Stated goal (verbatim)

I have an idea about a product that is like an AI-based assistant, specifically for GCC e-commerce, Amazon, Noon, and other platforms' resellers. It's basically a guide that makes people work based on my experience as an Amazon seller. It's something that walks through the entire process and helps optimize sales on Amazon and other marketplaces. It's basically this unified marketplace platform that helps with everything, like an end-to-end overwatch. Hicksfield is trying to be an end-to-end content creation machine. Here, it's an end-to-end Amazon selling platform that helps you do everything.

Are there products like this? I know that there are, for example, groups, YouTube channels, and social communities that give all of the guides, etc., related to that, to get familiar with the subject: read our lore, Amazon, how Amazon works, study. Is there an AI assistant product that helps with this?

("Hicksfield" = Higgsfield — the structural analogy Ilia already uses, not a competitor.)

Grill round 2 added the decisive clarification:

in general it shoudl cover droppshipping or sellers business, focus in those who do business by specifically importing goods how i do at SHIO. AMazon is the main one, but we shoudl ecopmpass busienss e2e.

Refined goal (post-Grill)

Establish what already exists for someone running a goods-importing resale business in the Gulf, and where an AI operator could own the parts nobody covers.

"End-to-end" is now defined as the business, not the console: company formation and trade licence → product selection → sourcing → freight → customs and IOR → VAT registration → product compliance and labelling → listing → advertising → operations → reorder. Amazon (.ae/.sa) is the primary channel; noon.com secondary; dropshipping covered but not the focus.

Three questions, descending in value:

1. What covers this spine today? Global seller SaaS starts at "you already have a product." Who, if anyone, covers the half before that — and the GCC-specific version of it?
2. What is the technical ceiling? 006 established the pattern is exports in, recommendations out, human pushes the change. Confirm or update, and map precisely which operations a third party can write back to on Amazon and Noon.
3. What is the smallest version that survives contact? Given no finding will stop the build, the Brief's job is walls, forks, and a shippable v1 — not a verdict.

Consumer of this Inquiry

• Reader: Ilia. Product/design background, owns SHIO LLC (Dubai), one SKU live on amazon.ae, completed the Dubai import journey himself, runs Claude Code daily, and has already built the lore corpus (100 practitioner sources → 385 published concept pages).
• Decision it gates: founding an AI startup on this thesis, and applying to GCC accelerators (added mid-Grill, 2026-09-05).
• Consequence of the startup framing: the Brief is now read by two audiences. Ilia needs walls and forks; a selection committee needs a market, a wedge, a "why now", a moat and a reason this founder. Every Area must therefore produce at least one claim that survives being quoted at a partner, which means primary sources and dated figures, not listicles.
• What "good" looks like: a named competitive set described by what each product does, not what its landing page claims; a definite Noon answer; an import-spine map showing uncovered steps; and a v1 narrow enough to demo.
• What fails: a survey of Amazon tools. A TAM with no seller in it. Any accelerator claim without a deadline and an eligibility rule.

Grill outcomes (binding)

• Beachhead: the aspiring GCC seller (0 → first sale), specifically the importer. The other three archetypes are cut from v1 scope. Scaling sellers, brands entering GCC and the SHIO-only case may appear in the Brief as expansion paths, not as v1 users.
• Differentiators being stress-tested (Ilia's pick): GCC/Arabic localisation · Noon + Amazon unified · guided process rather than a toolbox. He explicitly did not claim the lore corpus as the moat, which is the right call — but a Scout still tests whether a bare frontier model already knows this domain, because "guided process" implies encoded knowledge.
• "Based on my experience as an Amazon seller" is demoted. 006's own shio-context.md: "Ilia is a product/design person, not (yet) an Amazon operator." One SKU, one marketplace. The credible version of this claim is the import journey he personally completed in Dubai, which is exactly the spine above — not Amazon operating depth.
• v1 ceiling accepted: watch, advise, draft. The product reads exports/APIs, drafts listings and campaigns, enforces the sequence; the human clicks Apply. A2 and A6 are therefore design constraints, not kill gates.
• Kill criteria: none. Ilia pre-committed that no finding would stop him. Recorded, and the Inquiry is re-aimed accordingly. The Consolidated Brief will nonetheless carry a "strongest case against" section so the argument exists in writing once.
• Churn: design for graduation. The Brief must propose how a beginner becomes a retained operator inside the product — the thing a course cannot do and Helium 10 will not.
• Surface: undecided by design. A1 covers SaaS competitors, A4 covers community/bot competitors; the Brief recommends a surface based on where the beachhead already lives.
• Scope addition, unopposed: the Russian-speaking Dubai entrepreneur diaspora is treated as a distinct segment in A4/A5 — it is distribution Ilia has and an incumbent does not.

