There’s data
in the fabric.
Is there a deal?

Start with three different things. Fiber is the raw constituent, such as cotton or polyester; yarn is the continuous strand used to construct many fabrics; fabric is a sheet or tube of textile material; a garment is the shaped article someone wears. Cotton spinning inserts twist into prepared fibers. Weaving interlaces warp and weft yarn systems; knitting forms a structure of loops. A fiber-composition label therefore does not specify fabric structure or garment construction. The mechanics research explicitly distinguishes composition, woven/knit structure and fabric family because they convey different information. ^F02F05F12

Whiteboard a representative route as fiber preparation → spinning → weaving OR knitting → preparation/dyeing/printing/finishing → inspection → cutting → sewing → garment QA → packing/shipping. This is a learning route, not a claim that every connected manufacturer owns all these operations. Dyeing may occur at fiber, yarn, fabric or garment stage; garment finishing occurs after assembly; purchased yarn or fabric can enter midway. Nonwoven routes also exist and are outside the yarn-based example. CottonWorks separates these production disciplines in its technical learning hub and explicitly describes preparation of fibers, yarns, fabrics and entire garments. ^F01F03

Keep sample, batch and roll separate. For this report’s proposed data model, a sample is a small physical specimen or test item; a batch/lot is a production grouping defined by the factory; a roll is an individually identified length of fabric. Ask the factory what its word ‘lot’ actually means. One dye batch can contain several rolls; one roll can yield many cut pieces; a laboratory specimen represents only its sampled location. These are modeling examples, not observations about the Turkish connection. A fabric article/SKU identifies a specification; a roll ID identifies an instance. A test on one specimen is not proof that every meter of every roll meets specification. QA distinguishes color, drape, hand, durability and defects from garment fit and seam integrity. ^F04

The following is a recommended discovery schema, not a statement that these fields exist in any mill. Units are explicit proposed storage units; retain original units and test methods rather than silently converting ambiguous entries. The process families are grounded in CottonWorks; enterprise/operations separation follows the ISA overview; property capture follows the fabric-mechanics study. ^{F01F02F03F08F12}

Do not lose the unit system. English cotton count, Ne, is indirect: a higher number means finer yarn; the CottonWorks explanation distinguishes singles and plied yarns. ‘30’ alone is not a usable engineering feature. ^F02 Likewise, g/m² is mass per area, not thickness; in the mechanics dataset density and thickness are separate inputs. ^F12 A color difference value needs its formula, illuminant, observer and reference; this is a recommended metadata requirement, not a claim that existing sheets record them.

Everything in this example is synthetic, invented solely to explain schema. It is not a real recipe, factory record, measurement, robot trajectory or saleable training example.

{"fabric_article_id":"SYN-J01","structure":"single_jersey","composition_mass_pct":{"cotton":100}}
{"roll_id":"SYN-R01","fabric_article_id":"SYN-J01","knit_run_id":"SYN-K01","length_m":80}
{"dye_batch_id":"SYN-D01","roll_id":"SYN-R01","recipe_revision":"SYN-REC-A-r2"}
{"sample_id":"SYN-S01","roll_id":"SYN-R01","position_m":12.0,"test_method":"synthetic-method-not-a-standard","areal_mass_g_m2":180}
{"image_id":"SYN-I01","roll_id":"SYN-R01","position_m":12.0,"image_uri":"synthetic/I01.png","label":"hole","mask_uri":"synthetic/I01-mask.png"}
{"episode_id":"SYN-E01","fabric_article_id":"SYN-J01","t_s":0.25,"camera_frame_id":25,"gripper_xyz_m":[0.1,0.2,0.3],"action_frame":"robot_base"}

Recommended tables are article/specification, material-instance genealogy, process run, recipe revision, test observation, image annotation, business event and, only if relevant, robot episode/timestep. Splits and merges need an input–output bridge with quantities; do not force everything into one flat spreadsheet. Scope IDs by company/site/system, preserve original IDs, and store time zones. Distinguish a setpoint from the measured actual value and a planned order from actual production. ISA describes the need for enterprise-to-manufacturing information exchange; GS1 EPCIS provides a model for timestamped status, location, movement, chain-of-custody and aggregation events. Neither standard proves that a particular factory recorded its genealogy. ^F08F14

Leakage warning—recommendation: do not randomly distribute neighboring crops from the same roll into train and test. Hold out physical rolls, production periods and, where the task requires it, unseen articles/factories. Thousands of crops can still represent very few independent fabrics; TILDA explicitly generated images by relocating and rotating textile samples. ^F11

1804: programmable manufacturing instructions. The Science and Industry Museum dates Jacquard’s patent to 1804. Punched cards selected warp threads to raise, allowing patterns to repeat. Its account connects those cards to Babbage’s Analytical Engine and quotes Ada Lovelace’s 1843 comparison between algebraic patterns and woven flowers/leaves. These are instructions for producing a pattern, not labeled examples of successful and failed production. ^F05

1924/1931: statistical quality control. NIST dates Shewhart’s control-chart memorandum to 16 May 1924 and his book to 1931. This was a broader manufacturing development originating at Bell Labs, not an invention specific to textiles. It explains the distinction between detecting process variation and inspecting individual goods. A sequence of quality measurements can support control charts without being a modern ML dataset. ^F06

1960s–1970s: computer-aided design and manufacturing. The Smithsonian’s Gerber archive describes computer-driven apparel cutters developed during these decades and records development diaries beginning in July 1968. It also dates Lectra’s founding to 12 November 1973 and describes a system that calculated and plotted garment sizes. CAD is computer-aided design; CAM turns design into manufacturing instructions. Patterns, size grading and cutting paths had commercial utility before today’s AI. The archive supports this history; it does not support a claim that any particular mill retains reusable historical files. ^F07

Enterprise and shop-floor integration. ERP organizes business planning and logistics; MES manages manufacturing operations. ISA’s overview places them at different activity levels and dates the ISA-95 framework’s establishment to the mid-1990s. An invoice, a scheduled order and a sensor trace consequently answer different questions. Having an ERP is not equivalent to retaining synchronized machine histories. This is an inference from their different roles, not a measured statistic about Turkish factories. ^F08

1996–2025: benchmarks, digital twins and learning. TILDA version 1.0 dates to 1996; AITEX’s cited AFID paper is from 2019. NIST’s May 2021 digital-twin report discusses machine health, scheduling/routing and virtual commissioning, explicitly acknowledging implementation challenges. The 2024 fabric-mechanics paper learns simulator parameters from metadata/images; the 2025 cloth-deformation project publishes dynamic motion recordings. These are different technical tasks, not one universal ‘textile AI’ market. ^{F09F10F11F12F15}

The conceptual progression is instructions → measurements → linked operational records → calibrated models → learning on defined tasks, not ‘textiles invented computing, therefore old textile files are valuable AI training.’ A static 3D asset can be useful, but calling it a factory digital twin does not supply live connections, calibration or validated prediction. That distinction is a recommendation informed by NIST’s implementation framing. ^F09

Counts below are publisher/author-reported and were checked on primary pages or papers, not by downloading and counting archives. Publicly downloadable does not mean commercially licensable. No mirror’s license should override original rights.

The mechanics paper is especially important negative evidence against ‘a swatch photo uniquely determines physics.’ It treats raw measurement and simulator-parameter estimation as separate steps, discusses differing equipment/results and interoperability problems, and notes that very different parameter values can yield similar drape. Its authors include SEDDI researchers and describe a commercial extended dataset; distinguish their evaluated research results from their assertion of market representativeness. ^F12

Established uses include pattern repetition, apparel cutting/design, manufacturing coordination, quality evaluation and material digitization for apparel simulation. Those uses are documented by the museum/archive, standards overview, textile education and mechanics study. They are not evidence of external AI-dataset purchases. ^{F04F05F07F08F12}

Relative scarcity—reasoned assessment, not a market census: generic defect pictures are already publicly available. The harder-to-replace unit is a trustworthy pairing with context: independently measured physics plus images; intended recipe plus actual process plus resulting quality; full material genealogy; rare failures plus representative normal production; synchronized perception/action data. The cited datasets demonstrate some public supply and narrow coverage, but do not quantify private availability, Turkish access or buyer demand. ^F10F11F12F15

An RL environment additionally needs observations, allowed actions, transition behavior, reward/termination and resets. This is a conceptual requirement, not an assertion of available factory infrastructure. Cloth motion capture can inform simulator calibration; it does not by itself supply policy-learning interactions. ^F15 For a one-off product, recommend first identifying the actual folder and missing joins, then defining a narrow deliverable whose quality can be demonstrated. Do not start by promising a complete digital twin or inventing willingness to pay.

