Updates and corrections

Notable changes to the solver, the experiments, and any number this project reports. Corrections are listed as prominently as features: where a published figure changed, this says what it was and what it became.

0.1.6 (2026-08-29)

Superseded releases

Earlier tags are left where they are rather than moved, so the record of what each published survives.

• 0.1.0 published an E5b design that was scored but was not the best available: the correction it shipped discarded Adam’s final update, which the loop never scored and which is better. Its three E5b figures were never regenerated after that correction, so they drew a still earlier design’s numbers, and its gradient-checker sweep was sampled without a seed.

• 0.1.1 shipped e6_generalization.py with a NameError, so a documented README command crashed, and the same three stale figures.

• 0.1.2 to 0.1.4 are superseded by the corrections below.

Corrected

Every entry below is a correction made while building this entry, newest first, written as “it was X, it is now Y”. They are a record of what moved, not a statement of the current values: several entries quote figures that later entries supersede, and one entry records two corrections of its own that were made and then retracted. For what the entry publishes today, read docs/RESULTS.md, which is generated from the committed artifacts, or the artifacts themselves.

• The withdrawn claim turns out to be bounded by dimension, and the bound is measured. The control that withdrew “E5b is where the gradient pays” ran at 24 design variables, which is where CMA-ES is comfortable. Re-run at 93 on a design space nested by construction, so that the 24-parameter design is an exact interior point of the larger box and the best achievable loss cannot be worse, the ordering reverses: CMA-ES is 0.631x Adam at 24 and 3.773x the other way at 93, and giving CMA-ES three times the budget still leaves it 2.71x behind. The concession stands at 24 variables and does not extend to 93. It is two dimension points on one problem instance; the objective’s 48 residuals cap its local rank whatever the parameter count is.

• Three things in that experiment were designed to stop it flattering us, and two of them fired. A pre-registered middle point at 47 variables was dropped on its own gate rather than accommodated, because the embedded warm start moved the objective 10.1% there against a 5% threshold. CMA-ES’s best configuration at 24 sat on the edge of its step-size grid, so the grid was extended one step, which could only have strengthened the concession against us; it did not improve. And the first design of the experiment, at 96 and 192 variables with bump width held fixed, was abandoned before it ran: the dictionary’s numerical rank saturates near 52 at both, so those are the same problem to precision, and uniform maximum amplitude builds terrain metres taller than the launch height. That version would have manufactured the result it was testing for.

• The score Tesseract was removed from the gradient arm’s objective in the new control. It cost Adam an apply and a VJP per particle per iteration that the budget charged nothing for, and the size of that undercharge shrinks as the parameter count grows, so it would have biased the cross-dimension trend.

• The demo video’s optimizer card contradicted the figure printed above it. The card renders study_optimizers.png, which plots five-seed medians, but its caption was built from the single-run E5 values, so it said Nelder-Mead ends at 1.9e-02 while the line in the same frame visibly ended near 1e0. The caption reads the benchmark medians now, and the gap it states is 6.5 orders rather than 4.9. The error understated the result, but the card was the one place in the video that disagreed with its own image, and the README embeds a poster tile of it.

• Three counts of the studies and experiments were live at once. The landing page said seven and eight, the studies page six, the experiments page seven, and the changelog named “seven and six” as the pair that was wrong. All now say seven studies and eight experiments, and the studies page records that only six of the seven write an artifact the collector reads.

• The studies page still published the withdrawn wall-clock claim in its figure alt text, which is what a screen reader receives. The body text twenty lines below already stated the reversal.

• The E5b control table bolded CMA-ES’s fifth-percentile cell as a winner. The Adam learning-rate sweep reaches +0.600 m against CMA-ES’s +0.558 m, and a paired bootstrap on the shared draw puts the difference at +0.039 m with a 95% interval spanning zero. CMA-ES wins the objective; on the tail the two do not separate, and the page says that now.

• docs/writeup.md claimed “none is typed by hand” of a hand-written markdown file. The headline numbers in it are pinned to results.json by the drift suite, which is the true and narrower statement.