What is already established (do not re-research)

Inquiry 006-amazon-shio-guide (2026-08-19; 100 practitioner sources + web research) settled:

• The mid-2026 AI seller-tool stack — Helium 10 (Cerebro/Adtomic/Listing Builder/MCP), Data Dive, Jungle Scout, Superfuel (~$49/mo), Pixie, ZonGuru, sellerboard, AdLabs/Perpetua, Amazon-native (Experiments, Budget Rules, AI video, ads agent). Update and reclassify — do not rebuild.
• The integration ceiling. Amazon blocked OpenAI/Anthropic crawlers Aug 2025, sued Perplexity Nov 2025; Amazon Ads MCP is partner-only; every documented workflow starts from a hand-exported file.
• Adoption is low — 5–10 % of sellers using AI meaningfully, ~25 % "in some form"; practitioners call the next 12–18 months a window.
• amazon.ae mechanics — 15 % referral (home/personal care), 5 % VAT on top of every fee, 5 % customs duty, no Vine, no Brand Referral Bonus, Attribution contested, Rufus/Alexa for Shopping not confirmed on .ae, White Friday + Ramadan calendar, AUH1 fulfilment centre opened Sept 2025, amazon.sa cross-border enrolment friction.
• Three parked gaps that seeded this Inquiry: Arabic listing generation absent from all 100 sources; nobody totals a small seller's AI-stack cost; amazon.ae vs noon share exists only as third-party aggregate.

Scope

In: amazon.ae / amazon.sa / noon.com · AI assistant and agent products for marketplace sellers worldwide, with GCC availability as the discriminator · the full import spine (licence, sourcing, freight, customs/IOR, VAT, ESMA/SFDA conformity, Arabic labelling) · Amazon's and Noon's own first-party seller AI · SP-API / Ads API / Noon API and automation ToS · courses, channels, paid communities, agencies and bots, including RU-diaspora Dubai · whether a curated corpus beats a bare frontier model.

Out: shopper-side AI except where it changes seller behaviour · own-store platforms (Salla, Zid, Shopify) unless they have crossed into marketplace management · Egypt, Israel, Morocco, Turkey (GCC = UAE + KSA first; other GCC states noted opportunistically) · food and q-commerce (Talabat, Instashop) · the three cut archetypes as v1 users · building anything · SHIO's own account state, which remains a task for Ilia in Seller Central, not for a Scout.

Success criteria

1. Every product plausibly claiming end-to-end AI assistance for marketplace sellers, named and classified advisor / co-pilot / operator by what it can actually write back.
2. A definite answer on Noon: partner or seller API, yes or no, with a primary source.
3. A definite answer on what Amazon's own seller assistant does today and in which stores.
4. A map of the import spine showing, step by step, what software covers it and what does not.
5. Fork criteria (replacing kill criteria): which finding forces which change of shape.
6. A v1 recommendation narrow enough to build, plus the strongest case against it.
7. A shortlist of GCC accelerators with eligibility, cohort dates, terms and what each optimises for — enough to decide where to apply and what the application must claim.

Areas

A2 and A6 are design-constraint Areas; A8 serves the startup framing; the rest are landscape. 006's grounding dissolves the dependency between them, so all eight run in parallel.