1. What is the difference between 100% cotton, single jersey and a T-shirt? Expected answer: composition, fabric loop structure and assembled garment are different description levels. ^F01F02F12
2. A test report says 180 g/m². What object and method does it describe? Expected answer: ask for specimen/sample identity, roll/location, test method and timing; do not assign it automatically to all production. Thickness is a different variable. ^F04F12
3. Why might 20,000 image crops be less diverse than 500 fabrics? Expected answer: crops may share rolls/samples, lighting and texture; split by physical origin and task. TILDA explicitly includes relocation/rotation acquisitions. ^F11
4. Can recipe PDFs train a model to reduce dyeing failures? Expected answer: extraction perhaps; reliable outcome modeling additionally needs substrate, actual conditions and measured successes/failures. Preparation affects subsequent results. ^F03
5. Does synchronized video create an RL environment? Expected answer: no; establish actions/states/calibration and task dynamics, reward, termination and reset. Motion recordings can support calibration or perception without supplying those pieces. The public motion dataset illustrates the narrower asset. ^F15

Passing test: draw the route in section 1, mark where the offered files originated, identify their physical unit and join keys, and name one product they support and one they do not yet support. This is a recommended diligence exercise.

• Research cutoff and access date: 2026-10-25. Publication dates, named dataset versions and page metadata are distinguished; search crawl-age labels were not treated as publication dates.
• No proprietary Turkish factory data were accessed. Schemas, IDs, sample values, proposed products and diligence exercises are recommendations or explicitly synthetic illustrations, not statements of available inventory or buyer interest.
• Dataset sizes are primary publisher/author statements, not archive counts independently reproduced here. No dataset training, benchmark or license audit was performed.
• Unspecified licensing is an unknown, not permission. Article licenses, code licenses, academic-use statements and dataset redistribution rights must not be conflated.
• Scarcity assessment is relative to the reviewed public baselines, not a quantified supply/demand study. Technical feasibility does not establish a sale, budget or willingness to pay.

Recommendation. For a one-off data sale, the most defensible first prospects are specialist inspection, material-digitization and textile-software teams—not an assumed queue of frontier labs. They already offer products that use factory images, measured fabric properties, dyeing knowledge or structured material catalogs. Their public offerings establish relevant workflows, not willingness to buy your connection’s data. ^{M03M04M06M12M14}

A beginner distinction matters: a mill makes or processes fabric; an apparel manufacturer cuts and sews garments; a brand chooses, designs and sells products. An inspection system’s buyer wants fewer defective metres or claims, while a garment-development buyer wants a trustworthy digital material that reduces sampling uncertainty. Those are different deliverables, sales conversations and acceptance criteria, as the vendors’ product descriptions illustrate. ^M03M06

Known: established suppliers sell hardware-enabled inspection, software licenses/subscriptions, digitization/testing services, digital assets and decision-support tools. Inference: a small, rights-cleared dataset may be useful as an input, evaluation set or regional/material coverage extension. Unknown: which supplier has a funded external-data requirement, what it already possesses, and whether the Turkish factories can legally supply the needed records. Do not convert a product catalog into evidence of procurement demand. ^{M03M07M08M11M14}

İHKİB’s 2024 annual report, using Ministry of Trade figures, reports $19.705589 billion of apparel exports in 2024, versus $20.603008 billion in 2023, and total Turkish exports of $261.778131 billion. Its reported apparel share is 7.5% of national merchandise export value. The report’s apparel grouping includes HS chapter 61 knitted/crocheted products ($10.106bn), chapter 62 non-knitted products ($7.384bn), and chapter 63 other made-up articles ($2.216bn); rounded chapter totals differ slightly from the unrounded overall figure. Chapter 63 is not synonymous with clothing. These are export values, not domestic industry revenue or data spending. ^{M01, pp.3–4}

The 2025 report changes the underlying series: its TİM table gives 2024 apparel exports as $17.920376bn and 2025 as $16.773432bn; its E-Birlik destination table reports a 6.4% decline. It is unsafe to compare the 2025 $16.77bn directly with the earlier report’s 2024 $19.71bn and call that the annual decline. Its 2025 chapter table gives $8.783bn / $5.661bn / $1.980bn for chapters 61 / 62 / 63, respectively, and does not reconcile to the TİM headline. Its general-export share row and another product table also appear stale. Report the series and denominator, not a reconciled estimate invented from them. ^{M02, pp.3–5,7–8}

Both reports also present a broader “Apparel & Textile” export headline—$27.4bn for 2024, $26.2bn for 2025—with a less explicit statistical definition. Do not subtract apparel to manufacture a textile-raw-material figure, or add headlines across reports. A blanket HS50–60 total is not automatically raw materials: HS57 covers carpets and other textile floor coverings. ^{M01, p.2M02, p.2M21}

Commercial implication [inference]: Turkey is a credible place to search for relevant production diversity; none of these export values estimates the market for an AI training dataset. A realistic opportunity size must be bottom-up: named qualified buyers × a specific unmet data need × an actual quoted scope. No such buyer count or deal price has been verified here.

Reading the matrix: “Direct” means the service can substitute for a proposed dataset/digitization deliverable. “Adjacent” means it sells the downstream workflow and is a possible technical prospect, not a confirmed customer. Buyer roles and suggested pitches are recommendations inferred from the documented workflow; product claims are vendor assertions unless stated otherwise. “Undisclosed” means no numeric paid price established from the cited page, not that the vendor never quotes publicly.

There are at least 12 concrete commercial companies/services above, but not 12 independent data buyers: services share ecosystems, and public/free substitutes are labeled separately. Robotics is not added merely to inflate the prospect list: this evidence set establishes no robotics textile-data purchase.

This is not a wholly new market created by generative AI. Uster’s 12 February 2018 EVS acquisition-agreement announcement described automated locating, labeling, tracing and grading of fabric defects. Its September 2025 brochure combines equipment, software and adoption/outcome services. EVS should therefore not be listed as a cleanly separate new competitor solely because the older name appears in product material. The agreement announcement itself is not proof of transaction closing. ^M20M03

Hohenstein’s October 2020 factsheet already offered measured material parameters for 3D design. Smartex’s November 2022 announcement names Tintex, Familitex, Toraman Tekstil and Ekoten as customers and includes a Tintex founder’s economic-viability quotation. This is meaningful historical vendor-published usage evidence, including Turkish factories—not independent proof of savings, current deployment status or willingness to resell data. Its financing amount is not revenue. ^M11M05

Browzwear publishes attributed testimonials from Kookai, Perry Ellis, Odlo and Paladin, among others; Textura quotes PIC 3D Design House, Alpine Creations and Lolë on digitization. These support existing commercial workflows, but logo walls and testimonials do not establish order values, full adoption across a company, or demand for a third-party training corpus. No vendor revenue estimate is supplied here. ^M06M12

The monetization ladder is consequently: instrument/testing work → usable digital asset → software/workflow integration → ongoing intelligence or optimization. A one-off seller can participate at the first two levels without building an end-to-end factory system. That is a recommendation based on existing delivery patterns, not a claim that a transaction is guaranteed. ^M06M11M14

Recommended first pitch: “We can license a bounded, documented textile evaluation pack from consenting Turkish manufacturers. Which production/material failure do you currently lack evaluation data for?” Do not begin with “we have millions of images” when quantity, rights and capture consistency are unknown.