• The same file called the shipped method “the saltation adjoint”, colliding with the reverse-mode adjoint this project did not build, and said grid-reset autodiff returns a gradient that is “exactly zero”. That is true only on flat terrain; E5b optimises bump amplitudes on curved terrain, where the reset normal depends on the design and the gradient is nonzero and wrong.

• Smaller: the disk breakdown summed to 5.1 GB under a stated 9 GB total; the repro file’s skip arithmetic said 40 of 49 where it is 45; a drift-test docstring said CI does not install Sphinx, which it has since done; the figure script’s docstring said three figures where it writes six; the artifacts page credited scaling_repeats.json to a script that does not write it; and the three scripts backing the section 4c withdrawal had no run command anywhere in the README.

• The README’s performance table failed its own arithmetic. The refactor commit rewrote the ratio row and left the two rows above it at their pre-refactor timings, so the table printed 11.4 ms over 1.8 ms and labelled it 3.3x. Every column was wrong the same way, the VJP figures were about four times too large, and docs/RESULTS.md, which is generated, already published the right ones. The most checkable numbers in the entry disagreed with the repo’s own results page.

• A sentence the changelog listed as corrected was still standing. The README kept explaining the gradient charge as “measured against the 24 columns its objective actually differentiates”, which the previous entry retracted and study_optimizers.py retracts in its own docstring. Anyone who read the changelog and grepped the README found a correction claimed but not made.

• The arguments the refactor invalidated were rewritten rather than renumbered: the wall-clock section that called the ordering unresolved, the scaling sentence that said the measurement made a reverse-mode extension “a measured argument rather than a preference”, the O(n_params) framing in section 2, and the Future Work item that led on parameter count. Future Work keeps the adjoint on the justification that survives, large state dimension, and says plainly that the one we started with did not.

• The demo video baked in the old 32-second run time and the pre-refactor optimizer figure. Regenerated.

• The gradient was 4x more expensive than it needed to be, and a study asserted that it should be. The flow is affine in the state, so the RK4 variational update collapses exactly to a fixed 4x4 tangent map; the solver was instead calling ForwardDiff.jacobian four times per step to rebuild the same constant matrix and running RK4 on the full sensitivity matrix. Composing the tangent map across a smooth segment and applying it only at events takes the parameter dimension out of the inner loop. A gradient now costs 1.61 forward solves rather than 6.78, the scaling slope is 15 microseconds per parameter rather than 93, and at 581 parameters the VJP is 1.8x a forward solve rather than 7.9x. The primal is bitwise unchanged and the sensitivities agree to 6.5e-14 on every status branch.

• The wall-clock concession has reversed, and that is about us, not about gradients. With the gradient at 6.78 solves, CMA-ES was ahead on 4 of 5 seeds and we reported the ordering as unresolved. At 1.61 solves Adam is ahead on all five. We made our own gradient cheaper and the accounting followed; the pages now say so rather than presenting it as a finding about methods.

• The E5b control was re-run, because its budget derives from that charge. The matched budget fell from 14964 solves to 5040, leaving gradient-free 34% of what it had. CMA-ES still wins: 7.36e-03 against Adam’s 1.83e-2, held-out 200/200, fifth percentile +0.558 m against +0.426 m. The withdrawal of “E5b is where the gradient pays” therefore survives a threefold tightening of the baseline’s budget.

• A README sentence claimed the 6.8 charge was “measured against the 24 columns its objective actually differentiates”. It never was: vector_jacobian_product calls _run(want_sens=True) and slices afterwards, so it builds every column whatever the cotangent asks for. The charge was also published in prose and collected nowhere, so no guard could reach it.

