TUESDAY 25 AUGUST 2026 · 7 CLAIMS CHECKED · 0 SURVIVED THE RECEIPTS · ISSUE 16 OF 17
BUSINESS THE INFORMATION (ANONYMOUS SOURCING), SYNDICATED VIA REUTERS AND RE-REPORTED WORLDWIDE · CLAIMED 2026-08-2301/7
Nvidia may buy into the company that turns Nvidia's chips into $750 million of "revenue."
The buyer, the seller, and the compute are all the same trade, three days before Nvidia tells Wall Street how the quarter went.
01THE CLAIM
"Nvidia is in talks to invest in Perplexity as part of an equity funding round that would value the AI startup at more than $30 billion, with Perplexity's annualized revenue reported to have risen to more than $750 million from less than $250 million at the start of the year." [SOURCE ↗]
-4.8%Nvidia stock move over six sessions into its Aug 26 earnings print
02THE CHECK
THE PITCH. Nvidia is in talks to invest in Perplexity at a $30 billion-plus valuation, up 50% from last year's $20 billion, on annualized revenue that jumped from under $250 million to over $750 million in eight months.
THE CATCH. "Annualized" means one good month times twelve, never audited, and Perplexity has never published a real annual figure. Neither company confirmed the talks; the story runs on one anonymous source, everywhere, at once.
THE NUMBER THAT EXPLAINS EVERYTHING. 40. That is the valuation-to-revenue multiple, on revenue nobody outside Perplexity has shown a receipt for.
WHAT NOBODY SAYS OUT LOUD. the company floated to buy the equity is the same company selling the GPUs that revenue gets spent on. Central bankers have a name for that pattern and it is not a compliment.
03SAY THIS IN THE MEETING · 📸 SCREENSHOT IT
""Show me the audited annual, not the annualized run-rate, and tell me who's writing the check.""
If a vendor's own stock is propping up its customer's valuation, that "traction" is not a signal worth copying into your own pitch deck.
05🔮 OUR CALL · ON THE RECORD 2026-08-25
No signed deal announced by the Aug 26 earnings call. If one lands within 90 days, the number comes in closer to $25B than $30B, with a compute-purchase commitment attached. Hold us to it.
Flips toward "real" if Perplexity discloses audited revenue north of $500M independent of this round. Flips toward "worse" if the deal includes an undisclosed chip-purchase commitment.
RECEIPTS (6) · CONFIDENCE MEDIUM
every URL below answered a live HTTP check before publish · sweep 2026-08-28
▲benzinga.com⧉ · "Nvidia is considering participating in Perplexity's latest equity financing round, which could value the company at more than $30 billion"
●benzinga.com⧉ · "Nvidia and Perplexity did not immediately respond to Benzinga's request for comment."
●tradingview.com⧉ · "NVIDIA is effectively bailing out anyone in the AI industry as a means of inflating their valuations and keeping them buying compute"
●bis.org⧉ · "Chip makers and hyperscalers take equity stakes in AI labs or neocloud providers, who in turn commit to multi-year purchases of chips or computing power."
●en.cryptonomist.ch⧉ · "Nvidia first put money into the startup back in late 2023, well before generative search became a crowded and increasingly competitive category."
●pymnts.com⧉ · "The new funding values the company at $20 billion, according to multiple media accounts late Wednesday."
BUSINESS ASTROMECH (BEN LAMM AND GEORGE CHURCH, CO-FOUNDERS; ROUND LED BY BOB NELSEN) · CLAIMED 2026-08-2002/7
Twenty million dollars just told the world Astromech is worth $3.8 billion. Nobody checked whether the customers exist, because there aren't any yet.
A biology "operating system" with zero disclosed revenue just repriced into unicorn territory on a demo, a co-founder's name, and an unpublished benchmark.