These gates reflect documented product constraints: Uster emphasizes inspection settings and over-detection; Hohenstein specifies controlled, standardized measurements; CLO lists fabrics its service cannot reliably test/scan; Datacolor requires dye-set preparation and instrument/software integration. The procurement checklist itself is a recommendation, not an observed universal purchasing policy. ^M03M11M09M14

Keep a one-off offer bounded: agreed sample schema and acceptance task first; then a fixed dataset manifest, known exclusions, provenance/permission statement, delivery formats and explicit use license. Price after scope and rights are specified. A vendor’s SaaS fee cannot be multiplied into a supposed training-data budget. Avoid promising exclusive rights until every factory, brand/pattern owner, employee-record issue and software-output restriction is resolved; this is recommended diligence, not jurisdiction-specific legal advice.

Weak moat [inference]: a directory of Turkish suppliers, generic product photos, public trade tables or common digital swatches. They face established platforms, bundled CLO fabrics, free research information and existing digitization workflows. Public data may still carry license restrictions. ^{M08M10M16M18M19}

Stronger potential moat [inference]: lawful access to hard-to-capture paired observations—fabric image + calibrated mechanics; real defect + expert severity + production context; recipe/process + measured outcome—delivered reproducibly and compatible with a specific buyer’s stack. The incumbents’ measurement, database and integration offerings show why context matters, while also showing that they may already own substantial relevant data. A factory relationship is an access advantage only after permission and usable capture are demonstrated. ^M03M06M11M14

Negative evidence: substantial substitutes exist; mature vendors are not empty vessels waiting for generic textile data. CLO’s test exclusions show that “any fabric can become an accurate digital twin” is too broad. A public anomaly benchmark cannot be repackaged as unrestricted commercial training data. None of this proves that a niche dataset cannot sell; it narrows what must be distinctive. ^M03M08M09M19

Frontier labs: no verified textile-specific dataset purchase by OpenAI, Anthropic or Google DeepMind was established in this review. The reviewed commercial sources show textile software/hardware use, not those labs procuring textile training corpora. A targeted web search for those labs plus textile/dataset purchasing or licensing did not establish a transaction; that is a bounded search result, not proof no private purchases exist. Public AI capability, generic data partnerships, or robotics research would not alone establish textile-data budget, interest or procurement timing.

Decision rule [recommendation]: seek a named technical owner and a written unmet-data description before commissioning expensive capture. If several relevant specialists cannot identify a gap that your factory access uniquely fills, do not rescue the thesis with global apparel-market size or hypothetical frontier-lab demand. The realistic product is a narrow licensed asset solving a known evaluation or digitization problem—not a venture-scale “textile data platform” by default.

• Vendor descriptions and testimonials establish offerings and attributed usage, not independently verified ROI or external-data demand.
• Undated live pages are observations at the cutoff, not independently archived historical snapshots; crawl dates were not used as publication dates.
• No buyer-specific budgets, data-license deal sizes, factory dataset inventory, vendor revenue estimates or prospective interest are invented.
• WGSN current pages and Fibre2Fashion main page were blocked; accessible primary evidence and explicit limits are supplied instead.
• İHKİB annual reports have source/denominator changes and internal table inconsistencies; use the exact report/table labels.
• Textura prices come from an app linked by its official pricing page; confirm current commercial terms and annual billing separately.

Recommendation; research cutoff 2026-10-25. Sell a specific answer to “where does our model fail?”, not “access to Turkish factories.” No manufacturer assets, rights, buyer conversations or willingness to pay have been verified for this connection. The rankings below are conditional judgments, not market statistics.

A dataset is a collection of observations: images, documents, sensor readings or action sequences. A benchmark adds a task definition, held-out cases, authoritative answers and scoring. An environment also lets an agent act, changes state, exposes observations and supplies a reward or success check; it must reset reproducibly. In τ-bench, for example, agents use APIs against databases and are assessed against intended terminal states. Its authors explicitly note that a correct final database alone can miss a policy violation such as acting without confirmation. ^O07

A research demonstration shows something can work under the authors’ conditions. A purchasable deliverable needs lawful usage rights, a manifest, documented sampling, an agreed interface, acceptance criteria and a defined license. A video of a robot folding a shirt is neither a robot-training dataset nor evidence that the lab will buy someone else’s factory video. Physical Intelligence’s own description distinguishes web pretraining, robot interaction data and task-specific post-training. ^O10

OpenAI: historical general intake, not a textile purchase order. The November 9, 2023 Data Partnerships announcement described public and private training datasets. Accessible contemporaneous reporting quotes interest in large-scale data unavailable online, multiple modalities and especially material expressing human intention. It does not establish compensation. ^O02 The official announcement and data-partnership interest-form URLs remain discoverable, but both returned HTTP 403 and the announcement browser hit a challenge during this research. Accordingly, present-day form operation, terms and acceptance are unverified; use the link as a plausible outreach route, not verified procurement. ^{O01; O02} A small defect pack may fit an evaluation team better than that broad training initiative—inference, not an OpenAI request.

Anthropic: do not confuse grants with buying factory data. Its June 27, 2025 Economic Futures launch concerned AI’s effects on labor/productivity, grants, policy and measurement; it included research-institution partnership contact. ^O03 The July 22, 2026 fund agenda supersedes the easy-small-grant interpretation: it focuses on ambitious external intervention research, accepts specified institutional applicants rather than individuals personally, and says it will not directly fund projects below $1 million. Those are research-fund conditions, not dataset prices or a budget available to this seller. A linked proposal form is a real research mechanism but a poor route for a one-off textile-image sale. ^O04

Google DeepMind: real robotics research, selected partnerships. The March 12, 2025 Gemini Robotics announcement describes vision-language-action models, primarily trained on ALOHA 2 data, and names Apptronik and selected trusted testers. It is vendor-authored capability evidence, not an open factory-data buying program. ^O09

Physical Intelligence: unusually relevant, still conditional. Its October 31, 2024 π0 announcement includes clothing folding and robot-specific data. Its November 17, 2025 Recap announcement combines demonstrations, interventions and autonomous experience with outcome feedback; it explicitly invites collaboration with companies scaling collection from robots deployed in real applications, with research@physicalintelligence.company as contact. This is verified collaboration interest in that category, not verified interest in this manufacturer or a paid acquisition. ^{O10; O11}

Skild: human footage can matter, but compatibility is unresolved. Its January 12, 2026 blog claims learning new skills from human videos with a small amount of robot data; the same page acknowledges missing forces/tactile signals and the human-to-robot embodiment gap. These are vendor assertions, not independent validation or a public purchase commitment. ^O12

Recommendation. Rank by the amount of new engineering required if assets already exist, not by futuristic appeal. None is a confirmed buyer opportunity.

For rank 1, out-of-distribution (OOD) means a meaningful change from development data—another loom, lighting regime, material or production period—not just random files withheld. MVTec AD already supplies over 5,000 industrial images across 15 categories; AD 2 adds eight scenarios, over 8,000 images, lighting shifts and a private-label test portion. Both pages prohibit commercial use under CC BY-NC-SA 4.0. ^{O05; O06} This is negative evidence against selling “another generic anomaly dataset,” and a possible differentiation for independently owned, commercial-use-permitted, real-production evaluation data. It does not mean MVTec data can be repackaged, that all research competitors have the same restrictions, or that commercial buyers must pay for this new pack.

Agent/MES/ERP tasks. ERP manages business records; MES tracks manufacturing execution. Proposed example: “Allocate stock to an urgent order without mixing prohibited dye lots; explain whether the remaining quantity can ship.” A useful case needs the order, inventories, lot restrictions, inspection results and an adjudicated correct result. A multimodal variant can include a QC image, but only if it is actually necessary to decide. To become an environment, add tools such as get_lot, reserve_stock and place_hold, seeded state, policy enforcement, reset and terminal-state verification. τ-bench provides a public example of databases, APIs, policies and goal-state checks—not evidence of a missing textile-data market. ^O07 Recommendation: sell expert-grounded cases first; build a small executable sandbox only after a buyer specifies its harness and pays for that work.

