{"submission_id":"e0734bcf-f6a9-42e0-be95-f77556f4e23b","submission_revision_id":"b5a690a9-cf75-4cff-a490-23058d2379b8","revision_version":3,"submission_status":"under_review","submitted_at":"2026-10-08T13:41:42.066985Z","submitter_name":"Dan Krieger","submitter_email":"dan@raisi.ai","task":{"status":"submitted","version":3,"extra_fields":{},"submitted_at":"2026-10-08T13:41:42.066985+00:00","industry_domain":"Digital Marketing / Paid Social Campaign Operations","task_short_name":"raisi_fundraising_production_campaign_v2_1","operating_system":"MacOS","task_description":"Version 2.1 — five-concept production campaign; supported adaptive-media profile.\n\nObjective\nStarting with https://www.raisi.ai/, research Raisi's main fundraising offer for U.S. startup founders and turn it into a sourced, reviewed, measurable Meta campaign. Complete product/ICP/competitor exploration, positioning, copy review, native Canva production, read-only measurement audit, experiment design, saved-state reconciliation and authorized publication.\n\nStarting state and environment\nwebsite.txt is the initial business-discovery input. The supplied official logo and execution_context_v2_1.json provide brand and operating constraints; they contain no reference answers. Provision signed-in ChatGPT/Canva and an authorized Meta account with the intended identity, USD currency, documented timezone and existing dataset. Preflight the ability to store five concept ads, three Primary options per ad and three image formats per concept. Use a clean environment for any independent trial. Missing access or incompatible capabilities is an environment problem, not a low model score. This version records capabilities observed in the assisted reference; it was not frozen before an independent self-test.\n\nWork sequence\n1. Define the ICP before selecting positioning: company stage, founder job, triggers, pains, objections, exclusions and assumptions. Analyze the fundraising offer's mechanisms, benefits, plan/service boundaries and CTA journey. Cite official source URLs, dates and locators for material claims; distinguish vendor statements, observations and inferences.\n2. Compare at least two direct competitors and one substitute through official sources. Compare audience, promise, mechanism, proof, boundaries and CTA. Missing website wording is not proof that a competitor lacks a feature. Compare three positioning choices, justify the selected one and develop five distinct, source-supported creative hypotheses, C1-C5.\n3. Write exactly three Primary texts P1-P3, each three paragraphs, 80-150 words and <=1,200 characters, ending with a suitable CTA. Write five below-image Meta headlines H1-H5, <=40 characters, with two direct CTAs and three product/curiosity variants. Write five on-image headline/support pairs: headline <=60 characters and no more than three rendered lines; support <=100 characters. No separate short Meta Description set is required.\n4. Critique claim support, product clarity, specificity, ICP relevance, natural language, CTA expectations and all 15 Primary/concept-headline pairings. Retain genuine drafts, issues, revisions and rechecks. Freeze exact UTF-8 copy and hashes before design. No fabricated testimonials, product UI, investor endorsements, outcomes or funding guarantees.\n5. Create five concepts, each intentionally adapted to 1080x1080, 1080x1350 and 1080x1920: 15 editable native Canva pages in total. Use at least two meaningful layouts and the full official logo with preserved proportions. Keep headline/support objects independently editable. Export 15 opaque full-resolution PNGs and native PDF proof covering all 15 pages (three five-page PDFs are acceptable). Supply design/page IDs, source links, object/editability evidence, dimensions and file hashes. Record actual color/profile metadata; do not claim embedded sRGB certification without evidence.\n6. Inspect all 15 native exports for frozen text, logo, hierarchy, contrast, overlap and clipping. Inspect actual supported Feed, vertical and compact Meta adaptations; record coverage per concept/placement and fix observed core-logo/headline clipping. Distinguish unavailable/unloaded previews from visual passes. Retain before/after corrections. Meta may group format variations or select media adaptively: preserve C_i with H_i and all three P options, store all three correct images for each ad, and record actual groupings/crop settings. Do not claim deterministic per-placement locks or equal exposure when the interface does not provide them.\n7. Audit the real landing/CTA journey and existing measurement setup without submitting forms or producing fake leads. Document intended optimization event, business meaning, dataset, available event-source/freshness/deduplication evidence and attribution setting. Presence in a selector does not prove firing or attribution. Mark untested consent, controlled firing, unique qualified-meeting semantics and end-to-end attribution explicitly, with a concrete verification/remediation plan. Document UTMs, visible redirect/query behavior and unique concept reporting keys.\n8. Design a seven-day learning plan within USD 350: unit of comparison, outcome/denominator, minimum usable evidence, confounders and an inconclusive rule. Provide reproducible low/base/high sensitivity calculations for impressions/clicks/downstream events with labeled hypothetical CPM/CTR/CVR assumptions. Five concept ads with adaptive Primary text, placement and translation delivery are exploratory bundle screening, not an isolated headline A/B test. Do not manufacture a winner or promise statistical power.