01THE CLAIM
"Astromech raised $20 million at a $3.8 billion valuation, a roughly 90% markup in five months, and was reported across a dozen outlets as a $3.8B AI company while remaining pre-revenue with no announced customers and no prospectively tested forecast." [SOURCE ↗]
$20Mnew round raised - 0.53% of the headline valuation
$60Mtotal capital raised, lifetime
~90%valuation markup in ~5 months (from ~$2B in March 2026)
$0 / 0disclosed revenue / announced customers
20xcapital Chai Discovery raised for the same $3.8B sticker price
02THE CHECK
THE PITCH. Astromech, co-founded by Colossal Biosciences' Ben Lamm and geneticist George Church, raised $20 million at a $3.8 billion valuation, roughly a 90% markup on where it stood five months ago.
THE CATCH. Pre-revenue. No announced customers, no named pilot, no published forecast tested against real outcomes. The headline "100x speed" figure is Astromech's own internal number, never released for outside review.
THE NUMBER THAT EXPLAINS EVERYTHING. $20 million. That is the entire size of the check that set a $3.8 billion price tag, 0.5% of the number every outlet ran with.
WHAT NOBODY SAYS OUT LOUD. Chai Discovery bought the identical $3.8 billion sticker price in July with twenty times the capital. Same number, different company; that is a market pricing narrative, not product.
03SAY THIS IN THE MEETING · 📸 SCREENSHOT IT
""What's the check-to-valuation ratio, and who else paid the same sticker price this year?""
A "$3.8 billion" headline means nothing without the size of the check that produced it. Ask what percentage of the company actually changed hands before you update your read on who's winning.
05🔮 OUR CALL · ON THE RECORD 2026-08-25
No named customer or published forecast validation by end of Q1 2027. The $3.8B marker becomes the new reference point once the next round prices flat or lower. Hold us to it.
Flips toward "earned" if Astromech publishes a prospectively-tested forecast beating a stated baseline, or names a paying customer. Flips toward "worse" if the next round prices below $3.8B.
RECEIPTS (6) · CONFIDENCE HIGH
every URL below answered a live HTTP check before publish · sweep 2026-08-28
▲siliconangle.com⧉ · "We are building an algorithmic prediction solution"
▼thenextweb.com⧉ · "No forecast has been tested prospectively. Everything published so far looks backwards."
▼thenextweb.com⧉ · "No partner pilot has been named, and no customer has been announced in health, biosecurity, agriculture or conservation."
●thenextweb.com⧉ · "TNW reported in July that Chai Discovery raised $400mn at the same $3.8bn valuation, on a different problem and twenty times the capital."
●dealroom.co⧉ · "Biology runs the world and historically, we have only reacted to it"
▲pulse2.com⧉ · "The financing brings Astromech's total capital raised to $60 million."
OTHER ERIK BRYNJOLFSSON, BHARAT CHANDAR, RUYU CHEN - STANFORD DIGITAL ECONOMY LAB, 'CANARIES IN THE COAL MINE?' (AUGUST 2026 REVISION) · CLAIMED 2026-08-1203/7
The paper everyone cites to prove AI is destroying entry-level jobs opens by saying it found no such thing.
A 19% headline number and a "no economy-wide displacement" finding, both in the same study, both getting quoted by opposite sides.
01THE CLAIM
"Employment among workers ages 22-25 in highly AI-exposed occupations now stands about 19% below where it would be if it had kept pace with employment among similarly aged workers in less-exposed occupations - cited widely as proof AI has already destroyed a fifth of entry-level jobs." [SOURCE ↗]
19%counterfactual employment gap, ages 22-25 in AI-exposed occupations vs less-exposed peers
15%the counterfactual gap in July 2025, before widening to 19%
13%the gap as originally published in August 2025 - the number has grown twice
Mar-Apr 2022when AI-exposed job postings actually peaked and began declining, per EIG analysis
40 yearshow sharp the Fed's rate-hike cycle was that began the same month, per EIG
02THE CHECK
THE PITCH. A Stanford Digital Economy Lab paper finds employment among 22-25 year-olds in highly AI-exposed jobs now runs about 19% below where it would sit had it tracked less-exposed peers, up from 15% a year ago and 13% at first publication, cited widely as proof AI has already gutted a fifth of entry-level jobs.