• “E5b is where the gradient pays” was false, and is withdrawn. It was the entry’s answer to its own most awkward measurement, that Nelder-Mead matches Adam on held-out purity, and it rested on nobody having run a gradient-free method on the ensemble objective. Running it: at a budget matched in particle solves, CMA-ES reaches an ensemble design at 5.77e-3 against Adam’s 1.83e-2, with a fifth-percentile margin of +0.657 m against +0.426 m, on every one of nine grid runs, its median still beating Adam’s best. Because that gave CMA-ES nine configurations and Adam one, Adam was then swept over five learning rates fixed before the run; the ordering is unchanged, and the published lr = 0.004 turns out not to be Adam’s best either (lr = 0.001 reaches 1.21e-2). The sentence is gone from the README, the studies page and the write-up, replaced by what was measured: optimising the ensemble objective is what buys the robustness, and the gradient is not what makes that possible. Nelder-Mead is genuinely worse, so this is not a gradient-free win in general.

• Five published statistics had no code computing them. The four per-design rows on the studies page carried Wilson and bootstrap intervals that existed only as typed text, and the jitter-sweep McNemar p was quoted in two places with nothing producing it. study_design_table.py generates the table into an artifact the suite compares the page against, and the McNemar is computed where the sweep is. One published bound moved: Nelder-Mead’s upper interval is 0.52, not 0.51.

• The repeat-run figures were quoted from runs that were never recorded. A slope “re-measured at 81 µs”, a gradient charge of “5.5 forward solves” and a “ratio of medians of 1.3x” appeared across three pages with no artifact behind any of them. The scaling study is run twice now and both runs are committed: 93.1 µs and 83.0 µs, an 11% spread. The two figures that had no second run are gone rather than restated.

• make_e5_figure.py restated the entire E5 configuration instead of importing it, and its copy had silently dropped v_stop, so the published figure re-solved under the schema default rather than the value the design declares.

• The entry has eight labelled experiments and seven verification studies, not seven and six.

• A published README command crashed. e6_generalization.py used BIN_RUBBER and BIN_PET seven lines above the import that defines them, so it died with a NameError before doing anything. It is the seventh command a reader following the README types, and it shipped because CI ran the tests, the validators and the fuzzer but never an experiment. The four fast experiments run in CI now.

• Three figures published the numbers of a withdrawn design. The artifacts were restored and the prose rewritten, but study_robustness.png, design_comparison.png and e5b_purity.png were never regenerated, so the first read “983/1000 vs 1000/1000” against an artifact saying 997, and the second drew the ensemble design at 0.37 m below Nelder-Mead while the sentence above it said 0.43 m and “the widest tail of the four”. Two guards checked that a figure names an artifact and that the artifact exists; neither checked the picture was current. A test regenerates the artifact-only figures and compares bytes now.

• The 0.1.0 note had the defect backwards, in three places. The changelog, the annotated tag and the release body all said 0.1.0 shipped “a design the optimizer never scored”. It did not: 0.1.0’s design was scored, and was the best point inside the loop. The unscored one was Adam’s final update, which 0.1.0 discarded and which is better. The design shipping now is bit-identical to the one predating both corrections.

• The README’s justification for the gradient-checker settings was false. It claimed the run uses “distinct entries” against the CLI default’s repeats. The checker samples with replacement whatever the budget, and the payload has 14 differentiable elements over two output paths, so 150 checks are about 76 distinct comparisons. Seeding fixed repeatability, not representativeness: other seeds give 12, 20 and 21 failures at rtol = 1e-5 where this one gives 14. Both are stated now, and the seed is disclosed.

• The capture script could destroy the artifact it exists to produce. It wrote the file and then asserted the run was clean, so a failing run replaced the committed artifact with its own failures before raising. It validates first now, checks every endpoint reported, sums the sweep denominator from the rows it parsed rather than multiplying one of them, and fails loudly instead of silently dropping a tolerance that produced no output.

• Whether a degenerate crossing raises was documented two contradictory ways, and both were wrong. Tangency cannot reach the guard: the denominator floors near 1e-7 on metre-scale terrain, five orders above the 1e-12 threshold, and chatter cannot either, because the re-arm lift floors the approach speed at 1.4e-6. The guard is reachable, through a launch height small enough that the first impact speed falls under 1e-12. The second guard, on non-approaching velocity, fires only past the RK4 stability limit and is unreachable through the schema. All of that is written down now.

