Honesty about a study's own errors belongs on one page, not repeated above every number. A caveat loud enough to be unmissable becomes the subject, which is what happened to the previous version of this project. Everything below is a place this study was wrong and corrected itself, or a place two sources disagreed.
Every place two sources disagree: the values, the source ids, the value chosen, and why. Newest at the bottom. Rule 1 of the brief: a null is acceptable, a guess is a defect — and a disagreement recorded here is neither.
| id | date | field | A says | B says | chosen | why |
|---|---|---|---|---|---|---|
| C001 | 2026-09-06 | Bambu AMS 2 Pro price | AMS collection page: $299.00 | H2D product page add-on block: $269.00 (struck $299.00) | $299 | Resolved in Phase 1: on 2026-09-06 every AMS 2 Pro printer-family variant priced at $299; the $269 did not reproduce. It was a with-printer bundle discount. |
| C002 | 2026-09-06 | Creality CFS price | CFS product page: $319.00 | Same page: $139.00 | $319 | Resolved in Phase 1: the $139 is the CFS Workstand in the accessories rail on the same page, not a CFS variant. Tom's Hardware independently quotes $319 (2025-12-08). |
| C003 | 2026-09-06 | Bambu P1S price | Collection page "From $399" | Product page default $579 | $399 bare / $579 combo | The product page loads with a Combo variant pre-selected. This is the single most common price-capture failure and it recurs below. |
Every row here is a Phase 0 census figure that turned out to be wrong. The pattern is nearly always the same: a product page's default-selected variant is not the base product, or a global store was read instead of the US one. These are recorded as conflicts rather than quietly fixed because the census is the Phase 6 diff baseline.
| id | date | field | Phase 0 recorded | Verified in Phase 1 | why the first reading was wrong |
|---|---|---|---|---|---|
| C004 | 2026-09-06 | Bambu P2S Combo | $780 | $749 | $780 is the P2S Combo Game Starter Bundle variant (adds 3 filament rolls + a card-shuffler kit, struck from $829). Phase 0 read the default-selected variant. |
| C005 | 2026-09-06 | Bambu A1 Combo | $430 | $399 | Same failure: $430 is the A1 Combo Game Starter Bundle. |
| C006 | 2026-09-06 | Every Prusa price | e.g. CORE One+ $1,202.78 | $1,299 (and MK4S $799 not $740, kit $999 not $925, MINI+ $549 not $508) | prusa3d.com hides its price element with [data-ge-price] { visibility: hidden; } and a Global-e script substitutes the real US price. Phase 0 scraped the hidden pre-conversion value, so every Prusa figure was about 8% low. Corrected against each page's own JSON-LD offer. This changes conclusions: CORE One+ kit + MMU3 is $1,485.38 all-in, not ~$1,294, so it no longer fits the $1,300 budget — only the stretch. |
| C007 | 2026-09-06 | UltiMaker S3 | $1,095 / $1,152.63 | $2,999 | The two figures Phase 0 found are the UltiMaker PVA Removal Station in the same page's related-products carousel. Confirmed by the page's own JSON-LD (sku MMGMGK4S, $2,999, InStock) and by UltiMaker's own store at the same price. The "Save $491" banner is real and genuinely attached to the S3, off a ~$3,490 baseline — it just does not make it a $1,095 machine. The S3 fails hard filter (b) and is not scored. |
| C008 | 2026-09-06 | MakerBot Sketch Standard | $799 | $999 | $799 is a stale "Sketch Basic" marketing tile on makerbot.com whose own Buy button leads to a $999 SKU. msrp_usd left null rather than record either figure as list. |
| C009 | 2026-09-06 | Rat Rig V-Core 4.1 | $1,035.29 "US kit" | $1,803.45 pre-tax minimum | That price buys the structural and motion kit only. Rat Rig's own configurator states "Without this kit, only structural and mechanical parts are included" at the electronics step; the Electronics Kit is a separate $768.56 SKU. This removes Rat Rig from the budget entirely — it reads as an in-budget candidate and is not one. |
| C010 | 2026-09-06 | Anycubic, several | Kobra S1 Max $699, Kobra 3 Max $429, Kobra 4 $279 | $599 / $349 / $249 | Phase 0 read the global store. US buyers are redirected to a separate storefront (us.anycubic3d.com) with lower prices. Corrected throughout. |
| C011 | 2026-09-06 | Phrozen Arco | $930, sold out | $840, in stock | Phase 0 read global.phrozen3d.com. us.phrozen3d.com has the printer, Chroma Kit, outlet Chroma Kit and enclosure all in stock, verified by live cart-add and a $0.00 shipping quote to a Massachusetts ZIP. Arco + Chroma Kit is $1,243.12 all-in — inside the budget, not the stretch band. |
| C012 | 2026-09-06 | Creality K2 SE | "$499, CFS sold separately" | $499 = K2 SE Combo, CFS included; bare K2 SE is $279 and sold out | The $499 SKU is literally named "K2 SE Combo (CFS + Filament x2)". |