THE CATCH. The same paper's first listed finding: "We do not see widespread, economy-wide job displacement associated with AI." The 19% gap comes from slower hiring, not layoffs, measured against a comparison group, not the whole economy.
THE NUMBER THAT EXPLAINS EVERYTHING. March-April 2022. That is when AI-exposed job postings actually peaked and started falling, per a separate analysis, seven months before ChatGPT existed, and precisely when the Fed began its sharpest rate-hike cycle in 40 years.
WHAT NOBODY SAYS OUT LOUD. a growing headline number from a paper whose lead author is now publicly walking back the "AI apocalypse" framing is still getting quoted as the apocalypse case.
03SAY THIS IN THE MEETING · 📸 SCREENSHOT IT
""Which finding are you actually citing, the 19% gap or the 'no economy-wide displacement' line?""
A widening statistic and a softening author interview can both be true at once. Read a paper's own caveats before citing its headline number.
05🔮 OUR CALL · ON THE RECORD 2026-08-25
The 19% figure climbs again in the paper's next revision (this study has now been revised twice, each August), while its core "no economy-wide displacement" finding stays unchanged. Hold us to it.
Flips if a follow-up analysis isolates the AI-exposure effect from the 2022 rate-hike timing and the gap survives the correction. Flips the other way if the gap narrows once controlled for the broader hiring slowdown across all occupations.
RECEIPTS (7) · CONFIDENCE MEDIUM
every URL below answered a live HTTP check before publish · sweep 2026-08-28
●digitaleconomy.stanford.edu⧉ · "The adjustment appears to operate primarily through reduced hiring of young workers rather than increased separations."
▼agglomerations.eig.org⧉ · "vacancies for the highest AI exposure quintile of occupations peaked in March-April 2022 and declined sharply throughout the remainder of the year."
▼agglomerations.eig.org⧉ · "The Fed began its most aggressive cycle of interest rate hikes in forty years in March 2022, precisely when job postings in these sectors began to fall."
●brookings.edu⧉ · "Early research findings on AI's impact on the labor market are inconclusive, weak signals about the future"
●it.slashdot.org⧉ · "There is still no sign of economy-wide job destruction"
Inherent says its 27-billion-parameter model beats GPT-5.5. Its own paper says the model calls GPT-5.5 to do the work.
The "AI Scientist" that outperformed two frontier labs turns out to have one of them running inside it.
01THE CLAIM
"Faraday, a 27-billion-parameter 'AI Scientist' from Inherent Labs, outperforms Claude Opus 4.8 and GPT-5.5 on the task of replicating research papers." [SOURCE ↗]
73%share of in-distribution Replica tasks Faraday beats Opus 4.8 and GPT-5.5 on
60%win rate on out-of-distribution tasks (13-point drop off held-out data)
310total Replica benchmark tasks, drawn from 100 papers - all Inherent's own
$50Mseed round backing the launch
02THE CHECK
THE PITCH. Faraday, a 27B model from Inherent Labs, "outperforms Claude Opus 4.8 and GPT-5.5" at replicating research papers, in a launch that landed a TechCrunch feature and a live press cycle.
THE CATCH. Read the arXiv methods section and Faraday hands every coding subtask to GPT-5.5 Codex, "including at evaluation time." The 27B number describes the orchestrator, not the system doing the work being scored.
THE NUMBER THAT EXPLAINS EVERYTHING. 73% in-distribution, but only 60% out-of-distribution, a 13-point drop on held-out science tasks, on Inherent's own 310-task benchmark, judged by Inherent's own rubric, against baselines already a generation out of date by launch day.