• Light-mode contrast: the sweep widget’s two value colours sat at 2.55:1 and 2.90:1, and the footer provenance, sidebar captions and figure-source lines at 3.26:1 and 3.50:1, all under the 4.5:1 threshold for text. The text tones are darkened; the figure hues are unchanged.

• The artifacts page named “three independent oracles” where the write-up says two solver-independent ones plus a finite-difference gate, and rendered that on the live site.

• Four tests skip without Sphinx, not three; the run time moved 24.8 s to 31.6 s on a busier machine; and make_study_figures.py’s title and study_generalization_stats.py’s docstring were both left describing the withdrawn design.

• The E5b iterate fix was itself wrong, and is withdrawn. E5b’s numbers were corrected twice and both corrections are retracted; this states where they landed. Keeping the best point scored inside the loop discarded Adam’s final update, which the loop never scored and which is 0.4% better on the objective. So the correction published a worse design and moved eight numbers for nothing: the five below plus the five-ensemble McNemar p (1.5e-05 back to 0.0026), the jitter-sweep McNemar p (1.0 back to 0.50), and the median margin (+1.8 cm back to -2.7 cm). The design that ships now is bit-identical to the one that predates both corrections; what changed is that it is scored. The loop now scores the final iterate too and keeps the best of everything evaluated, which restores the original design with real provenance: purity is 997/1000, the worst case -0.35 m, the fifth percentile 0.49 m, the ensemble margin 0.43 m, and the jitter sweep 20 of 20. Every page that was changed for the bad fix is changed back.

• The gradient checker was sampling unseeded. The same sweep re-measured 15, 17, 19 and 21 failures at rtol = 1e-5 across four runs, so a published count was one draw. The seed is pinned and recorded in the artifact now, and two runs at the same seed agree exactly: 14 of 150.

• The headline checks and failures still mixed populations. The sweep was fixed but the headline pair was not: checks counted one endpoint while failures summed three, hidden only because the count is zero. Both are totals now, and the pages say “across the three gradient endpoints”.

• Two tests failed under the install the README documents. They import docutils, which requirements-repro.txt does not carry; CI installs Sphinx separately and hid it. They skip now, and the counts say three tests need Sphinx rather than one.

• The RK4 stability bound was a rounded guess, untested, and misordered. 2.7 rejected a genuinely stable band up to 2.785293563, the real root of z^3 + 4z^2 + 12z + 24; replacing the constant with 999 or with 0.001 left the suite green; and an infinite drag reported a stability violation rather than a finiteness one. The limit is computed rather than typed, straddled by a boundary test, checked after the finiteness loop, and stated in the schema descriptions the reference page is generated from.

• Berkowitz and Canny’s objective is not per-object rather than bulk. They score a per-trajectory metric alongside a feed-rate efficiency taken over all initial orientations. Both the write-up and the related-work page said otherwise, so making them agree had propagated the error rather than fixing it.

• The README’s figure tip named three scripts for what takes five, and called twelve figures eight.

• The median-margin interval was published with no code computing it. The robustness study bootstraps it now, and collect_results.py collects the McNemar p, the tail interval and the median interval it was already storing but never reading.

• The tolerance sweep counted failures over three endpoints against one endpoint’s checks. check_gradients.json recorded 81 failures out of 50 checks at rtol = 1e-7, which is impossible and was the tell. The README published “22 of 50” for a rate whose denominator is 150. The sweep records the endpoint count and both denominators now, and re-measuring gives 15 of 150 at rtol = 1e-5.

• drag and dt were validated independently, never their product. The integrator is fixed-step explicit RK4, so it is stable only while drag * dt stays under about 2.785. Past that a purely dissipative force gains energy, sensitivities change sign, and the run still reports status 0. Rejected at validation now.

• abstract_eval accepted a negative n_samples and returned a shape of (-3, 5), disagreeing with apply on the same input. n_samples is a plain int and stays concrete on the abstract path, so it is checked before the early return.