Source legend convention

Per-Area legends prefixed by Area: [A1-S1], [A3-W4]. Claims inherited from the predecessor Inquiry keep their original tags and are cited as 006:[L20] / 006:[W1].

Inquiry: is there an AI assistant that operates a GCC marketplace business (amazon.ae/.sa, noon.com) end-to-end — and if not, is the gap worth building into? Purpose: build / no-build decision for Ilia. Personal project. Predecessor: 006-amazon-shio-guide supplies the corpus, the .ae mechanics, and three parked gaps that seeded this Inquiry.

Log

2026-09-05T·· · — → Scoped · Goal captured verbatim. Prior Inquiry 006 read first
                    (findings 07-ai-tooling-for-sellers, 08-amazon-uae-specifics,
                    notes/shio-context) so no Scout re-derives the mid-2026 US tool stack,
                    the .ae rate card, or the "Amazon blocks AI agents" ceiling. Those are
                    carried into SCOPE.md as established, with instructions to update
                    rather than rebuild. Six candidate Areas.
2026-09-05T·· · Scoped → Triaged · Classified FULL. Stakes = build/no-build. Two Areas
                    (A2 platform risk, A6 technical ceiling) are kill-criteria and the rest
                    are landscape; grounding from 006 dissolves the dependency so all six
                    can run in parallel. Register EN. Output must be Consolidated Brief +
                    SUMMARY, and BenchmarkBrain validation is warranted before any build
                    handoff (per ORCHESTRATOR step 8: the Brief gates spend).
2026-09-05T·· · Triaged (open) · Intake returned ALL FOUR user archetypes selected
                    (aspiring seller / scaling seller / brand entering GCC / SHIO dogfood).
                    Flagged as the primary Grill target: four archetypes is four products,
                    four competitive sets and four go-to-market motions. Grill round 1
                    must force a beachhead.
2026-09-05T·· · Triaged → Grilled · Two rounds. R1 forced a beachhead (aspiring GCC seller,
                    other three archetypes cut), demoted "my experience as an Amazon seller"
                    against 006's own note that Ilia is "not (yet) an Amazon operator", and
                    accepted the watch/advise/draft ceiling for v1. R1 returned **no kill
                    criteria** — "nothing!" — which re-aimed the Inquiry from go/no-go to
                    walls-and-forks, and added a mandatory "strongest case against" section
                    to the Brief. R2 produced the decisive scope change: "end-to-end" means
                    the *import-and-resell business*, not the Seller Central console —
                    licence → sourcing → freight → customs → VAT → compliance → listing →
                    ads → reorder. That created A7 (import spine), the differentiation Area.
                    Also: design for graduation; surface undecided by design; RU-diaspora
                    Dubai scoped in unopposed.
2026-09-05T·· · Grilled (amendment) · Mid-turn user message: intends to found a startup and
                    apply to GCC accelerators. Added A8 (accelerators & funding), a
                    "Consumer of this Inquiry" section naming a second audience (selection
                    committee), and success criterion 7. Consequence recorded in SCOPE.md:
                    every Area must yield at least one claim that survives being quoted at
                    a partner — primary sources and dated figures, not listicles.
2026-09-05T·· · Grilled → Authorized · 8 Area Briefs written, 8 Scouts dispatched in
                    parallel. Model split by task type, not frugality: A1/A2/A6 on opus
                    (they require skepticism about what vendors and platforms actually do
                    vs claim), A3/A4/A5/A7/A8 on sonnet (enumerative research against
                    primary sources). A2 and A6 are design-constraint Areas; A7 carries the
                    differentiation claim; A8 serves the startup framing.