| C013 | 2026-09-06 | QIDI Plus4 | sold out | in stock, $649 bare / $799 combo | Phase 0 read the global store; us.qidi3d.com has it in stock shipping from a US warehouse. |
| C014 | 2026-09-06 | QIDI Q2 top variant | "$1,099 — probably a QIDI Box bundle" | a "Ship from: Brazil (pre-sale)" option | Not a bundle at all. The real Q2 Combo is $649. |
| C015 | 2026-09-06 | Sovol Zero | $359 | $389 | The $359 was the CA/UK variant price, not US. |
| C016 | 2026-09-06 | Kidoodle MiniBox A1 | $169.99 new | $224.99, and the cited offer was USED | The captured Amazon page reads "Buy Used: $224.99 / Condition: Used: Very Good". Used-market listings are never captured, so this is recorded but not offered as a new-unit price. |
| id | date | subject | A says | B says | chosen | why |
|---|---|---|---|---|---|---|
| C017 | 2026-09-06 | Bambu X2D noise | Bambu: "below 50 dB under typical conditions" | CNC Kitchen measured ~55 dBA at 1 m, >60 dBA on fast travel | the measurement | A vendor claim is never stored as measured. noise_is_measured exists for exactly this. |
| C018 | 2026-09-06 | Prusa CORE One noise | 3DPrint.com measured ~48 dB on its own unit | An eleven-page Prusa forum thread "Core One not usable inhouse"; an owner measuring 60 dB at 5 ft after full calibration and told by Prusa support "the Core One is a loud printer… It should not be expected to be quiet"; one report of 79 dB peaks | both, as a bimodal risk | The distribution is genuinely bimodal, not a single value with error bars. Averaging would erase the finding. Recorded as a risk in the practicalities dimension rather than a number. |
| C019 | 2026-09-06 | Anycubic Kobra X colour bleed | Wirecutter: swapping red to white "turned a significant portion of the model pink" | Tom's Hardware: multi-colour PETG including clear filament, "free of color bleeding using the default settings" | unresolved, recorded as the finding | Both are T2/T3 and both are recent. The disagreement is a direct consequence of the 10 mm cutter-to-melt-zone distance that also makes the machine fast and low-waste. For a buyer who wants cosmetic multi-colour parts this is the most important open question about the cheapest machine in the study. |
| C020 | 2026-09-06 | Flashforge Creator 5 "Zero Purge Waste" | Flashforge product title: "Zero Purge Waste" | Tom's Hardware (2026-07-13) weighed a 44 g prime tower against a 33 g model at default settings, and reports the low-purge behaviour is tuned only for Flashforge's own filament profiles | the measurement | A marketing claim in a product title is not evidence. This is the clearest case in the study of why purge waste needs T2. |
| C021 | 2026-09-06 | Snapmaker U1 enclosure | PCMag spec table: "Frame Design: Closed" | Snapmaker's own materials table lists ABS/ASA/PA/PC only "With Optional Top Cover"; the machine as sold has no lid and no filter | Snapmaker (T1) | PCMag is simply wrong about the machine as sold. Recorded as a source conflict so the error is visible rather than silently discarded. |
| C022 | 2026-09-06 | Geeetech M1/M1S architecture | One community review: "a bed slinger" | Another: "Fully enclosed, CoreXY" | neither; both fields left null and declared unverified | Two community sources contradicting each other on basic architecture is not evidence for either. |
| C023 | 2026-09-06 | Elegoo Centauri Carbon 2 reliability | Lifetime Amazon average 3.9 | In a 36-review sample, pre-2026-07-01 reviews average 2.52 and July–August 2026 reviews average 3.92 | the recent cohort, with the sampling caveat stated | The method lesson from v1: a lifetime average permanently carries the launch cohort and can never recover. Here it understates the machine as sold today. The reverse case also appears in this study — the Anycubic Kobra X's recent cohort looks worse than its lifetime average. |
| C024 | 2026-09-06 | InfiMech EX shipping to Massachusetts | The store's own rate endpoint on the newer listing: $0.00 | The same store's older listing of the same machine at the same price: $89.00 "Freight"; the newer listing declares a 0 g shipping weight | $89, the higher figure | Two live quotes to the same ZIP for the same machine. The $0.00 quote is almost certainly a store data error (0 g weight) that could be corrected without notice. Costing the optimistic figure would understate the price of a machine that already has thin support evidence. |
Rob's ruling on 2026-09-07: independent test evidence counts for glass fibre. These are the places where doing so puts a reviewer against a manufacturer. Both values are stored — the vendor fact in gf_named_by_vendor, the reviewer fact in gf_independently_tested — so neither overwrites the other and the tool can say which one it is answering from.