WHAT NOBODY SAYS OUT LOUD. this is "GPT-5.5 plus a wrapper" beating "GPT-5.5 alone," scored by the company that built the wrapper.
03SAY THIS IN THE MEETING · 📸 SCREENSHOT IT
""Which parts of the pipeline are your 27B model, and which parts are GPT-5.5 doing the heavy lifting?""
When a benchmark result includes a competitor's model as a component, the fair comparison is wrapper-vs-no-wrapper, not vendor-vs-vendor. Ask what's inside before you believe what beat what.
05🔮 OUR CALL · ON THE RECORD 2026-08-25
No independent third-party rerun of Replica within 60 days. If one happens, Faraday's edge over bare GPT-5.5 with a simple harness narrows to single digits. Hold us to it.
Flips toward Inherent if an independent lab reruns Replica (or a public benchmark like MLE-Bench) and confirms a comparable gap using open baselines. Flips toward "just a wrapper" if removing GPT-5.5 Codex collapses Faraday's score.
RECEIPTS (6) · CONFIDENCE HIGH
every URL below answered a live HTTP check before publish · sweep 2026-08-28
▼arxiv.org⧉ · "Faraday is provided with a frontier coding agent to use as a tool. A wrapper script runs the Codex CLI non-interactively."
▼arxiv.org⧉ · "GPT-5.5 in the final stage and for evaluation"
▼aiweekly.co⧉ · "Faraday is built on a 27-billion-parameter Qwen base and calls OpenAI's GPT-5.5 Codex for coding subtasks."
●superpowerdaily.com⧉ · "The disclosed evaluation includes no numerical scores, named test papers or methodology, limiting what can be concluded from the comparison."
▲techcrunch.com⧉ · "What was most interesting to us about this was not so much the result of beating those frontier agents"
▲inherentlabs.ai⧉ · "a 27B-parameter "AI scientist" agent that outperforms Claude Opus 4.8 and GPT-5.5 on the task of replicating research"
POLICY THE JAPAN TIMES (HEADLINE, AUG 19 2026), AMPLIFIED VERBATIM BY SLASHDOT THE SAME DAY · CLAIMED 2026-08-1905/7
"Japan to Require AI Firms to Disclose Training Data" is the headline. The actual code says nobody has to, and nobody checks if they do.
A soft-law draft became a hard-law headline somewhere between the Cabinet Office and the copy desk.
01THE CLAIM
"Japan to Require AI Firms to Disclose Training Data - Japan will mandate generative-AI training-data disclosure, including for foreign firms serving the Japanese market." [SOURCE ↗]
0statutory penalties in the underlying Cabinet Office draft code
0legally binding obligations imposed by the code
0government review of filed disclosures - opt-in list only
02THE CHECK
THE PITCH. Japan will require AI companies, including foreign firms serving Japanese users, to disclose their training data, according to headlines that ran across tech press worldwide on August 19.
THE CATCH. The actual instrument is a Cabinet Office IP Strategy Headquarters draft code, built explicitly as "comply or explain": no statutory penalties, no legal binding force, and the government does not review what companies file, only publishes a list of who opted in.
THE NUMBER THAT EXPLAINS EVERYTHING. 0. That is the count of penalties, binding obligations, and government reviews in the actual code behind the "require" headline.
WHAT NOBODY SAYS OUT LOUD. Japan's comply-or-explain corporate governance codes have driven real compliance before through reputational pressure alone, so "toothless" is not the whole story either, just do not call it a requirement.
03SAY THIS IN THE MEETING · 📸 SCREENSHOT IT
""Show me the penalty clause, or admit it's a headline about a draft.""
A "mandatory disclosure" headline about a soft-law code is a preview of a fight, not a settled rule. Do not build a compliance plan around a draft with no enforcement mechanism yet.
