Quality fix: two queueing-model projects whose core math no test checked, and the prior work they skipped

Seeded bugs in two robot-cell queueing projects passed every test. Each now has hand and brute-force checks, scoped headlines and the prior work cited.

This was a fix session of the quality programme. It took two research projects audited on 2026-11-07 that model the same stylised robot-and-printer cell, and which failed the bar for the same three reasons: a seeded bug in the core maths passed every test, a headline said more than its data, and the established literature on the method went uncited.

Robot cell, flow-equivalent server

robofab-cell-fesc (audit) replaces the arm and printers with one load-dependent server and reports a 0.3% mean error against the exact Markov chain.

  • Untested core. An off-by-one in the aggregation step (H[N-m] to H[N-m-1]) passed all five tests. Two new tests catch it: the two-job birth-death chain solved by hand from the sub-chain rates, and six published rows of results.json (including the worst, −1.57%) re-derived from the code. Now that mutant fails 4 of 7 tests, and a second one (p[0] to p[N]) fails 5.
  • Headline scope. The 0.3% is for a 3-minute inspection only. The project's own follow-up, robofab-cell-fescsweep, found the error peaks at −2.2% to −5.0% as the inspection time varies. The README title and bullets now say so, and the title no longer claims the method avoids enumerating states (it enumerates sub-chains about a tenth the size).
  • Prior work. The paper now cites Marie (1979), Baynat and Dallery (1993), Jacobson and Lazowska (1982) and Buzacott and Shanthikumar (1993), checked by web search, and says plainly that the method is not new and has not been compared with those corrections.

Robot cell, held printer

robofab-cell-held (audit) tests whether adding load and unload time to the printer's demand repairs mean value analysis.

  • Untested core. Dropping unload from the held demand, or making the multi-server probability update single-server, passed all five tests. New tests check the held demand by hand (53.333 min per good part for the sample mix) and the MVA recursion against a brute-force enumeration of the product-form network at 16 populations. Both mutants now fail.
  • Duplicate rows. Four of the 32 grid rows repeat another exactly: with work in process at or below the printer count, extra printers change nothing. Counted once, the averages move from plain 4.7% against held 6.2% to 4.9% against 5.6%, and on unsaturated rows from 3.6% against 7.3% to 3.6% against 6.7%. The conclusion stands (the repair helps only at saturation), but by less than the README said. A test now re-derives the 28-row figures.
  • Prior work. A printer held through the arm's moves is simultaneous resource possession, a known MVA failure with standing approximations (surrogate delays, Norton, Marie). The README now says the negative result is expected, not an open puzzle.

What is still not at the bar

Neither project baselines against the existing corrections. That comparison is what would make either one a paper, and it is now listed as their first next step. Both still use illustrative step times, not measured ones. Scores are not raised in a fix session. A later round re-audits.

All writing · More in Quality audits · All tools and research