AfterQuery - the untracked firm named in Sunday's McCaul report - reportedly raised at a $3.2B valuation, Y Combinator's fastest launch-to-unicorn ever, 10x its April mark. It named Nvidia, Legora, and Korean lab Motif as customers and claimed a $100M annualized run rate. The expert-data market's newest giant also carries a congressional China-exposure claim.
Mercor open-sourced its post-training recipe: a 397B-parameter RL run on public SkyRL with ~1,928 expert-created tasks, full training script, weights, and eval traces released - beating Opus 4.5 on its own APEX-Agents benchmark. Monday's evals-battleground warning just escalated.
Anthropic released Fable/Mythos 5.1 with zero data retention for eligible enterprise customers ahead of customer-controlled "Enterprise Frontier Safeguards" this fall. Data-handling guarantees are now a headline product feature - the procurement language our clean-provenance positioning speaks.
Meta Superintelligence Labs shipped its first real-time audio model, Muse Voice Transcribe (streaming ASR, 20+ speaker diarization, multilingual code-switching). Meta is building speech in-house - watch voice-data demand at this P0 account.
Government lane: Enabled Intelligence and Seekr are expanding AI data labeling for NGA and defense programs - cleared, specialized competitors are forming inside our Gov lane.
Do today: (1) upgrade AfterQuery to Tier 1 and brief GTM, (2) fold its $3.2B + McCaul naming into the clean-supply-chain one-pager, (3) run the Mercor APEX/SkyRL teardown (weights and traces are public), (4) watch whether zero-retention resets customer expectations for data handling.
The day’s material announcements, translated into the market shift and decision they create for AuraOne.
AfterQuery reportedly raised at a $3.2B valuation - 10x its April mark in five months - naming Nvidia, Legora, and Motif as customers. The model is AuraOne's exact positioning, and the company carries a McCaul China-exposure claim.
A 397B-parameter RL run on ~1,928 expert-created tasks, with the full script, weights, and eval traces released, topped its own APEX-Agents benchmark. The benchmark battleground flagged Monday just escalated.
Fable/Mythos 5.1 pairs zero retention with customer-controlled "Enterprise Frontier Safeguards." Expect procurement to start asking data vendors the same data-handling questions.
Muse Voice Transcribe brings streaming ASR, 20+ speaker diarization, and multilingual code-switching. A P0 account is building speech in-house - demand shifts toward edge-case data and evaluation.
Enabled Intelligence and Seekr are expanding AI data labeling for NGA and defense programs - company announcement dated Aug 27 (out of window, corrected after audit); retained as context on cleared competitors in the Gov lane.
Datatang's Nexdata multimodal ego-centric dataset (event Aug 26, carried as context) bundles video, IMU, SLAM, depth, and hand keypoints, backed by a 300-robot facility claiming ~5,000 hours per month.
TechCrunch (Sept 1): AfterQuery reportedly raised at a $3.2B valuation - 10x its $300M Series A mark from April, five months ago, and per YC partner Gustaf Alstromer the fastest launch-to-unicorn in YC history. Founders are 22 and 23 (YC W25). April claims: $100M annualized run rate; customers incl. Nvidia, Legora, Korean lab Motif. Model: doctors, lawyers, and specialists training models to work like professionals - "encoding the patterns, decisions, and reasoning of the world's best practitioners." First funding record in our system: flagged Monday only as an untracked McCaul-named firm; 48 hours later it is a reported $3.2B company.
The pitch is a direct competitor for expert-sourced post-training data, and $3.2B buys recruiting, brand, and price pressure. But it is also one of the four firms McCaul named as supplying Chinese AI labs while serving US/Pentagon clients - every dollar of that valuation sharpens our clean-supply-chain contrast. Its customer list is a targeting map of who buys expert data at scale.
Research: upgrade AfterQuery to Tier 1 today with full profile (done in this run). CEO/GTM: update the clean-supply-chain one-pager with the valuation; map our targets against its named customers. This week.
Mercor published a step-by-step 397B-parameter RL post-training guide (primary source, mercor.com): ~1,928 expert-created tasks post-trained Qwen3.6-35B past Opus 4.5 on its APEX-Agents benchmark and lifted 397B Pass@1 by 70% relative (16.11% to 27.29%). Full training script, model weights, and eval traces released; task format documented for proprietary data. First recorded observation of the recipe going public - the APEX Benchmarks hiring flagged Monday is visibly part of a deliberate strategy: own the benchmark, the recipe, and the data.
Mercor is converting its expert-data supply into public technical authority - the thing buyers read when choosing post-training partners. If the market learns post-training from Mercor's playbooks, Mercor defines and prices "expert data." Counter-narrative available: the recipe runs on ~2,000 expert tasks - quality and coverage are the lever, not scale.
Product/Research: execute the carried teardown with teeth - review the SkyRL recipe against our task taxonomy; decide this week whether AuraOne publishes a benchmark, a post-training-operations service, or both. Teardown by Friday.
Sept 1 release: cheaper tokens, fewer false-positive refusals, record Terminal-Bench 4.0 and Humanity's Last Exam scores - and zero data retention for eligible enterprise customers, with "Enterprise Frontier Safeguards" (customer infrastructure, customer-controlled monitoring) rolling out this fall. Anthropic: "never trained on enterprise data without explicit permission, and never will." System card candidly flags a slight misalignment regression vs Opus 5. First recorded observation: retention and customer-controlled deployment moved from contract language to headline feature at a frontier lab.