• The writeup contradicted the repository’s own related-work page on two cited works. DiffTaichi was filed under engines that trade away event-gradient fidelity, when it documents this exact failure and repairs it with a precise time of impact, and is the closest prior art here. Roussel et al. and Berkowitz & Canny were described as differentiating a smoothed model against a bulk flow statistic; both are non-gradient, and Berkowitz & Canny’s objective is per-object. The withdrawn “our contribution is the packaging” claim was also still standing.

• “tuned CMA-ES” described an untuned design. The CMA-ES design E5b scores comes from a single hardcoded sigma0 = 0.05, seed = 3 in e5_separator.py, not from the tuning grid. Only study_optimizers.py tunes.

• The E2b wall time was labelled “sampling” in the writeup and the generated table; it covers warmup plus sampling. The leapfrog count is the sampling phase alone, and experiments.md had paired the two.

• The landing page’s closed-form card floored the exponent, rendering 7.1e-12 as 10^-12 and claiming a tighter agreement than the artifact supports. It shows the measured value, like the oracle card beside it.

• The artifacts.md page said only check_gradients.json needs Docker; e2b_posterior.npz does too. Its regeneration note also credited the wrong commands for e3_rows.npy and timing.json.

• The second landing-page snippet is documented as running as written, but needs jax.config.update("jax_enable_x64", True); the component returns Float64 and JAX refuses without it.

• Test counts read 33 against 36 collected, in the README and getting started.

• The wall-clock optimizer comparison rests on one measured scalar, and that scalar moves. The gradient charge re-measured at 5.5 forward solves rather than 6.8 on a repeat run, which moves the ratio of medians from 2.3x to 1.3x. Every optimizer trace in the artifact is bit-identical between runs; only the timing moved. The pages already reported the wall-clock ordering as unresolved at n = 5, and they now say the charge itself is a measurement with spread.

• The landing page said the four checks take about five minutes; they take about twenty five seconds. The five minutes was the first-run Julia bootstrap, not the checks, and the figure came from no artifact. It is measured now by scripts/time_checks.py, which writes timing.json, and both the stat card and the prose read it.

• The landing page’s headline figures were hand-typed HTML. The drift test covers README.md, docs/writeup.md and docs/RESULTS.md, so the five numbers on the most-read page of the site were the one place a published figure could go stale silently. They are generated from results.json now. The other four were correct.

• The cost-scaling slope was quoted as though it were precise. It is wall-clock on one shared laptop: re-running the study moved 93 microseconds per parameter to 81, and R² from 0.998 to 0.997. The affine shape holds; the absolute figures carry ten to twenty percent of run to run spread, and the pages that quote them say so.

• The tolerance sweep in check_gradients.json was three hardcoded literals. They were written into the artifact whose entire purpose is that its numbers are measured, and the README quoted them. The script runs the sweep now, and running it changed the answer: at the same eps = 1e-6 as the headline run, the checker first fails at rtol = 1e-5 with 22 of 50, not at 1e-6 with 13. The earlier figure came from an ad-hoc sweep at a different eps.

• Four upstream fixes have merged since the last entry. tesseract-core #667 and tesseract-jax #236 are both merged, and issues #666 and #235 are closed; only the juliacall deadlock #234 is still open. The pages said all the Tesseract work was still open.

• A README claim about our own gradient checker was false. It said the check fails at rtol = 1e-5, and this entry claimed it passes there. Both were wrong: that sweep ran at a different eps than the headline configuration. At the published eps = 1e-6 the checker does fail at 1e-5. Superseded by the sweep-denominator entry at the top of this section, which re-measures it as 15 of 150 across the three endpoints.

• The results table published the statistic the prose calls wrong. The seeds are paired, so the ratio of medians (917x) pairs one seed’s CMA with another seed’s Adam; every prose page already used the paired per-seed median (347x) while the generated table led with the unpaired one. The table now leads with the paired statistic and keeps the unpaired one labelled as reference.

