The second edition: the vessel compiled, and the repository as a checked instrument - #18
Merged
Conversation
… byte for byte, on macOS and Linux Regenerating the tree on a second machine found four things that did not come back the same. Figures 2 to 14 had been drawn by matplotlib 3.9.4 and every one of them differs under the current release; the pair table photon/pairs.csv.gz carried the gzip timestamp of the run that wrote it, so it differed by eight header bytes on every run with identical content; the finite difference check of the adjoint compiler used a step of 10^-6, where rounding rather than truncation sets the residual, and read 9.1e-07 on one BLAS and 7.8e-07 on another; and two figures differed between operating systems, the Siegert curve because 1 + erf(u) cancels catastrophically for large negative u and the two libms round it differently, and the ampoule figure because its heat map was an embedded raster whose PNG encoding is not portable. requirements.txt now pins numpy, matplotlib and markdown, and every figure is redrawn under the pin. The two compressed tables are written with a zero mtime (levels.json.gz rewritten once with identical content). The finite difference step is 10^-4, where the check is truncation limited and reads 1e-08 everywhere; the theory supplement and the transport README quote the new figure. The Siegert integrand uses erfc(-u), and the ampoule's Green's function and its colour bar are drawn as vector cells. Under the pin every results file, table and SVG figure now reproduces byte for byte on macOS and on Linux; the raster PNGs still differ by a few pixels between platforms and are the one output not compared. The neutron documents and the generated candidate table also write their isotopes as superscripts (252Cf, 137Cs, 99Mo) as the rest of the work does.
…, the exact law at 64 sites, and the timing closure Open Problem 4 asked for the transport level Green's function of a real geometry and for the compiler of theory Section 11 to be run on it. /ampoule measured the 64 x 64 matrix; ampoule/compile.py now takes it as the fabric and does what a compiler does with the digital twin's instance restricted to the vessel's own lattice, eight rings of eight sites. The fabric is geometry: every measured coupling is the solid angle a 2 mm cell subtends at its neighbour times a 3 percent chance of interacting in 2 mm of plastic, the same probability for the ring, axial and diagonal classes; CsI cells take the receiver to its ceiling by the measured 31x. Way B places the 120 intended bonds exactly by apertures, and the 60 diagonal and far couplings each site receives, which pass no aperture, sum to 7.9x the instance in weight units and bury it: the sampled law saturates. The compiler's last pass, subtracting the predictable mean of the crosstalk from every bias through the collar, leaves a fluctuation 3.1x the intended drive. A diffusive shell cannot carry a nearest neighbour instance; the sight lines of the build note's channel plates are the synapse and the next transport must include them. The lattice's topology makes its exact Boltzmann law a transfer matrix over 256 ring states, and the 64 site twin reproduces it to counting noise at 3.3 sweeps per independent sample, the 8 site cost unchanged. Then the timing closure. A weight in a rate coded machine is a photon current, I_kj = G_kj r_j, and the precision law prices its read: a nearest neighbour bond carries 16.3 photons per second with CsI cells, a site's fan in 291, a 4 bit loop takes 881 ms, and the vessel as built draws 0.34 independent samples per second against the 38,462 the machine note priced on its proposal budget alone. The proposal budget and the synapse budget are different budgets, and the second binds by five orders of magnitude. Two measured levers (CsI at both ends, 964x in current) and two catalogue levers (cells packed face to face, 148x in solid angle; a terabecquerel core, 1,000x) reach 51,143 per second at 4 bits, 1.3x the priced figure, with no new physics. The README gains a section and figure 18, Open Problem 4 is marked done with what it asks for next, theory Sections 3.1 and 11.5 carry the synapse term, and the machine note, the build note, the ENIAC ledger, the ampoule and simulator READMEs and the transistor note are amended where the numbers say so. The same documents write 252Cf as a superscript and quote the adjoint check at its new step.
…entity gate in CI, the Dockerfile, the manifest, the glossary, the changelog reproduce.py runs every step in the order build.yml does and, with --check, fails unless every tracked file other than a PNG came back the same; build.yml runs the compile step, the tests, and the same gate on every push, so the claim that everything regenerates byte for byte is now enforced rather than stated. tests/ (48 tests, two seconds): the headline numbers the documents quote are read back from the files that compute them, which is the class of drift the 1.0.1 erratum corrected; every relative link and image resolves and every figure shown is drawn by a script; the house style holds (no dashes in prose, isotopes as superscripts, no edition notes inside the documents, the licence header on every script); the instruments are checked against things that are not themselves (the beta spectra against ICRP 107, the transfer matrix against brute force enumeration, the sampler against its exact law, the degree checker against the theorem, the adjoint Jacobian against finite differences, the committed fission matrix against the committed eigenvalue, the measured synapse against the solid angle ceiling); the reproduce script and the CI run the same steps; the requirements are pinned; the SVGs carry no raster and no date; the compressed tables carry no timestamp; and the committed inputs match data-manifest.sha256. The Dockerfile is the OpenMC 0.16.0 environment the neutron README referred to, on the x86-64 platform conda-forge builds it for. DATA.md names every input and its source. GLOSSARY.md indexes the terms the work coins to the places they are defined. CHANGELOG.md records the three releases, and CITATION.cff carries the version, 1.1.0.