Findings log

2026-09-05T·· · A4 knowledge-layer RETURNED. Three disconnected ecosystems, not one
                    incumbent: a mature English YouTube/agency layer monetising services
                    not subscriptions; a thin solo-founder GCC course layer ($65–$1,000, no
                    published cohort sizes) sitting on a free and adequate first-party
                    layer; and a large RU-Dubai population with the Amazon/Noon sub-niche
                    essentially unoccupied. No AI bot found sold into a GCC or RU-diaspora
                    seller audience. Retention in this layer is bought via human
                    accountability ($10K–$50K/yr masterminds), never designed into content —
                    direct ammunition for "design for graduation".
                    ESCALATION (time-critical, verified by Orchestrator against the primary
                    source, correcting the Scout's figures): Amazon Creators Foundry
                    (Creators HQ × Amazon Ads) closes 2026-09-09 23:59 GST. 20 places,
                    $10,000 Sponsored Ads incentive each + up to $100,000 DSP bonus media,
                    6-month account manager, FBA setup, Golden Visa nomination assistance.
                    Eligibility — UAE-based creator or UAE-registered business, valid Dubai
                    trade licence, ready-to-scale product — which SHIO already satisfies.
                    Relevant to SHIO, NOT to the AI startup; surfaced to Ilia immediately
                    rather than held for the Brief.
2026-09-05T·· · A3 gcc-mena-supply-side RETURNED. ESCALATION: Noon appears to operate
                    OAuth-based third-party integrator access (primary source dated
                    2026-07-27), which revives the "Noon + Amazon unified" pillar the Grill
                    had treated as at-risk. But the same Finding is evidence AGAINST the
                    pillar as a differentiator: Unicommerce (India-HQ, MENA ops) documents
                    a live amazon.ae + noon.com integration in one dashboard, and Omniful
                    (UAE/KSA, $5.85M seed) claims a native Noon integration. Prior art
                    exists; it is simply not aimed at beginners. Arabic listing-generation
                    AI still not found as a product — 006's absence finding reinforced, not
                    reversed; one UAE agency marketing "Arabic AI optimisation" turned out
                    to be manually prompting ChatGPT. No regional player serves pre-revenue
                    sellers. Relayed the Noon escalation to A6 mid-flight.
2026-09-05T·· · A5 demand-side RETURNED. No platform or regulator publishes an audited
                    count of active third-party sellers on amazon.ae, amazon.sa or
                    noon.com — the Area's headline is an absence. Best flow number:
                    Dubai Traders (DET + Dubai Chambers) onboarded 2,400+ new e-commerce
                    sellers via Amazon and noon in Dubai over 12 months (Sept 2024–Nov
                    2025). Platform targets (amazon.ae 100k companies by 2026; amazon.sa
                    40k SMEs by 2025) are pledges, not achieved counts. Importer-seller
                    profile essentially absent from public sources — reported as a gap, not
                    inferred. CLEAN NEGATIVE, quotable: every government programme found
                    (Monsha'at–Amazon, Dubai Traders, Kafalah, MBR Fund) funds SELLERS or
                    general SME finance, never a software company — so A8's funding path