• “1.1 to 5.1 orders” was CMA-ES alone, attributed to both gradient-free methods. Nelder-Mead spans 2.9 to 8.0. The published range is 1.1 to 8.0 and the collector computes it over both methods.

• The e3_bias figure drew the saltation reference as a literal 1e-15 line, 669x below what the artifact measures. It plots the measured column.

• E2b sampling time read “18 minutes” in four places against a measured 22.7, and the test count read 21 against 33 collected.

• The NUTS convergence statistic was the outdated one, and the number a reader actually needs was missing. gelman_rubin is the classic non-split R-hat, which cannot detect within-chain non-stationarity; it is split_gelman_rubin now. Effective sample size was never reported at all: 2000 draws buy 344 and 330 effective ones, which is the relevant figure given the sampler cost is a headline. Two chains estimate the between-chain variance on one degree of freedom, so R-hat here is stated as a weak check rather than a passed one. The wall-clock figure timed warmup plus sampling while the leapfrog count covered sampling only; the label says so now.

• The gradient-free designs had never been scored on the metric this project says matters, and doing so weakens a headline. E5 saved the CMA-ES and Nelder-Mead designs and nothing ever loaded them. Scored on E5b’s held-out scatter ensemble: Adam 0.995 purity with a 0.07 m fifth-percentile margin, CMA-ES 0.980 with 0.18 m, and Nelder-Mead 0.995 with 0.41 m: five orders behind on the objective, identical purity, wider margin. In engineering units the 3.9-order loss gap is 0.34 mm against 31 mm with bins 1600 mm apart. (This entry first quoted 12.6 mm, which came from the multi-seed benchmark median rather than the CMA-ES design E5b actually scores.) “The gradient-free methods never reach the design” is now stated as what it is, a statement about objective value, with the translation shown.

• The worst-case margin claim was backwards. It said the worst case moves “from inside the wrong bin to 0.09 m clear”. On the rebuilt designs it goes -0.12 m to -0.35 m, both inside the wrong bin. The ensemble objective buys the low tail, not the extreme, and the page says so now.

• Two statistical methods were wrong even where the answer held. The bootstrap resampled the two margin arrays independently although the same 1000 particles are scored under both designs; it now resamples particle indices and carries both together. Purity is a paired binary outcome, so non-overlapping Wilson intervals were the wrong criterion; McNemar is reported alongside, and gives p = 0.0026.

• The novelty claim was too broad and two thirds of it is withdrawn. A prior-art sweep found that saltation matrices have been used for optimal design explicitly (Kong et al., Proc. IEEE 2024), that a pip-installable package already pairs them with adjoint-gradient trajectory optimization, that SciMLSensitivity.jl implements exact event-time corrections with a bouncing-ball restitution-optimization example, that Diffrax gained pathwise event-time gradients in 2024, and that FMI has carried derivatives across a component boundary since 2014, with adjoint derivatives added for machine-learning consumers in 2022. Tesseract’s own fortran_enzyme example is the same boundary pattern. What is claimed now is only the conjunction: gradient-based design of passive geometry, with exact event times, against a per-object routing objective, with the nearest neighbours named.

• The gradient-checker headline measured a sampler, not a gradient. The CLI defaults to rtol = 0.1 and samples with replacement, so “0 failures / 1574 checks” was mostly repeated comparisons at ten percent tolerance. It now runs distinct entries at rtol = 1e-4: 0 failures / 50 checks on each of the three endpoints. It fails at 1e-5, where the finite difference itself stops resolving, so 1e-4 is the tightest honest setting.

• We were wrong about Diffrax, in its favour. This project claimed the restart-after-event pattern returned solver-dependent wrong gradients, citing three numbers from Diffrax issue #729. Reading the whole thread, the maintainer diagnosed the reproducer as a usage error (the jump time was passed to ClipStepSizeController as a plain float rather than closed over differentiably) and noted the vector field was not valid input; the reporter later said the example may not reproduce what they were chasing. Our own script repeated the same mistake. Measured correctly, Diffrax returns the exact gradient under all three solvers. The comparison now says so, and the remaining difference is that Diffrax has no reset map, which is ergonomic rather than a correctness gap.