05🔮 OUR CALL · ON THE RECORD 2026-08-25
The code finalizes roughly as drafted, comply-or-explain, by autumn 2026, with no penalty clause added. Hold us to it.
Flips if the final code adds statutory penalties or mandatory government review before the autumn 2026 target, or if Japan attaches binding force through separate legislation.
RECEIPTS (5) · CONFIDENCE HIGH
every URL below answered a live HTTP check before publish · sweep 2026-08-28
●yro.slashdot.org⧉ · "Japan is preparing a nonbinding "comply or explain" code that would urge generative AI companies"
▼mlex.com⧉ · "The non-binding code, discussed by an Intellectual Property Strategy Headquarters study group on intellectual-property rights in the AI era"
▼connectontech.bakermckenzie.com⧉ · "The Principle Code is explicitly framed as soft law. It does not impose legally binding obligations or statutory penalties."
▼nippon.com⧉ · "the government will use a "comply or explain" approach, under which it will set out a nonbinding code for generative AI businesses, including system developers and service providers, allowing them to choose either to comply with the code or publicly explain why they will not comply"
●resultsense.com⧉ · "It is non-binding and runs on comply-or-explain. Foreign providers offering AI services in Japan are covered too."
Nvidia's agent just went 100% on a benchmark built to resist that. The brain doing the reasoning is Anthropic's, and it scores 30% alone.
A perfect score on the public half of a test that has never been beaten on the half that counts.
01THE CLAIM
"NVIDIA AVO achieved a 100.00 RHAE score across all 25 environments in the ARC-AGI-3 public set, completing all 183 levels, demonstrating a frontier-level general-purpose architecture for long-horizon autonomous agents." [SOURCE ↗]
THE PITCH. NVIDIA AVO hit a perfect 100.00 RHAE score across all 25 environments and 183 levels of the ARC-AGI-3 public benchmark, pitched by Nvidia as proof of "a frontier-level general-purpose architecture."
THE CATCH. Nvidia's own post says the private and semi-private sets, the parts of ARC-AGI-3 nobody can rehearse against, remain unsolved by AVO and every other system. The reasoning inside AVO is Anthropic's Claude Opus 5, which scores 30.2% completely alone.
THE NUMBER THAT EXPLAINS EVERYTHING. 100.00 on the set you can practice against, 0 systems have ever cracked the set you can't.
WHAT NOBODY SAYS OUT LOUD. Nvidia's own post admits its comparison to the prior leaderboard entry "should not be interpreted as a controlled ablation." The architecture claim rests on a model Nvidia did not train.
03SAY THIS IN THE MEETING · 📸 SCREENSHOT IT
""Show me the private-set score, or tell me why there isn't one yet.""
"Frontier architecture" claims that credit a harness, not the underlying model, should be read as harness engineering, real work, but a different claim than general intelligence.
05🔮 OUR CALL · ON THE RECORD 2026-08-25
No system, including AVO, clears the ARC-AGI-3 private set before 2027. Hold us to it.
Flips if AVO or any system posts a verified private-set score above 50%, or if Nvidia publishes a true controlled ablation isolating the harness from the underlying model's contribution.
RECEIPTS (5) · CONFIDENCE HIGH
every URL below answered a live HTTP check before publish · sweep 2026-08-28
▼developer.nvidia.com⧉ · "They are not results on the semi-private or fully private competition sets."
▼developer.nvidia.com⧉ · "This should not be interpreted as a controlled ablation: the two systems differ in agent backend, observation representation, memory, context management"
▼eneralabs.com⧉ · "The private-set benchmark, which withholds environments not available in the public set, remains unsolved."
▼cryptobriefing.com⧉ · "Without private set results, the 100% score is impressive but incomplete as a measure of general reasoning ability."
●cryptobriefing.com⧉ · "AVO exists as a research demonstration for now, not a product you can buy or integrate."
OpenAI moved its own biorisk red line by 20 points and called the model that cleared it safe.