                    cannot lean on seller-enablement subsidy.
2026-09-05T·· · A1 global-ai-seller-landscape RETURNED. ~25 products read and classified
                    advisor / co-pilot / operator by write-back capability. **No end-to-end
                    operator exists anywhere — the Inquiry's central thesis holds.** Only
                    four products genuinely write to a platform, and every write is scoped
                    to ads, price, or a single listing field: Pacvue Agent (2026-04-14,
                    Amazon Ads only), CommerceIQ (enterprise), Profasee Ultra (price+bids,
                    from $698/mo), Superfuel (title + Item Highlights, from $0/$49).
                    Headline: the market leader cannot write — Helium 10's agent "Helium"
                    (launched 2026-08-24, five days after 006 closed) says in its own
                    release it is "focused on analysis and recommendations… including write
                    execution capabilities soon"; its MCP connector is read-only.
                    THE STRATEGIC LINE OF THE INQUIRY SO FAR: **the GCC gap is
                    go-to-market, not API.** SP-API supports AE (A2VIGQ35RCS4UG) and SA
                    (A17E79C6D8DWNP); noon runs a public partner API; yet ZonGuru and Scale
                    Insights exclude GCC in writing, Jungle Scout Catalyst lists 8 non-GCC
                    markets, only SellerApp lists UAE+KSA — and **zero of ~25 products
                    mention noon.com at all.** Pre-product is empty: only Jungle Scout's
                    Supplier Database touches sourcing; nothing covers freight, customs,
                    VAT, conformity, labelling or entity setup. Three claim-vs-evidence
                    contradictions resolved toward docs (Helium 10, Atomic One, Amazon
                    Seller Assistant — the last handed to A2).
2026-09-05T·· · A8 accelerators-funding RETURNED. 25 instruments checked against primary
                    sources. Structural finding: every UAE free-zone accelerator (Hub71,
                    in5, Dtec) requires a NEW dedicated trade licence — SHIO LLC cannot
                    transfer in, only sit as holding parent. No GCC programme runs an
                    e-commerce/seller-tooling vertical except 500 Global/Sanabil (Riyadh,
                    not Dubai — Scout corrected its own assumption), which is
                    traction-gated. Techstars has no live GCC accelerator. Best immediate
                    fits: MBRIF Innovation Accelerator (federal UAE, Cohort 13 portal live
                    now, NO incorporation gate — SHIO clears it today); Amazon SP-API /
                    Solution Provider Portal + AWS Activate (self-serve, no deadline, SHIO
                    qualifies now); Sheraa Startup Blueprint (free, rolling, idea-stage);
                    Antler MENAP Dubai ($180k/11% SAFE, cohort starts 2026-10-11 — tight);
                    Hub71 Access (Cohort 21 deadline Feb 2027, needs new ADGM subsidiary +
                    relocation). Hard gate found: KAUST TAQADAM requires a co-founder,
                    which a solo founder fails.
                    CROSS-AREA NOTE for synthesis: A8 found NO accelerator with a confirmed
                    Golden Visa pathway, while A4's Amazon Creators Foundry offers Golden
                    Visa nomination assistance directly. The fastest residency route found
                    in this Inquiry runs through the Amazon programme, not an accelerator.

