https://shissorobotics.com/changelog/ Every correction and update to a published result, newest first. Ravindu Somawansa2026-11-01T00:00:00Z tag:shissorobotics.com,2026-11-01:changelog-24 2026-11-01T00:00:00Z The plain first screen now covers the 22 engineering, method and post-mortem posts too (WS-084): a three-sentence summary in the form "We wanted to know… We did… This happened", the two boxes and tap definitions; every post on the site now has one. Four new gates in the experiment log (TOPReward below elapsed time, RACE not seen, VLA-Corrector (ZJU) monitor fail, execution gap not seen), each with a folder that recomputes it, and one finding: the $200 arm executes what it is told (18.9 % residual). (commit: site (this deploy); paper-vs-hardware 9d3128b) tag:shissorobotics.com,2026-11-01:changelog-23 2026-11-01T00:00:00Z Every post's reproduction checked cold from the public repository and Hub (WS-067): L0 29 / 29 pass, L1 6 reproduced, 0 contradicted, 12 waiting on the world-model environment, 11 needing a larger GPU, the arm, a private checkpoint or MAXN. Packaging fixes merged; no published number changed. Table: paper-vs-hardware rig/REPRODUCIBILITY.md, linked from the Rig page. (commit: paper-vs-hardware 49479a9; site (this deploy)) tag:shissorobotics.com,2026-11-01:changelog-22 2026-11-01T00:00:00Z The 15 register posts without a plot gain one (WS-083): every register post now shows its result as a figure with the pre-registered bar, every plotted number checked against the register. (commit: site (deployed 2026-11-01)) tag:shissorobotics.com,2026-10-31:changelog-21 2026-10-31T00:00:00Z Every register post gains a plain first screen for readers who build robots (WS-075): a three-sentence summary, the result plot where one exists (7 of 22 so far), an "If you build robots" and an "If you do research" box, and tap definitions for 36 terms; the posts' own text is unchanged (checked by hash at every build). The gate count is now read from the experiment log at build: 89 (PRIME and smooth chunking A1 / A2 added to the served log; PRIME counted as void). (commit: site (this deploy)) tag:shissorobotics.com,2026-10-31:changelog-20 2026-10-31T00:00:00Z Claims register audited against the papers' PDFs (WS-068): 21 of 22 papers re-read, 39 citations corrected to the section, table or figure, 3 values marked as not in the paper's text, and 22 paper-side values corrected (numbers, wording, conditions, and claims that were not in the paper); no verdict changed. The Thesis page states the audit date and coverage. (commit: site (this deploy)) tag:shissorobotics.com,2026-10-31:changelog-19 2026-10-31T00:00:00Z VLA-Corrector's claim re-pointed to its own method: a GRU verifier on observable history (our closest test, REAP v2's GRU floor: 2/8 at 3/42, AUROC 0.59), not latent drift; verdict unchanged (fail). (commit: site (this deploy)) tag:shissorobotics.com,2026-10-31:changelog-18 2026-10-31T00:00:00Z Paper 22, REMORA's readout, published (P1 pass, P2 and P3 fail: not adopted). (commit: site (this deploy)) tag:shissorobotics.com,2026-10-31:changelog-17 2026-10-31T00:00:00Z Site hygiene, no content change (WS-066): ISO post dates, the site's icon on every page (favicon.ico, PNG, apple-touch), own descriptions and titles of at most 60 characters where the site name pushed them over, social cards on the legal page, breadcrumbs and isBasedOn (the arXiv link) in the structured data, no arXiv link for a paper without an id, and an Atom feed of this changelog at /changelog.xml. (commit: site (this deploy)) tag:shissorobotics.com,2026-10-30:changelog-16 2026-10-30T00:00:00Z The thesis is rewritten (what survives when a paper leaves the lab: the hardware people can buy, the cheapest baseline first, the bar before the run), on /thesis/ and the landing, its figures generated; the old text was the blog index from before the register. /findings/ opens with the disagreements. (commit: site (this deploy)) tag:shissorobotics.com,2026-10-30:changelog-15 2026-10-30T00:00:00Z /architecture/ removed (it redirects to /findings/; the gate decisions it listed are in the experiment log) and the post list is one page (/blog/page/N/ redirects to /blog/). (commit: site (this deploy)) tag:shissorobotics.com,2026-10-30:changelog-14 2026-10-30T00:00:00Z Every paper in the register gains its goal and its own hypotheses (each marked with the verdict of the claims that tested it here, or not tested here) and its tested claims as triples, every number re-read from its source file at build; /findings/ gains "Where the papers and this rig disagree" (5 contradictions found by the build). WS-061. (commit: site (this deploy)) tag:shissorobotics.com,2026-10-30:changelog-13 2026-10-30T00:00:00Z Posts 19 (Adaptive Action Chunking) and 20 (PACT-VLA), published overnight with the old "Paper vs hardware N:" prefix, take the paper's name as their title like the other paper posts. (commit: site (this deploy)) tag:shissorobotics.com,2026-10-30:changelog-12 2026-10-30T00:00:00Z Site restructured into Papers · Findings · Rig · Thesis (WS-060). The 20 paper posts drop the "Paper vs hardware N:" prefix: each title is now the paper's name as it stands in the register (URLs and dates unchanged); the series is labelled "Papers tested". /what-to-try-first/ and /systems/ are merged into /findings/ with nine method findings. The gate-protocol post's verdict count, typed by hand, was stale (16 papers); it now points to the scoreboard. (commit: site (this deploy)) tag:shissorobotics.com,2026-10-29:changelog-11 2026-10-29T00:00:00Z Rig prices are now what I paid, in euros: SO-101 pack (follower, leader, wrist camera) €350, Orin NX €950, C920 €50. Site-wide hardware total ≈ €1,350 (was ≈ $930 at list prices); mat, table and cables are not priced yet. (commit: site (this deploy)) tag:shissorobotics.com,2026-10-29:changelog-10 2026-10-29T00:00:00Z Gate count 73 → 74: the HWM gate row was added to the experiment log; the failed count (20) is unchanged (HWM is not proven). The work-stack figures are now dated (29 October). (commit: site (this deploy)) tag:shissorobotics.com,2026-10-28:changelog-9 2026-10-28T00:00:00Z Added REKI trained on my data and run on the arm: 19/20 on each instruction (SmolVLA 9/10), oscillation 1/40; the memory-off control is not run. Cost $13.05. (commit: site (this deploy)) tag:shissorobotics.com,2026-10-27:changelog-8 2026-10-27T00:00:00Z Corrected the arm price: $300 → $230 (SO-101 follower + leader list price); site-wide hardware total now ≈ $930 at list prices, from the rig page. (commit: site 1ca365b) tag:shissorobotics.com,2026-10-26:changelog-7 2026-10-26T00:00:00Z Added LeRobot's identity and exact RTC modes as baselines; the 71–73 % is scoped to naive async. (commit: site 221452c) tag:shissorobotics.com,2026-10-26:changelog-6 2026-10-26T00:00:00Z Corrected: the prefix overwrite uses RTC's soft prefix weights without the gradient, not the hard mask. (commit: ravediamond.github.io 743a1f4) tag:shissorobotics.com,2026-10-23:changelog-5 2026-10-23T00:00:00Z ReAP v2 (ramp target, no time feature, common horizon, shuffled-action control): verdict moved from not proven to fail. (commit: ravediamond.github.io 472eeb3) tag:shissorobotics.com,2026-10-23:changelog-4 2026-10-23T00:00:00Z Oscillation share restated over all 69 episodes (20.3 %). (commit: ravediamond.github.io 472eeb3) tag:shissorobotics.com,2026-10-23:changelog-3 2026-10-23T00:00:00Z Our port claim withdrawn: upstream's own trainer on our data ignores the instruction too (path A). The cause is the data. (commit: ravediamond.github.io 25e0356) tag:shissorobotics.com,2026-10-18:changelog-2 2026-10-18T00:00:00Z Two Jetson gates voided the day they were run: G11-A timed a random encoder, G12 timed a stand-in model. (commit: ravediamond.github.io 7eb2e2f (post, 2026-10-22)) tag:shissorobotics.com,2026-10-18:changelog-1 2026-10-18T00:00:00Z Correction row: the W1 latency benchmark had loaded the checkpoint with strict=False; its timing stands, its latents do not. (commit: ravediamond.github.io c87d807)