…ame on every platform Regenerating the committed tree in a Linux container found one file still moving: the finite difference check of the adjoint Jacobian read 9.6e-09 on macOS and 2.8e-09 on Linux. The residual at a step of 10^-4 is a few parts in 10^9, and its digits are set by which BLAS inverted the 144 by 144 transport matrix. The check now reports the bound the residual satisfies, below 10^-7, which is the claim the theory makes (machine precision) and which every platform agrees on; the supplement, the transport README and the changelog say so.
…con Mac Verifying the recipe on this Mac: the conda solve and install succeed under x86-64 emulation, openmc --version reports 0.16.0 inside the image, and numpy imports once the emulated OpenBLAS is pinned to one thread; without the pin the import spins. The header now says so, and leaves the threads alone on a real x86-64 host, where OpenMC uses them.
…rom them The currents of the timing closure take one photon out per interaction as the most a site can send, which is a ceiling; the scatter share of the cell material sits below one. The sample rates in the table and the top rung of the ladder are therefore ceilings as well, and the section, the README, the machine note and the build note now say so: the priced figure is within reach of the catalogue levers rather than reached.
…d it The two neutron tally files recorded the cross_sections.xml path of the machine that ran OpenMC. The field now names the library itself, the official ENDF/B-VIII.0 distribution that neutron/data.py fetches and verifies, which is the provenance that matters; the manifest follows.
When the tree does not come back the same, the CI step and reproduce.py now print the diff of the text outputs that changed, so a number that moved on one platform can be read off the log instead of guessed at.
G_fab times G*/G_fab is G* up to rounding, which is exactly zero on one architecture and a few parts in 10^16 on another; the CI gate on x86-64 caught the difference. The line now reports the bound the identity satisfies, below 10^-15, which reads the same everywhere.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this adds
The vessel, compiled. Open Problem 4 asked for the compiler of theory Section 11 to be run on the Green's function a real geometry gives, instead of on the toy fabric.
ampoule/compile.pyruns it on the 64 × 64 matrix thatampoule/measured, with the digital twin's instance restricted to the vessel's own lattice of eight rings of eight sites. The fabric turns out to be geometry: every coupling is the solid angle a 2 mm cell subtends at its neighbour times a 3 percent interaction probability, and CsI takes the receiver to its ceiling. Way B places the 120 intended bonds exactly by apertures, and the 60 couplings each site receives without an aperture sum to 7.9× the instance in weight units and bury it; the collar can subtract their mean and not their fluctuation, which is 3.1× the intended drive. The 64 site sampler is checked against an exact transfer matrix over the 256 ring states and needs 3.3 sweeps per independent sample, the 8 site cost unchanged. Then the timing closure: a weight in a rate coded machine is a photon current, a nearest neighbour bond carries 16.3 photons per second with CsI cells, reading a site's fan in to 4 bits takes 0.88 s, and the vessel as built draws 0.34 independent samples per second against the 38,462 the machine note priced on proposals alone. The currents are ceilings, so the rates are too. Two measured levers and two catalogue levers carry the figure back within reach.The README gains the section and figure 18, Open Problem 4 is marked done with what it asks for next, theory Sections 3.1 and 11.5 carry the synapse term, and the machine note, the build note, the ENIAC ledger and the ampoule, simulator and transistor documents are amended where the numbers say so.
The repository as an instrument.
reproduce.pyruns every step in the CI's order and, with--check, fails unless every number, table and vector figure comes back byte for byte;build.ymlruns the same gate, the compile step and the tests on every push.tests/holds 48 tests: the documents' headline numbers read back from the files that compute them (the class of drift the 1.0.1 erratum corrected), every internal link resolved, the house style enforced, the instruments checked against things that are not themselves, the reproduce script and the CI kept in step, and the committed inputs matched againstdata-manifest.sha256.requirements.txtis pinned, and under the pin every output reproduces on macOS and Linux. ADockerfilecarries the OpenMC 0.16.0 environment the neutron README referred to;DATA.mdnames every input and its source;GLOSSARY.mdindexes the terms the work coins;CHANGELOG.mdrecords the releases;CITATION.cffcarries the version, 1.1.0.What it corrects
Regenerating the tree on a second machine found four outputs that did not come back the same: figures 2 to 14 had been drawn by an older matplotlib, the pair table carried a gzip timestamp, the adjoint check's digits depended on the BLAS, and two figures differed between operating systems (a cancellation in the Siegert integrand, an embedded raster in the ampoule figure). All are fixed; the raster PNGs remain platform specific and are the one output not compared. The two neutron tally files named the directory of the machine that ran OpenMC as their cross section provenance and now name the library. Hyphenated isotope designations in the transport documents are written as superscripts, as the rest of the work does.
Checks
python reproduce.py --checkpasses on macOS in 45 s; the same gate and the 48 tests pass on the committed tree in a Linux container under the pinned requirements.