2026-09-05T·· · A2, A6, A7 STALLED. All three Scouts killed mid-run by a session rate limit (HTTP 429, "resets 6:50pm Asia/Dubai"), along with six sub-agents they had spawned. No finding files written for these Areas. RECOVERY: two of A6's sub-agents completed and reported before the kill, and their output was the most valuable evidence of the Inquiry. Preserved by the Orchestrator to notes/recovered-noon-api-evidence.md and notes/recovered-noon-terms-and-verticals.md rather than lost. Substance: noon's partner API is real, official and publicly documented (official PyPI SDK noonapi published from a noon.com address, 51 endpoints, no login wall, OAuth flow built explicitly for multi-tenant integrators, an in-product "noon-approved integrator" dropdown). AND noon Seller T&Cs clause 8.3 makes "any automated tool of any type to scrape data" a material breach that SURVIVES TERMINATION (clause 7.5) — the structural twin of Amazon's Agent Policy. Also corrects A3: Anchanto is a demonstrated negative on noon, not an unverified one.

2026-09-06T·· · A6 technical-ceiling RECOVERED FROM DISK. The original Scout wrote its
                    Finding (37KB, 45 sources, all sections) at 15:02 before the rate limit
                    killed it — the "failure" notification was misleading. Orchestrator
                    stripped a stray tool-markup artifact from the file tail; content
                    untouched.
                    **A6 OVERTURNS THE INQUIRY'S WORKING ASSUMPTION.** Its verdict: a third
                    party CAN honestly promise a credentialed *operator*, not merely an
                    advisor, on BOTH marketplaces — create/edit listings, set price and
                    stock, build campaigns, change bids, add negatives, pull Search Query
                    Performance on amazon.ae/.sa, and run catalogue/price/stock/offers/
                    partner-fulfilment on noon across AE/SA/EG.
                    Specific corrections to 006 and to the Grill's accepted ceiling:
                    (1) "exports in, recommendations out" described PRACTICE, not the
                        ceiling — practitioners hand-export because they never built an
                        SP-API app, not because the API refuses.
                    (2) Amazon Ads MCP is NOT partner-gated; a Direct Advertiser self-applies
                        in ~1 business day, and the EU endpoint advertising-ai-eu serves AE
                        and SA profiles.
                    (3) SP-API registration gates on corporate identity + a live public
                        website, NOT revenue or an existing seller base — a pre-revenue
                        startup qualifies. None of the operations this product needs
                        requires a restricted PII role, so the heavy architecture review is
                        avoidable by scope choice.
                    (4) The crawler block and the API are DIFFERENT DOORS: robots.txt on
                        .ae/.sa blocks ~101 agents incl. ClaudeBot/GPTBot, closing the
                        shopper-facing web, and says nothing about SP-API.
                    New walls found, all design-relevant:
                    - SP-API AUP 4.4 forbids aggregating data across sellers to provide or
                      sell to others → CROSS-SELLER BENCHMARKING IS NOT BUILDABLE. Kills any
                      network-effects/"compare to other GCC sellers" pitch.
                    - AUP 3.3/3.7 forbid holding a seller's credentials → OAuth or secondary
                      user permissions only.
                    - AUP 2.4 makes AI disclosure a CONTRACTUAL obligation.
                    - Customer Feedback API is unavailable on AE/SA → reading reviews on
                      amazon.ae is possible neither by API nor by crawler. A real,
                      GCC-specific gap.
                    - noon has NO advertising, reviews, buyer-messaging or search-term API;
                      reporting is its weakest surface.
                    - noon integrator access has a HUMAN GATE ("contact noon"), no published
                      eligibility criteria and no SLA.
                    - Agent Policy verbatim text NOT obtainable (BSA is a JS app returning
                      sign_in); widely-repeated blog specifics (12-month audit trail,
                      180-day retention, 20% price-change rule, certification programme) are
                      UNVERIFIED and must not enter the Brief.
                    CROSS-AREA: noon's Content API is bilingual BY DESIGN — "Submit
                    bilingual content (English and Arabic) and track completeness and QC
                    status independently per language." The Arabic *plumbing* exists
                    first-party; what remains unserved is generating good Arabic copy.
                    Synthesis must sharpen the Arabic differentiator accordingly.
2026-09-06T·· · Workflow #1 (wf_326f1998-fbf) TOTAL FAILURE. All 14 agents killed by the
                    same session rate limit (reset 11:50pm Asia/Dubai); 1.7M subagent tokens
                    and 717 tool calls spent, zero outputs written, notes/raw/ left empty.
                    Lesson recorded below. Relaunched leaner as wf_c94099cc-29f covering
                    only the two genuinely missing Areas (A2, A7) with 7 researchers + 2
                    writers, instructing researchers to create their notes file EARLY and
                    append as they go so partial work survives interruption.