Material physics/world models. Bending is resistance to curving; stretching is extension under force; response can differ by direction. MIT’s 2013 fabric project paired 30 fabrics’ measured properties with 90 RGB videos under three wind strengths, illustrating why matching physical measurements to visual response is more valuable scientifically than fabric names alone. ^O08 DeepMind’s August 5, 2025 Genie 3 announcement describes interactive generated worlds but also constrained actions and only minutes of interaction. It does not establish calibrated cloth-mechanics accuracy or a textile data procurement program. ^O13 Inference: paired mechanical tests and controlled deformation could support physical reasoning or simulator validation. Their relevance to a particular frontier model is unknown until that team confirms representation, accuracy and task requirements.

Robot manipulation. A human folding video lacks robot joint states, issued actions and gripper feedback. Skild’s video-learning claim means this footage is not categorically useless; it does not erase missing signals or prove arbitrary CCTV will transfer. ^O12 Physical Intelligence’s Recap result specifically uses corrections and outcomes from the policy’s own robot experience. ^O11 Recommendation: approach with access to diverse fabrics, repeatable tasks and a lawful collection site, not a claim to possess robot demonstrations. Require a partner’s capture specification before recording.

Process decisions/RL. Proposed dyeing example: state = material/bath conditions; action = a temperature or dosing change; later outcome = quality and resource use. An offline RL dataset logs transitions and outcomes; an interactive simulator predicts consequences of new actions. They are not interchangeable. D4RL’s authors explain that narrow behavior coverage, hidden human cues and biased data can undermine offline learning and that evaluation in real systems is difficult. ^O14 Inference: if operators intervene only on difficult batches, naive correlations can misattribute bad outcomes to the intervention. Without linked actions/outcomes, call the records observational process data, not an RL asset. Never promise safe control improvements from a historical dump.

Recommendation: start with a private textile inspection evaluation pack, conditional on an asset/rights audit and a buyer naming a real gap. An illustrative discovery offer is “expert-adjudicated pass/fail and localization cases from one production process, with a clean holdout from another period or camera setup.” Do not promise this exists before inspection.

Proposed sample acceptance criteria—not observed inventory or universal standards:

• A buyer-agreed preview of roughly 50–100 independently selected cases, including good material, common defects and difficult negatives. Preserve natural frequencies separately from any intentionally balanced challenge subset. A small sample is for fit, not precise rare-defect performance claims.
• Original-resolution assets, stable anonymized IDs, checksums and CSV/JSON manifest; explicit defect taxonomy, image dimensions, capture conditions, fabric family and group IDs. Buyer decides whether masks, boxes or image-level labels are sufficient.
• A textile expert labels every preview case; a second reviewer independently checks the preview; disagreements are adjudicated and uncertain/ambiguous cases marked rather than silently forced. Define labels against written inspection criteria, not the seller’s intuition.
• Rights/provenance register covering each asset source and permitted use; no customer names, faces, order references, protected designs or recipe secrets without necessary permissions. Redaction must preserve decision-relevant relationships. Obtain appropriate legal review rather than treating anonymization as a license.
• Frozen grouping and split rules, documented exclusions and missingness; score script/interface and a sample output; task-specific metrics agreed before bulk preparation. For inspection, consider recall at an agreed false-alarm rate, per-defect results and localization quality where labels support it. Avoid a single headline accuracy on an artificially balanced set.
• Buyer accepts the preview’s usefulness, scope and labeling quality in writing; full-pack volume, quality tolerances, delivery format, permitted uses and remedy for defective records then enter a paid statement of work. No guarantee that every buyer model improves.

A second-best deliverable is 20–30 proposed expert-verified decision cases with input bundle, applicable policy version, expected result and scoring rationale. That count is a discovery scope assumption, not an inventory estimate. Do not describe static case files as a runnable environment.

Recommendations grounded in the public holdout patterns above. AD 2 keeps some ground truth private; τ-bench separates task annotations from agent-visible tools and policies. ^{O06; O07}

Split by physical roll/lot/session, and where relevant manufacturer or time—not adjacent image crops or video frames. Deduplicate originals and transformed variants before splitting. Preserve group IDs internally so redaction does not break leakage control. Hold out entire workflow templates or customer/order families when evaluating agent generalization.

Separate training licenses from evaluation licenses. Keep final labels, answer-bearing reports and future outcomes out of model inputs; in historical decision cases include only information available at decision time. Audit filenames, folder names, OCR, document revisions and metadata for accidental answer clues. Do not use a publicly posted demo as a private test case.

Record previous publication, licensees, authorized training exposure and version hashes. Request non-training handling for private test material; do not upload it to public model playgrounds merely to label it. If a buyer trains on a pack, retire that version as an uncontaminated evaluation set for that buyer. A nonexclusive benchmark cannot honestly guarantee universal non-exposure. Agree limited scoring access or an evaluation-only license if secrecy is the value; an elaborate hosted evaluation service is unnecessary unless requested. Document unknown historical exposure explicitly.

Recommendations, not observed deal terms. Offer (a) a nonexclusive internal evaluation license to a fixed pack; (b) a separate training license if authorized; or (c) a paid, fixed-scope preparation/evaluation pilot with staged acceptance. A pilot purchase must specify what is delivered even if the buyer concludes the data is not useful. Avoid a recurring-platform promise for a one-off goal.

Illustrative algebra: let N be prepared records, c variable preparation cost per record, F fixed legal/extraction/expert setup cost, S support/delivery cost and q the fraction accepted. Total cost C = F + Nc + S; effective preparation cost per accepted record = C/(Nq). For price P, contribution before tax = P − C. For target margin m, required P ≥ C/(1 − m), assuming no other costs. These are accounting identities, not evidence buyers will pay.

Purely hypothetical planning inputs: N=1,000, c=€1.50, F=€1,200, S=€300 and q=0.90 yield C=€3,000 and €3.33 per accepted record. A hypothetical 40% target margin implies a €5,000 price floor; that is not a quote, comparable transaction, market rate or predicted revenue. Replace every assumption with supplier quotes and buyer feedback. Include founder time, translation, rejects, taxes, payment delays and manufacturer revenue share in the actual model.

Nonexclusive licensing preserves resale potential but no repeat sales are assured; contamination can reduce evaluation value. Exclusivity sacrifices future licensing and may require restrictions on prior transfers and future collection. Define duration, field of use, territory, affiliates, training/derivative rights and existing licenses precisely. Recommendation: prefer limited, time-bounded or evaluation-only exclusivity over blanket permanent exclusivity; demand enough contracted value to compensate actual opportunity cost, not an imagined premium.

Recommended milestones; these are decision gates, not a guaranteed sales cycle.

Days 1–2: ask the connection for an inventory interview, a screen-shared inspection and a rights-owner map. Record actual modalities, counts, date range, normal/defective coverage, group IDs, experts and permission constraints. Do not transfer sensitive samples yet. If no lawful, outcome-bearing asset exists, stop the data-sale plan.

Days 3–4: send a one-page capability sheet, without raw confidential data, to a small named-role shortlist. Start with inspection-model/evaluation leads for wedge 1; enterprise-agent evaluation or ERP/MES AI teams for wedge 2. OpenAI’s historical program is a plausible general route, Anthropic Economic Futures is not a data-buying route, and PI’s stated robot-deployment collaboration category applies only if the factory meets it. ^{O01–O04; O11} Google DeepMind and Skild are technical-fit prospects, not verified purchasers. ^{O09; O12}

Days 5–7: seek calls with the person owning evaluation failure and procurement. Ask: What exact task and failure slice is missing? What public/internal substitute is insufficient? Training or evaluation? Which rights and geography restrictions? Required formats, groups and labels? How will usefulness be judged? Can an expert-grounded sample unlock a paid pilot? Who signs, from which approved process, and on what timeline? Are there vendor onboarding, security or tax requirements that rule out a quick transaction? What makes this a definite no?