\n9. Configure one NEW campaign, one ad set and five concept ads in the authorized account. Set USD 50/day for the whole campaign, a separate USD 350 campaign spending cap and a future 168-hour start/end schedule with timezone/UTC offsets. Enforce United States location; distinguish founder suggestions from audience controls and expansion. Keep the homepage destination, all eligible platforms and automatic translations enabled. Verify identity, CTA, objective/event, dataset, attribution, UTMs, all stored English text and 15 selected images. Build intended and observed state records, reopen the editor after saving, repair mismatches and recheck affected fields. A daily budget is an average, not a strict daily ceiling; 50 x 7 alone does not prove a saved cap.\n10. Publish only with separately provisioned live-spend authorization for this instance. The owner's reference-run authorization is already consumed by that reference; it does not authorize another funded replay. Without authorization, preserve the draft and report the blocked publication step without claiming full live completion. After authorized publication, independently reopen budget/cap/schedule and record actual object IDs/status. Distinguish publication, scheduled, processing/review, approved and delivering. Do not alter existing campaigns, account billing/limits, production tracking/site code or send outreach.\n11. Package original outputs, native links/exports, copy review/freeze, audit, experiment assumptions, configuration/state differences, prompts, model/harness contributions, timestamped screenshots, failures and assistance. Screenshots/logs are the chosen record; continuous video is not required by this task. Preserve original evidence and redact secrets/unrelated account data.\n\nCompletion\nJudge the source-supported creative set, correct saved configuration and evidence, not exact reference wording or guaranteed leads/ROAS/funding. A valid read-only tracking audit can report gaps. An unsupported preview is not a pass. Live completion additionally requires actual authorized publication and readback. The attached reference is assisted and includes explicitly documented QA/measurement limits; it is not a new independent difficulty result.","software_licensing":"Commercial","evaluation_criteria":null,"verification_method":"Version 2.1 evaluation — inspect genuine native outputs and persisted Meta state. Creative solutions may differ. Use numeric_metrics_v2_1.json, objective_qa_v2_1.csv and Raisi_ALE_V2_1_Reference_Evidence_Repaired.zip. Hashes verify reference-file integrity, not the only acceptable creative answer. These acceptance criteria are task-authored, not ALE's official scoring system.\n\n1. Research: dated official-source claims with boundaries; explicit founder ICP; two direct competitors and one substitute; three positioning options and five distinct supported hypotheses. Trace every material advertising claim. Flag invented proof, unsupported exclusivity and incorrect software/service or CTA expectations.\n2. Copy: exactly 3 Primaries, 3 paragraphs each, 80-150 words, <=1,200 characters and final CTA; 5 Meta headlines <=40 characters, 2 direct CTAs/3 product-curiosity variants; 5 on-image headline/support pairs <=60/100 characters and <=3 headline lines. Inspect critique/revision/freeze sequence and all 15 allowed Primary/concept combinations. Count and inspect meaning; do not reward merely generic but length-compliant text.\n3. Canva: open all 15 active native pages across the three ratios and verify editable headline/support text, intact official logo and at least two meaningful layouts. Inspect all 15 native PNGs and PDF proofs covering the pages. Verify dimensions, opacity, actual export provenance, exact frozen text and legibility. Compare original/R2 manifests; historical tall pages are not active assets. Record color-profile limits honestly.\n4. Integration and visual QA: one campaign, one ad set, five concept ads. Each C_i uses H_i, P1-P3 and its correct square/portrait/tall assets. Reopen both global and per-media overrides; a global headline does not prove a media override is correct. Record adaptive groupings and crop settings. Inspect real representative Feed/vertical/compact previews per concept where supported. A blank or unrendered preview remains unverified, not passed. No equal-delivery or deterministic-placement assertion without evidence.\n5. Measurement: verify read-only evidence for the intended event/dataset, attribution and actual CTA route. Separate event availability, recent receipt, aggregate deduplication and freshness from controlled firing, consent, qualified-meeting semantics and end-to-end attribution. Require explicit gaps and actionable next checks; correct diagnosis receives credit. Reject fabricated leads/events or misattributed historical totals. Verify unique concept UTMs and distinguish visible query loss from proven attribution loss.\n6. Experiment: check formulas and budget arithmetic in the sensitivity model; assumptions must be hypothetical or sourced. Require meaningful unit/outcome/denominator, uneven-allocation and translation/placement confounders, minimum evidence and an inconclusive rule. Do not accept a causal winner from a small adaptive screening campaign or planned numbers.