Zero-retention and on-infrastructure deployment reset buyer expectations about data flows everywhere - including data vendors. Procurement will start asking vendors the same questions: where does our data go, who touches it, what is retained. Our consent-documented, US-clean supply chain answers that questionnaire best. The misalignment candor keeps the independent-evaluation opening warm.
GTM: add a data-handling section (retention, residency, consent records) to enterprise/gov collateral this week. Research: track the Safeguards fall rollout as an eval-on-customer-infrastructure trigger.
Meta AI Research introduced Muse Voice Transcribe, "the first real-time audio perception model developed by Meta Superintelligence Labs" (primary source, research.meta.ai): real-time streaming ASR, diarization across 20+ speakers, endpointing, multilingual with seamless code-switching, and biasing by language/keyword/context. Coverage highlights five Indian languages and a Mac dictation surface. First recorded observation of MSL shipping an audio model - Meta's voice capability is now productized, not just research.
Threat edge: a P0 account building speech in-house can internalize data needs we might serve. Opportunity edge: real-time multilingual code-switched ASR is exactly where off-the-shelf speech data fails - accent coverage, domain audio, diarization edge cases, and evaluation sets become recurring purchases as MSL pushes language coverage. Voice is heating from both the model and application ends.
Research: track Muse language-coverage expansion as a demand calendar for multilingual speech data and evaluation. GTM: add the MSL speech org to the Meta account map alongside the robotics contacts. This week.
LlamaIndex and Kaggle publicly launched ExtractBench, a document-extraction benchmark and leaderboard for AI agents (Sept 2, llamaindex.ai primary source). Installed 2026-09-03 after independent audit: replaces the original item 5 (Enabled Intelligence + Seekr), whose company announcement was dated Aug 27 - outside the coverage window; the Aug 31 article was follow-on coverage. That item is retained as Gov-lane context.
Low-end extraction evals are commoditizing in public; enterprise-grade evaluation stays open. A public leaderboard is also a cheap public artifact - the same play Mercor just ran with APEX.
Product/Research: decide whether AuraOne participates in ExtractBench (submission = public artifact in the eval conversation). This week.
Anthropic: Fable/Mythos 5.1 with zero data retention (Top 5, item 3). The Register reports Anthropic is asking partners to help fund safety work, and TNW confirms external cyber testing has resumed under the new hardening - both consistent with Monday's alignment post. Buyer per SSOT: Brett Wittmershaus (existing thread). Watch: Enterprise Frontier Safeguards fall rollout.
Meta Superintelligence Labs: Muse Voice Transcribe (item 4) plus the data-center robot pilots (Sept 1 report, item 3; confidence Medium pending Meta/vendor confirmation per audit). Two physical-world build-outs at one P0 account in one week.
Mercor (frontier-adjacent): 397B RL recipe release (item 2) - its models took top corporate-law and fourth overall on APEX-Agents; its expert-data positioning now has public receipts.
AfterQuery: customer names on record (Nvidia, Legora, Motif) - overlay against our target list.
Per the customer exclusion rule, OpenAI (canceled pilot) and xAI (excluded) are not covered as customer news.
China AI data services market sized at RMB 7.83B for 2026 on agentic AI demand (InfotechLead, Aug 31) - the China-side data market is scaling in parallel; watch for cross-border procurement friction. Confidence: Medium.
Wetour Robotics demonstrated an sEMG-vision system targeting force/occlusion blind spots in Physical AI training (Sept 1) - new capture modality; watch for dataset productization. Confidence: Medium.
LlamaIndex + Kaggle launched a document-extraction leaderboard for AI agents (Sept 2) - evals infrastructure keeps commoditizing at the low end. Confidence: High.
TechPolicy Press (Sept 1) on MTurk's displaced workers - the labor-force narrative around the Sept 30 shutdown is building; expect press to name migration winners. Confidence: High.
Anthropic partner-safety-funding ask (The Register, Sept 1) - if frontier labs start charging partners for safety programs, eval budgets formalize further. Confidence: Medium-High.
Context resurfaced this window (not new): Datatang Nexdata ego-centric dataset (event Aug 26), Skild S1 (Aug 18) and HappyRobot Series C (Aug 4) syndications, and McCaul re-waves - all excluded from the Top 5 per the recency rule.
MTurk migration: closure Sept 30 unchanged; the migration offer remains the P0.
Evidence-backed revenue candidates discovered in today’s intelligence. Every candidate must be deduplicated against the Revenue & GTM workbook before it is created or routed.
P0 - Do now
P1 - This week
P2 - Monitor
1. The expert-data labor model is now hyper-funded (AfterQuery's 10x in five months). Capital alone does not buy provenance; our differentiation must be documented trust, not just expertise.
2. Post-training methodology is being open-sourced by competitors (Mercor/SkyRL). If buyers learn the craft from rivals, we must publish or be defined. The decision on a public benchmark/ops artifact cannot slip another week.
3. Data handling is becoming a product feature (Anthropic zero retention). Vendors will face the same scrutiny; our consent-documented supply chain should be sold as a feature, not disclosed as a footnote.
4. The Gov lane is professionalizing around us (Enabled+Seekr for NGA; Innodata on SHIELD). Pick the channel strategy now: direct, prime, or partner.
5. Voice is heating at both ends - Meta building models in-house, HappyRobot-class production fleets above. The durable contract is multilingual evaluation and edge-case coverage, not raw audio.