Days 8–10: only with concrete technical feedback and provisional rights clearance, prepare the agreed preview; ask the buyer to run it or specify a controlled evaluation. Seek an accepted pilot scope and written acceptance criteria—not compliments or a generic “send more.”

Days 11–14: agree license, fee, manufacturer share, payment milestones and delivery obligations. Target evidence is an authorized buyer’s funded pilot/order or a specific procurement step, not a partnership-form submission. If useful feedback is absent, pause preparation; silence from a tiny outreach sample is not proof no market exists.

Go only with rights, existing distinctive assets, checkable ground truth, a named buyer problem and paid scope acceptable relative to true costs. No-go if ordinary records lack permissions or outcomes, buyers only want free research access, required capture infrastructure is absent, or a quick sale depends on building an RL environment before anyone specifies it. The honest result may be that the connection is valuable for introductions and future collection, but not yet a sellable AI dataset.

• No manufacturer inventory, legal title, buyer interview, purchase order, buyer budget or licensing comparable was verified. Ranking and product specifications are recommendations.
• OpenAI primary pages blocked retrieval; historical claims rely on explicitly attributed contemporaneous secondary reporting. Live form operation and current intake terms remain unverified.
• Vendor capability announcements support technical direction, not independently proven performance, dataset purchasing intent or transferability to textile-factory footage.
• All numerical package sizes, economic inputs, margins and validation milestones are illustrative assumptions, not market rates or promised outcomes.
• Publication dates were checked against page dates/metadata where available; undated dataset pages are marked. Sources after 2026-10-25 are excluded. Publicly downloadable does not imply commercially licensable.

Research cutoff and access date: 2026-10-25. Known means supported by the cited source; inference means an engineering interpretation; recommendation/design means proposed here, not an observed factory capability or buyer requirement. Every number in the episode below is invented for education. No source establishes that this connection has usable machine logs, robot trajectories, rights to redistribute them, or a willing purchaser. Mutable documentation supports interface descriptions, not claims that a particular release existed on an unverified date. Dated releases and papers anchor historical availability; vendor performance superlatives are not adopted.

An RL policy chooses an action from observations/history. The state is everything the environment needs to evolve; the observation is the permitted view of that state. An action changes something; the transition computes the next state; the reward scores the result. A rollout is the resulting sequence of observations, actions and rewards. Reset creates a fresh episode. Termination means the task reached its own terminal condition; truncation means an external limit stopped it. Gymnasium explicitly separates these and defines reset(seed=...) -> (observation, info) and step(action) -> (observation, reward, terminated, truncated, info). ^E01

Example: the full scheduling state includes the queue, machine setup and clock. The observation exposes current orders and documented constraints, not tomorrow's actual breakdown. An action chooses the next order. The transition advances the clock and consumes capacity. Reward penalizes delay; termination means all orders finished. Evaluation is running a frozen policy on fresh, held-out episodes—not updating the policy on those answers. This example is a proposed design using the Gymnasium contract. ^E01

Static logs are not an interactive environment. They show the action somebody actually took and its observed successor. If the agent chooses a different action, the next logged row is not automatically its consequence. A simulator must implement that missing counterfactual transition. Offline RL learns from fixed historical data without further online collection; this is a distinct, difficult research setting, not a shortcut that makes arbitrary alternative actions observable. ^E02 RLlib separately documents trajectory storage, simulator-based evaluation and off-policy estimation; its importance-based evaluation uses logged action probabilities. ^E10

Distinguish five products:

Recommendation: use a bandit only when delayed capacity, inventory and later decisions can reasonably be excluded. Scheduling with shared machines is naturally sequential. Never call an ordinary ERP export a digital twin.

Recommendation: choose one task buyers can already run—such as mapping an inspection image plus specification to a defect label and cited acceptance rule. Start with a negotiated sample, not an ingestion platform.

Factory permission + task definition
 -> restricted immutable raw export + provenance ledger
 -> redaction / unit normalization / duplicate grouping
 -> expert labels + second-review disagreements
 -> group/time split + frozen answer key
 -> versioned Parquet/JSONL + image files + manifest
 -> buyer model adapter (read-only cases)
 -> isolated scorer (private labels)
 -> per-slice report + licensed encrypted delivery

The deliverable is cases, taxonomy, annotation instructions, adjudication evidence, schema, frozen split IDs, scorer and baseline results. Keep uncertainty and “cannot determine from available evidence” labels instead of forcing false certainty. A model's predicted label and explanation are evaluated against a human-defined task; neither is a machine command. These are design recommendations, not claims that labels or photographs already exist.

LeRobot demonstrates a useful existing organization for robot data: tabular signals, video and feature/episode metadata. Its v3 documentation describes Parquet, MP4 and schema metadata; a dated October 24, 2025 release announcement corroborates v3 availability. ^E04E15 Use ordinary Parquet/JSONL for document/QA products; forcing them into a robotics format adds no buyer value. For genuine robot trajectories, prefer buyer-requested LeRobot or RLDS rather than inventing another ecosystem. Open X-Embodiment represents datasets as sequences of RLDS episodes and specifies action semantics; matching column names alone does not align coordinate frames or absolute/delta/velocity actions. ^E05

Recommendation: only build after a buyer asks for multi-step planning/tool-use evaluation, agrees on allowed tools and score, and supplies or approves the underlying rules. A deterministic discrete-event model needs no physics engine. Tools might be list_orders, inspect_machine, and schedule_next; they operate on an episode-local database, never a production ERP.

flowchart LR
  F[Rights-cleared rules and calibration evidence] --> V[Versioned scenario generator]
  V --> R[Reset with scenario and seed]
  R --> S[Private authoritative state]
  S --> O[Observation serializer]
  O --> P[Buyer policy or agent]
  P --> A[Typed action validator]
  A --> T[Deterministic transition]
  T --> S
  T --> G[Private reward and terminal verifier]
  G --> L[Append-only rollout and hashes]
  O --> L
  A --> L
  L --> E[Held-out evaluator and baseline comparison]

Text equivalent: approved scenarios initialize state; the policy sees only serialized observations; the validator checks requested actions; the transition mutates authoritative state; a separate verifier scores outcomes; the recorder captures replay inputs; held-out evaluation compares policies without exposing answers. Training is optional. To train, collect rollouts with policy version and, where needed, log-probabilities; compute rewards/returns; update on training episodes; checkpoint; repeat; evaluate separately. RLlib provides experience ingestion and rollout/evaluation infrastructure, while verl documents function-based LLM rewards using responses and ground truth. Neither supplies textile transition rules or validates industrial rewards. ^E10E11

Do not start with Ray clusters, model training, Kubernetes, a hosted environment marketplace or LLM-as-judge. A single local process, SQLite or in-memory state, a CLI and JSONL traces suffice for this design. Add distributed execution only after measured throughput needs and an actual buyer requirement. A valid JSON tool call is not proof of a good schedule; score authoritative outcomes.

Educational design, not a factory process model. One non-preemptive machine; all orders available at minute zero; one order per action; no breakdowns, parallelism or wet-processing chemistry. Integer minutes avoid floating-point replay ambiguity. The current color family is blue. Changing family costs one minute. Each order must run exactly once; invalid IDs or repeats leave state unchanged, give reward −10 and consume one action attempt. A fixed external attempt cap truncates unfinished episodes; partial completions do not count as success.

The minimal reset specification is {scenario_id:'toy-01',seed:17,env_version:'toy-1',reward_version:'r1'}. Its fixed fixture contains A: blue, processing 3, due 3; B: red, processing 2, due 5. The seed is recorded but does not change this fixed fixture. A generated variant would record the PRNG name/version and sampled orders; a seed alone does not reproduce code or dependency changes.

Total return −8, makespan 6 minutes, total tardiness 1 minute. Reverse order produces B at minute 3, A at minute 7, total tardiness 4 and return −15. These are arithmetic consequences of invented rules, not empirical performance. For two orders enumerate both permutations as the optimality baseline; for larger instances compare earliest-due-date and an exact solver where tractable. Never interpret a better toy score as factory savings.