• The wall-clock optimizer comparison over-charged the gradient. The charge of 8.5 forward solves came from a VJP over every input, 77 sensitivity columns, but the separator differentiates only its 24 bump amplitudes. The benchmark now measures the charge against the inputs it actually differentiates: 6.8 solves. CMA-ES is 2.3x ahead on the ratio of medians under wall-clock accounting, not the 24x previously published.

• E6 and the generalization study evaluated the design past its own horizon. They redeclared t_final = 3.0 where E5 designs at 2.2, so the separator was scored two impacts beyond the point its output is defined at. At its own trained value e = 0.8 the design lands at 4.3996 against a 4.4 target; evaluated at the longer horizon it read 3.7262, which is the number that had been published. Both now import the design’s configuration instead of restating it. The point design is indecisive under jitter at 2 of 20 restitutions, not 5 of 20, still including its own trained value; at e = 0.85 it is now unanimous where two different samples had disagreed.

• “Reported and fixed upstream” was not true. Two Mosaic harness fixes are merged; the three Tesseract issues and the two PRs against them are open.

• The writeup said the solver “raises an explicit error at degenerate crossings”. It does not: the guard on the saltation denominator is effectively unreachable, and the real behaviour is a large finite gradient and then a silently missed event. Stated as a limitation now.

• Zhong et al. (NeurIPS 2021) was filed under “learned contact models”. Its contact model is analytic; only its parameters are learned.

• The Hiskens & Pai parameter augmentation is equations 13 to 16 with the augmented sensitivity system at 33, not 62 to 63.

• drag and v_stop were unvalidated while e > 1 was rejected for exactly the same reason. Negative drag added energy without bound and was accepted.

• E5 and E4 reported a design they had not scored. Three optimization scripts returned the parameter vector one update past the last point they evaluated, so the objective in the trace belonged to a different design than the one saved, plotted and quoted. They now return the best evaluated iterate. The E5 landing errors move from 0.38 mm and 0.34 mm, which were measured on the unevaluated design, to 0.23 mm and 0.42 mm. The optimizer benchmark’s own loop was already correct, so the 917x eval-accounting figure is unaffected. (The wall-clock figure changed later, for the separate reason recorded above.)

• The performance table did not match its artifact. It listed 4.9 ms and 41.6 ms for 77 parameters where scaling_result.npz records 2.34 ms and 19.82 ms, and labelled the column with the wrong configuration. The table is now generated from the artifact.

• E2b’s diagnostics rested on nothing. The divergence count, r-hat and throughput were prose. The run now records them: 23,440 leapfrog steps over two chains in 23 minutes of warmup plus sampling, 0 divergences, non-split r-hat 1.0089 and 1.0119 (now reported as split r-hat, 1.0061 and 1.0074). The previous estimate of “roughly 10,000 solver calls per chain” is replaced by the measurement.

• In-figure series labels were below the WCAG AA contrast threshold for text (4.42:1 and 3.20:1 on white). Label colours are darkened; plotted line colours are unchanged, so the figures are otherwise identical.

Added

• An artifacts and provenance page indexing every committed artifact against the script that writes it and the claim it backs, with a test that fails if an artifact is added without a row or a row points at a file that is gone.

• Documentation site at https://impact-adjoint.vercel.app, with a component reference generated from the Tesseract schemas, an interactive step-size sweep driven by the committed E3 artifact, and a comparison against Diffrax, MuJoCo MJX, Brax, DiffTaichi, NVIDIA Warp and Dojo.

• Every figure now names the artifact it was plotted from and the script that drew it.

Baseline

First complete entry: three Tesseracts, eight experiments, seven verification studies, twelve golden tests, two solver-independent oracles plus a finite-difference gate, and CI on every push. The suite has grown since; the README states its current size.