2026-09-06T·· · Authorized → FINDINGS-COMPLETE. Workflow wf_c94099cc-29f returned 9/9 agents successful; A2 and A7 written. All eight Areas now have Findings. A2 (1,261 lines, 83 sources): Amazon has shipped extensive free seller AI and almost none of the agentic layer has reached the Gulf. Seller Assistant is documented on .com/EU/.in and ABSENT from the .ae and .sa help catalogues — demonstrated against a working same-host control. Amazon's free AI ad-creative suite (Creative Agent, image, video, audio) lists four country sets and UAE/KSA appear in none, including on Amazon's own Arabic-language UAE page — "the clearest open lane found." Live in the Gulf: Enhance My Listing, Ads Agent behind an AMC+DSP+account-executive gate, Campaign Manager beta by invitation, and the global Ads MCP. Amazon's own disclaimer scopes Seller Assistant to "selling in the Amazon store," excluding professional/legal/financial advice — so the entire pre-Amazon import spine and every pre-account user are outside it BY AMAZON'S OWN WORDS. noon documents exactly ONE AI feature (a category classifier) across 243 help articles and its whole API index, and STATES IN WRITING IT WILL NOT TRANSLATE SELLER CONTENT. Amazon's posture is regulated welcome, not hostility: no listing fee, no revenue share, an LLM-addressed page naming Claude — alongside §1.3 "Use of Agents" in the UAE developer contract, a 25-seller cap for unlisted apps, and a security floor raised 2026-08-25. Resolves A1's Seller Assistant contradiction; partly closes A6's Agent Policy gap from the developer side. SHELF LIFE: Amazon Accelerate runs 22–24 Sept 2026 with the Seller Assistant author speaking — re-check A2 §1.2/§2.1 after 2026-09-24. A7: the spine is 22 steps and THE FIRST 18 ARE EXECUTED BY HUMANS. The only software upstream of the listing is government transaction portals (Dubai Trade/Mirsal 2, EmaraTax, SABER, FASAH) that execute a decision already made, plus single-vendor configurators that only ever return their own answer. Amazon's answer to the entire middle leg is the Service Provider Network — a directory of 850+ humans and a contact form. DEMONSTRATED ABSENCE: all 107 English URLs on sell.amazon.ae enumerated; zero pages on importing, customs, duty, HS codes, IOR, freight, conformity, Arabic labelling, VAT or the trade licence. Cost figures nobody publishes: AED 8,100–18,200 fixed setup and AED 23,000–34,700 all-in to a first sale on a stated worked example, both explicitly floors, with the residence-visa fee and accredited-lab test report parked rather than guessed. Freight given as a band because five 2026-dated sources disagree ~5x on the same lane. THREE A7 ESCALATIONS: (1) KSA incorporation is capital-impossible for the beachhead — MISA's 13th-edition Investor Guide sets 100%-foreign trading at SAR 30,000,000 capital plus a SAR 200–300m five-year commitment; noon publishes an automatic opt-out UAE→KSA lane needing "no KSA trade license or KSA VAT registration" because noon is the importer, while Amazon publishes no path at all (/grow/expand-internationally → 404). (2) Arabic is a STATUTORY DUTY, not a localisation feature — Federal Law 15/2020 as amended requires Arabic invoices and Arabic product information, specifications, contract terms and warranty, with penalties to 2 years custodial and AED 2m. That statute is what 006's parked Arabic gap actually sits on. (3) Amazon asks for NO TRN and NO conformity certificate at registration, so a seller can be fully live on amazon.ae while non-compliant with the FTA, MoIAT and the Ministry of Economy — the structural argument for the product, presented as a fork because whether FBA/FBN absorbs the mainland role is unresolved by any government or marketplace source. 2026-09-06T·· · Findings-Complete → (synthesis prep) · Launched wf critique-and-cluster: four adversarial lenses over all eight Findings (cross-Area contradictions · load-bearing claims on weak evidence · unmet success criteria and tooling-induced gaps · steelman the case AGAINST), then a cluster proposal. Per ORCHESTRATOR step 7 the synthesis itself stays in the main thread and the clusters go to the human via AskUserQuestion before the Consolidated Brief is drafted.