Small runnable idea for the parent: implement this fixture and reset/step in a dependency-light Python educational demo, returning copies of observations; enumerate the two policies and emit a replay trace. Gymnasium wrapping is optional; its observation/action-space and reset/step contracts are the reference. ^E01 No robot, cloud account, RL training or real factory data is necessary.

Recommended contract: each source record has record_id:string, pseudonymous factory_id/machine_id/lot_id:string, event_time_utc:timestamp, available_at_utc:timestamp, source_uri:string, source_sha256:string, rights_id:string, task_id:string, units:map<string,string>, nullable quality flags and annotation provenance. Preserve original-language text beside translations; preserve capture/label timestamps. Missing is nullable, not a made-up zero. For robot records add camera timestamps/intrinsics/extrinsics, coordinate frames, actuator units, action frequency, synchronization offsets, calibration ID, teleoperator policy, interventions and success/failure labels. DROID's project reports 76,000 demonstration trajectories, standardized robot/camera hardware and later calibration corrections—evidence of collection/calibration effort, not proof your contact owns equivalent data. ^E06

A rollout row should include episode_id:string, step:uint32, observation:object, action:object, next_observation:object, reward:float64, terminated/truncated:bool, policy_id:string, nullable behavior_action_probability:float64, and private before/after state hashes. Probability is genuinely unknown unless logged or recoverable from a known behavior policy; do not fabricate it. RLlib's documented off-policy estimators require appropriate behavior-policy probability information. ^E10

Split before augmentation or normalization. Keep every image crop, adjacent video frame, repeated inspection, lot and derived task in the same group. Publish separate evaluations for held-out factories, future time windows, machines and cloth/fabric families; these are distinct axes, not a claim that one small dataset can support all four. If only one factory exists, explicitly say cross-factory performance is unmeasured. Fit normalization only on training data. Hold out scenario templates/rule combinations as well as seeds so generated test tasks are not superficial copies.

Manifest proposal: dataset_version:string, schema_version:string, created_at:timestamp, license_id:string, files:array<{path:string,bytes:uint64,sha256:string,rows:uint64?}>, splits:map<string,array<record_id>>, split-policy text/hash, source snapshot IDs, taxonomy/annotation-guideline hashes, code commit, dependency lock hash, environment/reward versions, scenario-generator hash and seed list. Treat a manifest as immutable; corrected labels create a new version with a change log. DVC tracks data via Git-versioned metadata and content checksums, but is optional: a signed SHA-256 manifest plus archive is enough for a one-off handoff. DVC's ordinary MD5 tracking is not the cryptographic delivery signature proposed here. ^E12

Replay must match scenario bytes + code/dependencies + seed + ordered actions + reward version, then compare observations, flags, components and state hashes. API model outputs can be nondeterministic: replay recorded actions first, and separately measure policy variability across repeated runs. GPU physics requires its own tested reproducibility tolerance; do not promise bitwise equivalence from a seed alone.

Recommended acceptance report: static QA—per-class precision/recall, confusion matrix, defect localization metric if labeled, extraction exact-match with unit tolerance, evidence correctness, abstention coverage/risk and inter-annotator disagreement. Scheduling—completed-feasible episode rate, constraint violations, total tardiness, makespan, setup count, return, invalid-action rate, tool calls, latency and cost. Report episode counts and uncertainty by held-out group, not only an overall mean. Pre-agree thresholds with the buyer; this research invents no universal passing score.

Reward-hacking examples: reward each tool call and the agent loops; reward outputting “pass” and it passes defects; penalize tardiness only and it never finishes late orders; expose future outcomes and it memorizes them. Mitigations: authoritative state-based scoring, one-time completion accounting, explicit unfinished/invalid metrics, capped attempts, independently recomputed terminal checks and adversarial policies that try duplicates, malformed tools and premature exits. Keep absolute completion/feasibility alongside reward so a convenient scalar cannot hide failure.

Task leakage includes answer keys in filenames, expert rationales in prompts, future measurements joined into current observations, duplicated lots across splits, test cases used to tune rewards, and model access to private verifier state. Separate agent/evaluator processes and files; emit only allowlisted observation fields. Training environments may expose reward, but held-out evaluation should withhold detailed feedback until the run is frozen. Keep test labels private under a buyer evaluation arrangement when possible.

Counterfactual validity is separate from scoring correctness. Exact replay verifies software consistency. Matching real logged schedules checks only factual trajectories. To argue interventions transfer, seek expert-reviewed causal assumptions plus authorized comparisons under changed schedules; confounding and unsupported actions remain unknown. Offline/bandit literature motivates specialized policy evaluation, not unconditional causal guarantees from ordinary logs. ^E02E03E10

SoftGym already provides public deformable-manipulation simulation benchmarks and identifies high-dimensional, partially observed dynamics as challenges. This is negative evidence against selling a generic cloth simulator merely because cloth is involved. ^E14 MuJoCo documents contact as an approximation with configurable softness/friction; Warp v1.6.0 provides CPU/GPU simulation primitives and cloth/self-contact examples; Isaac Lab exposes separate observation/action/reward/termination/event managers. These are implementation tools, not industrial cloth-fidelity certificates. ^E07E08E09

Consented fabric specimens + calibrated mechanical/contact measurements
 -> identified material model + uncertainty bounds
 -> cloth mesh / collision / self-contact + robot / gripper / camera assets
 -> simulator + sensor timing / actuator model
 -> randomized calibrated parameter scenarios
 -> Gym-style observation / action / reward loop
 -> policy training or evaluation + synchronized trajectory export
 -> held-out cloth / scene / machine tests
 -> controlled robot-cell trials under safety supervision
 -> sim-to-real discrepancy report and licensed assets

Recommendation: define the physical task first—single-layer pickup, corner localization or folding, not “all textile manufacturing.” Calibrate mass per area, stretch along material directions, bending/shear, damping, friction, thickness, gripper slip and contact against measured tests. Compare drape, pull/slip, fold geometry and trajectories on unseen specimens. Evaluate camera occlusion/noise, timing and action semantics. Randomize plausible measured ranges of mechanics, lighting, camera pose and initial folds; randomization widens variation but cannot validate the wrong equations. Document unsupported phenomena, such as seams, entanglement or wet adhesion, instead of silently simulating them as generic cloth.

Open X and DROID are useful public baselines, not evidence of textile-specific coverage, permission to relicense every contributing asset, or purchased demand. DROID itself describes hardware/labor investment and safety logistics. ^E05E06 A possible differentiated product is measured, rights-cleared fabric/contact calibration with paired held-out manipulation trials—but that is an inference, and specimen access alone does not deliver it.

Recommendation: deliver an encrypted archive or buyer-controlled private storage containing data, schema, split IDs, signed manifest, reproducible local scorer, dependency lock, licenses/attributions, provenance/rights ledger and limited acceptance report. Pin dependencies; avoid untrusted pickle/model execution; use JSON/Parquet and ordinary media formats. Run buyer agents as unprivileged processes with read-only cases, no production credentials, denied network egress by default, bounded CPU/memory/time and an allowlisted local tool API. Treat factory documents as untrusted content, not executable instructions. Keep evaluator secrets outside the agent filesystem; a container sharing those secrets is not privilege separation.

Agree rights separately for factory exports, worker images/voices, buyer/order information, standards text, simulator assets and third-party datasets. A repository software license does not automatically authorize resale of every embedded dataset; Open X explicitly distinguishes software and other materials and points to contributing datasets. ^E05 Define commercial training/evaluation rights, redistribution, exclusivity if any, retention/deletion, permitted derivatives, confidentiality and maintenance scope. Public examples must be synthetic or explicitly cleared.

Do not connect exploratory policies to loom controls, dyeing equipment or robot actuators. NIOSH identifies robot risks including crushing, trapping and unexpected contact near people. ^E13 Physical deployment requires a separate safety-engineered project with competent local professionals, guarded trials, independent interlocks and emergency stopping; simulator reward is not a safety system.