\n7. Saved state and publication: compare intended state with independently reopened values: USD 50/day TOTAL, separate USD 350 cap, future 168-hour schedule/offsets, U.S. control, identity, homepage, CTA, objective/event/dataset, attribution, translations, eligible placements, 15 image files and frozen English options. Document corrections, including end dates lost on budget-mode change. Confirm new objects only, valid per-instance authorization and actual publication/readback. Do not infer delivery/approval from Scheduled, Processing or enabled switches. If provisioning/authorization is missing, record environment invalidity and incomplete live endpoint, not model inability.\n8. Evidence: require original exports, native access, prompts, timestamps, model/harness/operator contributions, coverage/mismatch records and screenshots. Keep credentials and unrelated account data out. Report measured completion separately from an independent trial score.\n\nActual current reference\nMeta's receipt at 2026-10-08 03:32:43 UTC (October 7 Eastern) confirmed 1 campaign, 1 ad set, 5 ads published. Campaign 120253865661450729; ad set 120253865661460729. The five published rows initially showed Processing; campaign/ad set were Scheduled. Reopened published budget: USD 50/day plus separate USD 350 campaign cap. Reopened schedule: October 8 2026 09:00 to October 15 2026 09:00 Eastern, 168 hours. No approval, delivery or performance is claimed. See PUBLICATION_REPORT.md, outputs/meta/publication_receipt.json and evidence 61-63 in the ZIP.\n\nThe active inventory is five concepts x three formats: 15 PNGs, 15 editable native pages and three five-page native PDFs. There are 3 Primary texts and 5 Meta headlines; the three long descriptions are Primary fields, not short Description fields. Exact global copy readback passed 25 field comparisons across 5 ads. Each ad's 3 correct active files and unique UTM were reopened. R2 tall pages preserve the same wording/logo inside safer content bounds after an observed Feed crop problem. Per-media inherited-headline and save-persistence errors were repaired and recorded.\n\nReference limits\nNative renders were reviewed, but the Meta placement-preview matrix is partial; blank/lazy-loaded cards were not counted as passes. Independent colorimetric sRGB certification is not claimed. The measurement audit observed historical aggregate events/deduplication/freshness, not controlled firing/consent/qualified-meeting semantics or end-to-end attribution. The published adaptive-media interface differs from a deterministic placement lock. The source package preserves intermediate findings; PUBLICATION_REPORT.md supersedes its old pending-budget/unpublished statuses. Native links retain existing permissions and require evaluator access. This is a genuine executed reference with disclosed gaps, not a claim that every quality gate passed.\n\nDifficulty status\nNo independent self-test has been completed for this version. The reference used ChatGPT Work GPT-6 Astra/Extra High for research, copy, Canva and measurement planning, with Codex operating the authenticated local browser and supplying observations/corrections. It cannot be presented as a blinded model attempt or assigned a lower difficulty percentage. The earlier five-square-image pilot had 67/100 observed artifact completion, was access-limited and exposed to prior Raisi context; it belongs only to that earlier scope. SELF_TEST_STATUS_v2_1.md preserves this history. The optional score is therefore left blank for v2.1; no official last-exam/near-term classification or days-long expert execution claim is made.\n\nMeasurement evidence repair — 8 October 2026\nAll five original measurement UI text captures cited by outputs/measurement/observed_audit.json have been restored at the exact archive-root paths. All six audit evidence references now resolve. MEASUREMENT_EVIDENCE_REPAIR.md, evidence/measurement_capture_provenance.json and outputs/measurement/claim_evidence_map.csv document source-file creation/modification timestamps, original and packaged hashes, disclosed privacy redactions and claim-to-source line references. File timestamps are local capture-save provenance, not independently certified capture times. The original audit report timestamp and all historical metric values remain unchanged. No fresh conversion test or funded replay occurred; prior measurement and preview limits remain in force.\n","software_and_version":"ChatGPT Work web (observed GPT-6 Astra / Extra High) with Canva connector; Canva web; Meta Ads Manager and Events Manager web; local Google Chrome on macOS; Codex local-browser operator. Reference captured October 7-8, 2026 UTC. Exact app builds and operator model were not independently recorded. No cloud browser was used for the new Meta