2026-09-06T·· · Findings-Complete → SYNTHESIZED. Critique pass (4 adversarial lenses +
                    cluster proposal) saved to notes/critique/. It caught the Inquiry's
                    worst hole: the two central negatives (A1 "no end-to-end operator exists
                    anywhere", A3 "no regional player serves pre-revenue sellers") had been
                    reached WITHOUT testing the three closest candidates, because A2 and A7
                    discovered them after A1 and A3 had returned. Closed:
                    - **qeen.ai** — not a competitor. AI performance-marketing agency for
                      Shopify/Magento/Salla/Zid storefronts; zero marketplace integration
                      (0 hits for "amazon"/"noon" across its site), 0 of 18 spine steps,
                      $4,000/mo ad-spend floor. Its Arabic dialect generation is genuinely
                      strong — a graduation partner, not a rival. A7's escalation conflated
                      three separate announcements.
                    - **Portmind** — B2B logistics infrastructure layer, not the seller's
                      journey. Does not close the spine gap.
                    - **Jarvio** — classified by the Orchestrator directly (its Scout died on
                      a rate limit). IS a real operator: "takes action: pricing changes,
                      listing edits, inventory orders, customer responses, ad adjustments";
                      $149/brand/month; declares .ae/.sa/.eg across 24 stores per Amazon's
                      own Appstore record. BUT setup requires "connect your Amazon account"
                      and it covers ZERO pre-listing spine. **A1's headline is narrowed, not
                      retracted:** the defensible claim is "no product covers the business,
                      and every operator starts after the hard part is done."
                    Critic's method finding, adopted: the Areas that LOST WebSearch (A2, A6,
                    A7) produced the strongest evidence — forced onto contracts, sitemaps,
                    llms.txt, government PDFs and controlled probes — while the Areas that
                    spent their search budget (A3, A4, A5) lean on syntheses and
                    aggregators. Discount by METHOD, not by return order.
                    Human confirmed at the synthesis gate: two-stage product with an explicit
                    graduation edge · delete the TAM and lead with the absence · all six
                    clusters in the proposed order · pricing shape deliberately left open
                    ("don't go that deep").
                    Artifacts: consolidated-brief.md (six clusters, six contradictions
                    resolved, mandated "strongest case against" with a verdict on which
                    objections actually land), SUMMARY.md (plain-language front door),
                    tasks.md.
2026-09-06T·· · Process correction from the human, recorded: "the research wastes too many
                    tokens… you are going too deep." Four workflows and eight Scouts had
                    spent ~5M subagent tokens by this point. Subagent use stopped; Jarvio
                    was classified in the main thread with one fetch, and synthesis was
                    written directly. FRAMEWORK LESSON: DiscoveryBrain has no budget gate.
                    Triage decides Light vs Full but nothing caps fan-out, and the
                    Orchestrator seat has no instruction to report spend or check in on it.
                    A Full Inquiry should declare an agent/token budget at Scope and report
                    against it at each milestone.

Incident

2026-09-05T·· · WebSearch budget exhausted (200/200) with 5 Scouts still in flight. A4
                    used 72 tool calls, A5 66, A3 52. Remaining Scouts retain WebFetch but
                    lose discovery search. Surfaced to Ilia with the settings.json fix
                    (CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION), and instructed A6 to park
                    budget-blocked items explicitly rather than let a tooling limit
                    masquerade as a demonstrated absence. Areas returning degraded must be
                    marked as such in the Consolidated Brief and are candidates for re-run
                    in a fresh session — the Briefs are already written, so a re-run is cheap.

Framework lessons (to fold back into the repo)

1. Explore cannot be a Scout. SKILL.md and ORCHESTRATOR.md step 6 both prescribe subagent_type: Explore for Scouts, but the Explore agent has no Write tool, and the Area Brief lifecycle requires each Scout to produce findings/<area>.md. Dispatched general-purpose instead. The framework should either say general-purpose, or say that Explore Scouts return content for the Orchestrator to write.

2. Kill criteria may come back empty, and that is a real answer. The lifecycle has no handling for a human who pre-commits to building regardless. Recording it and converting kill criteria into fork criteria (which finding forces which change of shape) preserved the Inquiry's value; a "strongest case against" section carries the adversarial load instead. Worth encoding in lifecycle/scope-grill.md.

3. Scope Grill round 2 changed scope more than round 1. The ≥2-round floor earned its keep here: round 1 tuned the segment, round 2 discovered that "end-to-end" meant something entirely different from what the stated goal implied. Evidence for keeping the floor at 2.

ORCHESTRATOR.md orders Triage (step 2) before Scope (step 3); lifecycle/milestones.md has Scoped as the entry criterion for Triaged. They contradict. In practice the intake question ("Light or Full?") has to be asked with the scope questions to avoid two round trips, so the milestone table is the one that matches reality only if "Scoped" means "goal captured", not "SCOPE.md written". Worth resolving in the framework.