A safety threshold that says 50% one quarter and 30% the next is not a stricter bar, it is a bar that stopped holding still.
01THE CLAIM
"OpenAI's GPT-5.6 System Card states GPT-5.6 Sol scores below its indicative biorisk threshold for protein-binding capability, using '30% as an indicative threshold, based on a survey of 20 independent experts.'" [SOURCE ↗]
0times the word 'survey' appears in the April GPT-5.5 card, despite both new thresholds being credited to a 20-expert survey
118 daysa pass@1 value sat mislabeled as pass@4 across two published system cards before an Aug 19 correction
3.7xsize of the Aug 19 correction (0.4% to 1.48%), which shrank the apparent generational jump from 19x to 5.1x
02THE CHECK
THE PITCH. OpenAI's GPT-5.6 System Card says the model "scores below" its indicative biorisk threshold for protein-binding capability, now set at 30%, "based on a survey of 20 independent experts."
THE CATCH. The prior card, published 11 weeks earlier, set that same threshold at 50%. The word "survey" appears zero times in that April document, despite both thresholds now credited to the same 20-expert process.
THE NUMBER THAT EXPLAINS EVERYTHING. 118 days. That is how long a mislabeled score sat published across two system cards before an August 19 correction that shrank the model's apparent generational jump from 19x to 5.1x.
WHAT NOBODY SAYS OUT LOUD. the protein threshold got stricter, which cuts against a bad-faith reading, but a threshold that moves 20 points in 11 weeks on a retrofitted justification is not a fixed line, it is one drawn after the fact.
03SAY THIS IN THE MEETING · 📸 SCREENSHOT IT
""Which threshold, dated when, and what changed the number since the last card?""
A safety margin measured against a threshold that moves is not a fixed margin. When a lab reports "below threshold," ask which version of the threshold, and when it last changed.
05🔮 OUR CALL · ON THE RECORD 2026-08-25
OpenAI adjusts at least one more Preparedness Framework threshold within two system card cycles, by mid-2027, without a standalone announcement. Hold us to it.
Flips toward "normal calibration" if OpenAI publishes the raw 20-expert survey data and shows the methodology was genuinely unchanged between cards. Flips toward "worse" if a third threshold moves in the permissive direction without disclosure.
RECEIPTS (5) · CONFIDENCE HIGH
every URL below answered a live HTTP check before publish · sweep 2026-08-28
▲deploymentsafety.openai.com⧉ · "August 19, 2026: We corrected GPT-5.5's pass@4 score on the hard-negative protein binding prediction evaluation from 0.4% to 1.48%."
●securebio.substack.com⧉ · "identifies an approach that successfully evades certain screening systems, but would be inconvenient for a malicious actor to carry out in practice"
▼arxiv.org⧉ · "The 2025 OpenAI Preparedness Framework does not guarantee any AI risk mitigation practices"
▲kingy.ai⧉ · "OpenAI believes the family is capable enough in biological and chemical domains to trigger stronger safeguards."
THAT IS THE RECORD FOR ISSUE #16. NEXT VERDICT DROPS 9PM AEST.
You just read the free check. Members get the full autopsy: evidence trail, steelman, every source. 2 months free.START 2 MONTHS FREE ↗
THE AI BS REPORT
Don't take the AI industry's word for it.
“A viral claim on X said the mystery stealth model 'Ox Alpha' beats Claude Fable 5 and GPT-5.6 Sol on the DeepSWE coding benchmark, based on an 8/10-task sample (effectively >80%).”
TRUE, BUTBen Davis, @davis7 on X, amplified by AGTP (@AGTPinsights) and Coin Bureau (@coinbureau) to a much larger audience without the sample-size caveat
That's today's lead check. One a day, receipts attached, in your inbox before someone repeats it at you. Free.
No spam. Unsubscribe in one click. Receipts or it didn't happen.