launch.","specific_requirements":null,"expected_completion_time":null},"files":[{"file_id":"3c133c56-9ee1-4b4a-846d-372ed7639c01","filename":"website.txt","category":"input_materials","content_type":"text/plain","size_bytes":22},{"file_id":"be360081-90f1-4487-8b49-7fbc9b15b986","filename":"raisi-logo.png","category":"input_materials","content_type":"image/png","size_bytes":44302},{"file_id":"4f44551b-ca4d-46c8-a217-53f2f09e493d","filename":"INPUT_README_v2_1.md","category":"input_materials","content_type":"text/markdown","size_bytes":972},{"file_id":"e1b8ef3f-3fb5-4202-8e17-dffad0df515e","filename":"execution_context_v2_1.json","category":"input_materials","content_type":"application/json","size_bytes":1916},{"file_id":"2560521b-8935-43e3-8945-53471638b55f","filename":"SELF_TEST_STATUS_v2_1.md","category":"reference_output","content_type":"text/markdown","size_bytes":1497},{"file_id":"043b624c-f319-485d-bf5a-86cef3e3f26c","filename":"MEASUREMENT_EVIDENCE_REPAIR.md","category":"reference_output","content_type":"text/markdown","size_bytes":3662},{"file_id":"2038f9a6-9003-4994-b410-b45320ccd889","filename":"objective_qa_v2_1.csv","category":"reference_output","content_type":"text/csv","size_bytes":2786},{"file_id":"f20f9352-20b1-4c63-94fe-81e85fd9203d","filename":"numeric_metrics_v2_1.json","category":"reference_output","content_type":"application/json","size_bytes":2812},{"file_id":"afbdc957-4a12-4e58-addd-d0d7015e8262","filename":"SUBMISSION_README_v2_1.md","category":"reference_output","content_type":"text/markdown","size_bytes":2669},{"file_id":"30bfa760-a250-4243-8d9c-2272be8af7e1","filename":"Raisi_ALE_V2_1_Reference_Evidence_Repaired.zip","category":"reference_output","content_type":"application/zip","size_bytes":7029355}],"deep_review":{"overall_verdict":"Accept","review_text":"**Overall Verdict:** Accept — Revision 3 provides a coherent professional task and a readable, internally consistent assisted reference, with measurement repairs verified and remaining limits disclosed.\n\n**Completeness Check:**\n\n- Instruction: Complete — Defines the research, creative production, measurement audit, experiment, configuration, publication, and evidence requirements, including authorization boundaries.\n- Inputs: Complete — Website discovery input, official logo, input guidance, and operating constraints are supplied; reference answers are explicitly excluded from independent participant inputs.\n- Toolchain: Complete — Commercial tools, authenticated environment requirements, capability preflight, and fresh publication authorization are specified.\n- Reference output: Complete — Includes research, copy review/freeze, creative exports, native editability receipts, measurement evidence, experiment calculations, saved-state records, and publication evidence. It is explicitly a partial-verification assisted reference.\n- Rubric: Complete — Provides inspectable criteria, quantitative constraints, evidence locators, and distinctions between publication, delivery, measured results, and independent difficulty.\n\n**File and Task Reachability:**\n\nAttachment classification: `reachable`. All ten manifest attachments were downloaded and inspected: website.txt, raisi-logo.png, INPUT_README_v2_1.md, execution_context_v2_1.json, SELF_TEST_STATUS_v2_1.md, MEASUREMENT_EVIDENCE_REPAIR.md, objective_qa_v2_1.csv, numeric_metrics_v2_1.json, SUBMISSION_README_v2_1.md, and Raisi_ALE_V2_1_Reference_Evidence_Repaired.zip. Their byte counts match the manifest. The archive contains 281 readable members, passes container integrity checks, and matches all 280 declared member hashes.\n\nInspection confirmed 15 correctly sized opaque PNGs, three five-page Canva PDF proofs, editable headline/support objects documented across all 15 active pages, compliant Primary/headline counts, and exact copy in 25 recorded field comparisons. All six measurement audit evidence paths resolve; all eight claim locators match the restored captures. The sensitivity calculations reproduce correctly.\n\nTask classification: `reachable` in the specified provisioned environment. The workflow has a viable production and evaluation path through native creative assets, persisted campaign settings, and timestamped evidence. Each new live instance requires its own authorization; the reference authorization cannot be reused.\n\nLive Canva permissions, current Meta state, and external source pages were not independently accessed. Full webpage bodies and historical local-only screenshots remain unreviewed. Packaged scripts were not executed. These limits do not invalidate the supplied evidence. Publication screenshots substantiate the recorded daily budget, separate cap, schedule, and Scheduled/Processing statuses; they establish neither approval nor delivery. Placement coverage, controlled conversion testing, consent, qualified-meeting semantics, attribution, and independent difficulty remain explicitly unverified.\n\nLicense keys, accounts, tokens, and other credentials are not required from the contributor. Confidential, proprietary, or internal material may be evaluated as submitted evidence without requiring additional disclosure.\n\n**Special Screening:**\n\nNone. No affirmative evidence supports a screening label. Public benchmark provenance was not independently checked because external search was outside the permitted evidence boundary.\n\n**Recommended Changes:**\n\nThe restored captures materially resolve the prior measurement evidence gap. The submission also preserves genuine corrections and clearly separates assisted execution from independent scoring.\n\nNo contributor changes are required. Evaluation should retain the disclosed limits and award completion only for supported artifacts and recorded state; unavailable previews and untested measurement behavior must remain unverified.","reviewed_at":"2026-10-08T13:54:10.046825Z"},"claim":null}