213)

lanl/bngsim#1217 · opened

on Oct 22, 2026

MIR-JIT'd code dies with SIGILL under Rosetta (x86_64 Python on Apple Silicon) on some models, at -O2, on the old MIR pin and the new

lanl/bngsim#1216 · opened

on Oct 22, 2026

MIR JIT: each context now reserves 128 MB of address space for code; under RLIMIT_AS (ulimit -v) that caps a process at ~60 live JIT'd models per 8 GB

lanl/bngsim#1215 · opened

on Oct 22, 2026

log10(p) or log2(p) in a derived rate constant declines the analytic sensitivity RHS for the whole model, which falls back to the difference quotient

lanl/bngsim#1214 · opened

on Oct 22, 2026

codegen copies ExprTk functions, table calls and implicit products it has no C for into the source: cc pays a failed compile, the MIR JIT refuses

lanl/bngsim#1212 · opened

on Oct 22, 2026

ODE runs with stops: the look inside flat stretches (#998) misses steps inside a call to an output row, stretches with a delayed event pending, the stretch after a stop that restarts nothing, and stiff models

lanl/bngsim#1192 · opened

on Oct 22, 2026

Compartment sizes that read each other through conditions still run silently where #1134 does not refuse: NaN when the conditions read the sizes by name, and one of two consistent states for a ring through a value and a condition

lanl/bngsim#1185 · opened

on Oct 22, 2026

NFsim: a linear table function indexed by time is held between its x values, so a ramping rate lags where little fires

lanl/bngsim#1179 · opened

on Oct 21, 2026

SSA, and rm on a model that reads the clock, lose waiting times to rounding at a large absolute t_start (z = -287 at 1e12, silent)

lanl/bngsim#1059 · opened

on Oct 15, 2026

2

The differenced state Jacobian steps every species by 1.5e-8 of the largest concentration: dY_ss/dp with jacobian="fd" is 7% off beside an unrelated species at 1e7 and 44% at 1e8, with no error

lanl/bngsim#1002 · opened

on Oct 9, 2026

2

Tracking: forward-sensitivity runs that are refused where a derivative exists (refusals to lift)

lanl/bngsim#961 · opened

on Oct 1, 2026

4

Navigation Menu

Repository navigation

Issues

A smooth narrow pulse (a Gaussian or sech² in a clock) is stepped over: Y = 0 for 2.212 at every tolerance, with no warning

The clock-rate column of a narrow kinked pulse stops converging at 2e-5 to 9e-5 from rtol 1e-10 (w <= 0.005), where central differences of plain runs reach 1e-7

A moving size whose rate bngsim could know is refused since #1204: rateOf of a constant rate, a floor that holds, a cycle of constant rules

A rate-law equality the state sits on is judged at the run's start: an event that lands on it mid-run is still silent, and a stretch that outlasts the run is refused

#1133's read-ahead still costs 1.15-1.34x at 20001 rows after #1200 (1.08-1.21x with sensitivities): read ahead only where a trigger can cross

A fixed-time event and a state trigger a column moves, on one instant, fire as two batches: S jumps 3.75 to 4.25 across k = 1 and the run returns dS/dk = 1.25 with no refusal

A sensitivity run stopped by another root within 1e-12 before a sample still records that row before the event (#1182's read-ahead needs t_now == t_out)

.net writer: an assignment-rule species is written as a $-fixed species beside a same-named observable or function; bngsim reads the rule, run_network holds the initial value with no message

amici_sens: judge each sensitivity parameter on its own scale (after #1167)

Proposal: a differential harness that runs each model through every path (sensitivity vs plain, vs central differences, analytic vs FD Jacobian, KLU vs dense, compiled vs interpreted, C vs de_DE locale) and requires the same answer, as a nightly beside the parity checks

The analytic Jacobian of a size read through a diamond of rules takes exponential time to set up: 3.1 s at 12 rules, 35.8 s at 16, 53.5 s at 18, where jacobian="fd" builds in under 10 ms

An exact equality on a state, if(Bobs==0, kb, 0), is silently wrong on KLU with the FD Jacobian: Y(10) off by 1e-3 at rtol 1e-8 and 4.7e-3 at 1e-10, where dense and the analytic Jacobian are right

S4 agreement check: Koo2013's final state at rtol 1e-6 swings with the output grid in every engine, so its 5% verdict is luck (bngsim 4.8% off RoadRunner today, 1.1% in July)

Under a comma-decimal LC_NUMERIC, the .net loader refuses `0.5*kf` and reads a group factor of 0.5 as 0: its number parsing follows the C locale

Make the full suite a clean signal under BNGSIM_CODEGEN_JIT=mir: 93 forced-MIR failures hide new ones (cc-artifact tests, engine-name asserts, #1212, #1213)

MIR-JIT'd code dies with SIGILL under Rosetta (x86_64 Python on Apple Silicon) on some models, at -O2, on the old MIR pin and the new

MIR JIT: each context now reserves 128 MB of address space for code; under RLIMIT_AS (ulimit -v) that caps a process at ~60 live JIT'd models per 8 GB

log10(p) or log2(p) in a derived rate constant declines the analytic sensitivity RHS for the whole model, which falls back to the difference quotient

codegen copies ExprTk functions, table calls and implicit products it has no C for into the source: cc pays a failed compile, the MIR JIT refuses

ODE runs with stops: the look inside flat stretches (#998) misses steps inside a call to an output row, stretches with a delayed event pending, the stretch after a stop that restarts nothing, and stiff models

Compartment sizes that read each other through conditions still run silently where #1134 does not refuse: NaN when the conditions read the sizes by name, and one of two consistent states for a ring through a value and a condition

NFsim: a linear table function indexed by time is held between its x values, so a ramping rate lags where little fires

SSA, and rm on a model that reads the clock, lose waiting times to rounding at a large absolute t_start (z = -287 at 1e12, silent)

The differenced state Jacobian steps every species by 1.5e-8 of the largest concentration: dY_ss/dp with jacobian="fd" is 7% off beside an unrelated species at 1e7 and 44% at 1e8, with no error

Tracking: forward-sensitivity runs that are refused where a derivative exists (refusals to lift)