Stop-building rule: before a buyer approves a task, sample schema, permission path, scorer and acceptance criterion, do not build a cloth simulator, full digital twin, distributed RL infrastructure, general factory connector or autonomous deployment. The appropriate one-off engineering is the smallest rights-cleared artifact a buyer can actually ingest and evaluate. Technical compatibility with an AI framework is not evidence of demand.

• No textile manufacturer data, legal permission, robot hardware, engineering validation or buyer demand was independently observed. Architectures are recommendations, not ready-made inventory.
• The deterministic example is fully synthetic and educational. Replay correctness and toy reward gains do not demonstrate counterfactual validity or financial impact.
• Mutable documentation has no verified original publication date where date is null. Source notes distinguish editorial dates, rebuilds, release anchors and older version-pinned interfaces. No post-cutoff claim is intentionally used.
• Original DROID and other project claims are authors’ or vendors’ reports. Public baseline availability does not convey resale rights or textile-specific coverage.
• No code was executed, simulation validated, model trained, or empirical evaluation performed in this research assignment.

Scope: research cutoff and access date 2026-10-25. This is a practical screening guide, not legal advice or a legal clearance. No factory files were inspected; asset availability, permissions, costs and buyer demand remain unknown. Known identifies sourced facts; Inference connects them; Recommendation describes proposed actions, not observed industry practice.

Known: Türkiye's Ministry of Trade describes clothing production across Istanbul, Izmir, Bursa, Denizli, Gaziantep, Tekirdag and other provinces. Its home-textile report identifies towel/bathrobe concentrations in Denizli and Bursa, curtains/upholstery in Bursa, and blankets in Uşak; some producers integrate yarn, fabric, design, dyeing, finishing and sewing. These are 2024 ministry publications with mainly 2021–2022 statistics—not a census of your contact's factory. Their quality/competitiveness language is promotional, not independent performance evidence. ^D12D13 DENİB's dated association profile independently confirms its institutional focus on Denizli textile/apparel exports, especially home textiles; it does not establish that a member has licensable AI data. ^D14

Inference: an integrated mill may connect measurements across stages; a sewing subcontractor may only possess brand tech packs and production records. Ask which operations happen here, not merely “Do you manufacture textiles?” A cluster gives routes to introductions, not permission to aggregate members' records. ^D12D13D14

Known, with attribution: May 2025 reporting relays exporter/manufacturer association leaders' concerns about input/labor costs, financing, exchange-rate competitiveness and weaker capacity utilization. These are attributed industry complaints, not independently audited factory results. A January 2024 IHKIB statement describes sustainability-transition work and pressure from brands' price focus. ^D15D11

Recommendation: frame a one-off transaction as a bounded, paid export of an already-existing, approved asset—not a factory digitization project or promised AI transformation. Cost pressure does not prove willingness to sell secrets or buyer demand. Start with unbranded fabric-only inspection images plus existing quality labels, or a documented measurement table, only if rights and usable volume exist. Avoid collecting workers' activity, new recipes or brand collections just to create inventory.

Recommendation: require four separate yeses: access → disclosure → commercial AI use → onward licensing. A login, owner friendship, camera permission or NDA is not all four. Turkish copyright law distinguishes ownership of a physical copy from intellectual rights and requires written, individually specified economic rights; commercial-code provisions address unauthorized exploitation of entrusted work products and unlawful disclosure of manufacturing/business secrets. ^D04D05

Use this rights matrix as a contract-review worksheet, not a declaration of ownership:

Every row is conditional: machine ownership does not automatically settle software contracts; paying a photographer does not by itself answer downstream rights; worker consent cannot authorize a brand's design. Relevant legal anchors are the copyright, secrecy and personal-data rules, but the decisive factory/vendor/brand contracts are unknown. ^D01D04D05

Known: Law 6698 defines personal data as information about an identified or identifiable natural person. It requires purpose limitation, proportionality, retention limits, a processing condition, information to individuals and security; Articles 7 and 11 address deletion and rights. “Anonymization” means impossible linkage even through matching with other data—not simply replacing a name with “Operator 17.” Article 6 separately regulates special categories, including health, union membership and biometric data. ^D01

Recommendation: photograph fabric against a controlled background. Inspect image pixels and filenames, EXIF/GPS, OCR-readable screens, badges, packing labels, reflections, watermarks, audio and linked spreadsheets. A visible logo is not necessarily personal data, but may reveal confidential brand/order information; removing it does not clear design or contract rights. A face photo should not automatically be called biometric identification data; nevertheless an identifiable face is personal data. Prefer excluding people entirely. ^D01D04D05

Known: the 2024 Article 9 amendment introduced a tiered cross-border framework: a valid Article 5/6 processing condition plus an adequacy decision; absent adequacy, appropriate safeguards with enforceable rights/effective remedies; absent those, specified incidental exceptions. Safeguards include Board standard contracts, approved binding corporate rules and approved undertakings. The old Article 9(1) continued in parallel only until 1 September 2024. The implementing bylaw entered into force on 10 July 2024. ^D01D02D03

Known: making data accessible to a foreign controller/processor counts under the bylaw's transfer definition. Standard-contract text must not be altered; signed contracts must be notified within five business days, with authorization documentation. Safeguards also address onward transfers. One-off frequency is a description field, not automatic eligibility for an exception. ^D02D03

Recommendation: before foreign cloud storage, annotation, remote review or delivery, have Turkish counsel map exporter/importer roles, destination, subprocessors, processing purpose, lawful basis, transfer mechanism and filing responsibility. Do not assume an EU GDPR standard clause substitutes for the Turkish form, or assume adequacy exists for the chosen country. Verify the Board's current position at signing. A “global perpetual AI license” cannot erase people's statutory rights. ^D01D02D03

Known: EDPB guidance makes GDPR territorial scope processing-specific: an EU establishment's relevant activities can bring processing within scope regardless of where individuals are located; non-EU actors may be covered when offering goods/services to people in the Union or monitoring their behavior there. Citizenship alone is not the test. A non-EU factory is not automatically directly subject simply because it has an EU customer; an EU controller can impose processor obligations contractually, and international-transfer rules may also matter. ^D06

Recommendation: if GDPR applies, counsel should determine purpose/legal basis, transparency, controller-versus-processor roles, applicable processing agreements, retention/rights and transfer routes separately from KVKK. Selling fabric photographs to a European company is not itself proof that every Turkish factory activity falls within GDPR. ^D06

Known: ESPR is a framework. The Commission's 16 April 2025 working plan prioritizes textiles/apparel with 2027 indicative adoption of product measures. Its textile DPP page says textile-specific passport details will be established through a future delegated act, indicating Q4 2027 planned adoption—not an already-effective universal textile passport mandate. Dates can change. ^D07D08 A DPP is structured product information accessed through a data carrier, such as a QR code; it is not permission to scrape or resell confidential supply-chain records. ^D08

Inference: requests for composition, product IDs, source evidence and lineage could make well-documented data more useful. That supports asking about existing traceability systems, not claiming AI firms must buy this dataset or selling an invented “DPP compliance database.” No evidence here establishes such demand. ^D07D08

Known: AI Act Article 53 places copyright-policy and public training-content-summary obligations on general-purpose AI model providers. The Commission's July 2025 guidelines distinguish such providers from other actors and narrow-purpose models; the official Article 53 text accessed identifies its consolidated version as 27 July 2026. ^D09D10 Recommendation: supply accurate origin, license and rights-reservation information that an in-scope buyer can use; do not label every dataset seller a GPAI provider or promise that buying licensed data alone proves AI Act compliance. Negotiate permitted regulatory disclosures without casually publishing trade secrets. These rules do not establish a buyer budget. ^D09D10

Recommendation — ask before exporting:

• Reality: which asset exists today, who maintains it, on what system, over what dates? Digital originals or paper? Can an authorized employee export without new software or vendor permission?
• Unit: one image, fabric, roll, lot, test or production run? How many distinct units rather than near-duplicate frames? Which materials, constructions, machines and lighting conditions?
• Meaning: what does each column/label mean, in Turkish and English? Units, tolerances, test method, instrument, calibration, missing-value codes? Who made the label and how was disagreement handled?
• Linkage: can an image genuinely join its test/defect record? Can repeat shots, rolls and lots be grouped for non-leaking train/eval splits? What fails or is absent?
• Rights: origin, author, brand restrictions, OEM/software export terms, personal/confidential content, subcontractors and authorized signatory? Obtain clause references through secure counsel review, not public chat.
• Burden: staff time, redaction effort, export fee, release approvals, continuing support expected? Who is paid, invoices and receives funds? Tax treatment remains for local advisers.