| id | date | subject | source A | source B | chosen | why |
|---|---|---|---|---|---|---|
| C025 | 2026-09-07 | Flashforge Creator 5 Pro, glass fibre | Flashforge's own supported-filament list names carbon grades only, never glass | A reviewer ran PAHT-GF on the machine | both, stored separately: gf_named_by_vendor false, gf_independently_tested true | The sharpest case in the study of a vendor and a reviewer disagreeing about the same material. Collapsing them into one boolean would lose the thing that matters — that running glass fibre here is unsanctioned but demonstrated. The tool's GF filter is vendor OR tested, so this machine now passes it; the materials_now score still reflects the vendor position, and that gap is deliberate and visible rather than silently reconciled. |
| C026 | 2026-09-07 | Elegoo Centauri Carbon 2 and Centauri 2, glass fibre | Elegoo's lists name carbon-reinforced grades and are silent on glass | PCMag ran ABS-GF on the Centauri Carbon 2; Tom's Hardware printed PET-GF on the Centauri 2 | gf_named_by_vendor null, gf_independently_tested true | Silence is not refusal, so the vendor field stays null rather than false. Four Elegoo configurations and the Centauri 2 Combo now pass the GF filter on the reviewers' evidence alone. |
| C027 | 2026-09-07 | Snapmaker U1, glass fibre | Snapmaker's own materials position is silent on glass, and the four stock hot ends are plain stainless | 3DPrint.com ran PETG-GF after fitting the $49 hardened set | gf_named_by_vendor null, gf_independently_tested true | The test is real but conditional on a purchase the buyer would have to make. abrasive_ready_now is false for this configuration and carries that condition, so the two fields together say "demonstrated, but not on the machine as sold". |
| C028 | 2026-09-07 | Elegoo CANVAS (coverless), glass fibre through the MAGAZINE | Elegoo marks ABS/ABS-CF, ASA/ASA-CF, PET/PET-CF, PETG-CF, PA/PA-CF and PC NOT RECOMMENDED on the coverless unit | Tom's Hardware ran PET-GF through that exact combo and called CANVAS "pretty much bulletproof" | the vendor: accepts_gf stays false | Deliberately NOT extended to the magazine layer. The reviewer's run was a low-temperature glass fibre, while the vendor's exclusion covers the hot, abrasive grades Rob wants loaded; and accepts_abs is false regardless, so abs_gf_in_magazine is false either way and no outcome turns on it. Recorded because the asymmetry with C025–C027 is a judgment call, not an oversight. |
Not evidence disagreeing with evidence: instructions disagreeing with instructions. Recorded here under the same rule, because a spec followed in a way its author would not recognize is a silent decision, and this is where the study keeps the ones it had to make.
| id | date | subject | A says | B says | chosen | why |
|---|---|---|---|---|---|---|
| C029 | 2026-09-08 | 04_NOTRBOOKLM_EXPORT.md §2, Leaderboard shape | "Split into two tables of ≤ 8 columns each, both keyed by rank and configuration" | the column list in the same sentence names nineteen columns; two tables that both repeat rank and configuration hold fourteen | three tables of ≤ 8 columns, each keyed by rank and configuration | The column list is the content and the split is the formatting. Dropping five columns to obey the count would have lost slots, magazine type, enclosure and the status flags, all of which the prompt sheet's own infographic prompts ask for by name. |
| C030 | 2026-09-08 | 04_NOTRBOOKLM_EXPORT.md §2, length of 00-source-of-truth.md | "under roughly 3,500 words so the whole file fits comfortably in context" | the same section's content list measures about 6,200 words at this data size, with every line trimmed: ten leaderboard rows nineteen columns wide, six configurations head to head with a rationale line per judgment, 26 known issues on the top five, 24 unverified fields, 45 glossary entries, 78 brands in the coverage summary | the content, at about 6,200 words | The cap's stated reason is context, and 6,200 words is roughly 8,000 tokens: a fraction of one notebook source. Cutting to 3,500 would mean dropping mandated sections or truncating the known issues, and the file has to answer the fact-check prompt on its own. The build fails over 7,000 words so the file cannot drift further without someone deciding to. |
| C031 | 2026-09-08 | preset name mmu30 | the spec names the preset mmu30, which reads as "the magazine weighted 30" | build/scoring.py raises the magazine to 25 (with value to 5 and practicalities to 0), and that is the ranking the study published | the name from the spec, the weight from the study, and the note line beside it says 25 | Renaming the preset would break the prompt sheet the buyer pastes from; re-weighting it to match the name would change a published ranking to fit a label. The one place the two can be reconciled honestly is the note line, which states exactly what the preset changes. |
| C032 | 2026-09-08 | table width | 03_DELIVERABLES.md §4: no table wider than six columns, and 04 says everything else in §4 still applies | 04 §2 sets eight for 00-source-of-truth.md specifically | eight in 00, six in every other file in the folder | The narrower rule is about documents that get read aloud and rendered in a PDF. 00 is a lookup table for a machine, and its own section says eight. Both limits are enforced by the builder rather than left to judgment. |
| C033 | 2026-09-08 | notebooklm-prompts-3d-printer.md chat rule 6, default weights | the sheet: reliability 25, multi-material 15, value 15, openness 15, upgrade path 10, materials now 10, software UX 5, practicalities 5 | build/scoring.py after OC-1 and OC-2: the same, but upgrade path 5 and materials now 15 | the study's weights; the sheet is corrected | Both sum to 100, so nothing about the sheet looked wrong. It is also the only line on the sheet that hands the notebook arithmetic to perform, and it would have produced a weighted score that disagreed with every table in the export. |