Exact request, to send privately:

Please do not send production folders, employee information, customer names, recipes or confidential designs. First send only an authorized description: asset type, approximate distinct-unit count, date range, format, columns/units, label definitions, capture/test procedure, who created it, and which third-party approvals may be needed. If an authorized company officer clears a small disclosure, please prepare up to 12 unbranded, person-free fabric images already owned/cleared for this purpose, including ordinary and defective examples only if already available and permitted, plus a matching redacted table of at most 20 rows. Use new sample IDs; remove original IDs, brand/customer/order/operator identifiers, prices, locations and confidential settings, including embedded metadata. Preserve non-confidential technical units and explain every removed field. Include one blank/redacted form showing the schema, capture conditions, label source and known limitations. Confirm in writing that this sample may be reviewed solely for feasibility under an agreed NDA; no training, onward sharing or public upload is authorized. If real samples cannot be cleared, provide a clearly labelled invented schema and no real records.

The numbers above are proposed inspection limits, not a claimed available dataset or statistically representative sample. An invented schema explains structure; it is never evidence of real quality, volume or marketability. Keep any approved originals in a restricted data room. Do not upload proprietary samples to public AI tools. ^D01D04D05

Known: the WIPO-hosted Turkish copyright text requires written, individually specified economic rights (Article 52), addresses further transfers (Article 49), and separates copy ownership from IP (Article 57). It is a historical consolidated English text, not a substitute for checking current authoritative Turkish law. ^D04

Recommendation: ask counsel to turn the following into a short, asset-specific agreement:

1. Parties and chain: actual licensor, authority to sign, agent's written marketing/collection mandate and commission; list necessary brand/OEM/author approvals. An NDA protects confidentiality, not a missing license.
2. Deliverable: immutable version, manifest/hashes, formats, unit counts, label schema, redactions, exclusions, acceptance window and material-error remedy. No invented quality warranty.
3. Uses separately: paid feasibility evaluation; commercial model training/fine-tuning; testing/benchmarking; preprocessing/annotation/derivative features; deployment and distribution of trained models. State whether public benchmark publication is forbidden or specifically licensed.
4. Onward rights separately: named buyer affiliates/contractors under confidentiality; raw-data redistribution; sublicensing and marketplaces. A buyer wanting raw-data resale needs that express right throughout the chain—not merely permission to train.
5. Term and economics: territory, nonexclusive/exclusive field and duration, fee/payment trigger, currency/tax/invoicing, who pays preparation. Avoid broad exclusivity over unrelated factory records.
6. Retention and exit: evaluation deletion, production retention, backups, subcontractor deletion, complaint/takedown workflow, future training stops and treatment of existing weights/outputs. Do not promise technically effortless model unlearning or immunity from statutory deletion duties.
7. Security and liability: authorized users, storage countries, no public uploads, incident cooperation, evidence-backed rights representations, dispute procedure and proportionate negotiated remedies. Commercial terms must not contradict mandatory transfer safeguards. ^D01D02D03D04

Recommendation — attach a dataset card: collection purpose/date/site scope; asset creators; source systems; permission evidence IDs and permitted-use scope; privacy legal basis/notice/consent record where consent is relied upon, version/date/withdrawal route; approvals without publicly exposing workers; redaction history; units/instruments; labelers/adjudication; duplicates/splits; material and defect coverage; missing data; known bias/limitations; version and maintainer. Keep private rights evidence separate from the public marketing card. Never label unknown provenance “consented.” ^D01D03D04

Recommendation — stop the affected asset or deal when: a contact offers a personal USB export without company authorization; a brand/vendor restriction blocks the intended use; rights holders cannot be identified; secrecy restrictions remain unresolved; personal data cannot be excluded and lawful processing/transfer cannot be established; signatories cannot support the grant; or onward rights conflict with the buyer's required use. These stops follow directly from access, secrecy, copyright and privacy constraints, not hypothetical fines. ^D01D02D04D05

Commercial/technical stops: no existing export, no trustworthy labels/joins, mostly duplicates, cleanup requiring an open-ended integration, unsupported volume promises, or no buyer willing to evaluate the defined asset. These are proposed fast-sale decision rules—not evidence that all Turkish factories fail them. Also stop a pitch resting on “all textiles already need DPP.” ^D07D08

First-call script: “We are exploring one small, authorized data license, not taking your customer files or installing machinery. Which stages do you actually perform? What existing digital records help your team judge quality? Could you describe one fabric record without identifying a customer? Who created the images, measurements and labels? Which brand, vendor or worker-related restrictions apply? Who can authorize a limited review and later commercial AI use? Could that person approve a redacted schema first? What preparation burden and payment arrangement would be acceptable? We have no committed buyer, price or guaranteed sale. Please send no files today; let's identify the responsible technical and commercial/legal contacts.”

Recommendation — ADHD-friendly format, no medical claims: one visible timer, one question per card, short spoken answers, physical objects and written outputs; skipping/repeating a card is allowed. Use household/unbranded swatches, not proprietary factory assets.

1. Your roommate can download 50,000 brand garment photos. Sell tomorrow? No. Download capability proves neither copyright grant nor contractual disclosure authority. Identify company signatory and brand/creator restrictions first. ^D04D05
2. Factory owner approves CCTV because “workers are mine.” Cleared? No. Workers have personal-data rights; the employer is not owner of their identity. Exclude people or assess lawful purpose, notices, rights and transfer requirements. ^D01D02
3. Names removed; exact shift times and operator code remain. Anonymous? Not established. Other data may reconnect the record to a person. Assess linkage or remove the fields; do not assert anonymization. ^D01
4. Fabric-only images are cleared; export comes from licensed inspection software. Enough? Not yet. Review OEM/software terms and export/label rights; do not presume either vendor ownership or factory freedom. This is a contract fact still to establish. ^D04D05
5. An American buyer only views personal-data samples remotely on a Turkish server. No transfer? Wrong. The bylaw includes accessibility to an entity established abroad, not only downloading. Establish the applicable route before access. ^D02
6. Buyer may train but wants to sell the raw images onward. Same permission? No. Training and raw-data sublicensing are distinct proposed grants; verify downstream authority explicitly. ^D04
7. A prospect says every textile already requires an EU DPP, therefore this dataset is guaranteed to sell. Agree? No. Textile-specific requirements are staged through future delegated measures; indicative adoption is not a universal effective obligation, nor proof of dataset demand. ^D07D08
8. A narrow defect-data seller must publish a GPAI training summary? Not merely for selling data. Article 53 concerns GPAI model providers; determine actual role. The buyer may request provenance for its own duties without transferring those duties wholesale to every supplier. ^D09D10

• Factory agreements, asset counts, actual resale rights, authority to sign, buyer interest and budget are unknown. No proprietary data was collected or displayed.
• Legal sources identify issues, not a guarantee of lawful commercialization; current Turkish originals, applicable foreign law and actual contracts need qualified review.
• Undated institutional pages are qualified as such and corroborated where possible by dated documents; no material published after the cutoff is relied upon. Plans remain plans.
• No examined source establishes that a small Turkish textile dataset has sold to an AI buyer at any particular price. Regulatory pressure and industrial capability are not evidence of demand.