You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.

Post in Your Voice

Sunday, July 5, 2026 Intelligence for Portfolio Executives closing the AI Wage Gap.

This week Mira Murati's Thinking Machines Lab published a result with Bridgewater that should reorder how you think about every AI dollar you'll spend this year. They took an investment news-filtering task — the kind of judgment-heavy triage Bridgewater analysts do all day — and ran the frontier models at it. GPT, Claude, Gemini. The frontier average: roughly 50%. Coin-flip territory.

Then they had experts write careful prompts for those same frontier models. Mid-70s — better, still under the 80% threshold where Bridgewater's analysts would actually trust the output.

Then they fine-tuned Qwen3-235B — an open-source model, not a frontier one — on Bridgewater's own judgment, using TML's Tinker platform. 84.7% accuracy. At 13.8x lower cost.

Read the shape of that result, not just the number. The biggest, most expensive models in the world scored a coin flip on real domain work. A mid-size open model carrying one firm's encoded judgment beat them all, past the trust threshold, at a fraction of the price. Murati's framing was exact: "experts improving AI that empowers experts."

For two years the entire enterprise AI conversation has run on one assumption — when the output isn't good enough, wait for the bigger model. This week that assumption got a price tag and a scoreboard, and it lost.

———

If you've been reading this newsletter for the last month, you can feel the arc completing.

Three weeks ago, the 400,000-session study: lawyers and managers completing agentic tasks at nearly the same rate as engineers. The syntax moat drained; domain judgment became the scarce input. Last week, Karpathy's line: expressing your goals to the agent is now the bottleneck. The constraint moved from the model into your chair, and the fix was configuration — teaching the machine your context once instead of a hundred times.

This week is the third link, and it's the biggest. Your judgment doesn't just direct the model anymore. Your judgment can be the model.

That's what Bridgewater did. They didn't hire prompt engineers or wait for GPT-6. They took the thing only they have — thousands of instances of their analysts' actual calls, what mattered and what didn't and why — and pushed it into the weights of a model they control. The expertise stopped being an input you retype every morning and became an asset that runs at machine speed and machine cost.

And a second data point landed in the same seven days, from the opposite direction. A CMU paper made the rounds arguing that most "AI agents" in production are just scaffolding — orchestration code and conditional logic wearing an agent costume. The label hides enormous variation in what's actually underneath. Put the two findings together and the message is uncomfortable for anyone who's been buying AI by the pound: neither model size nor the "agent" label is where the value lives. The value lives in whose judgment the system encodes, and how honestly.

———

Here's where it meets the wage gap.

The market premium was never for access to AI — everyone has access, and as of this week access got dramatically cheaper. Sonnet 5 shipped as the default at $2 per million input tokens, near-Opus performance at 40% of Opus cost. Frontier capability is trickling down the stack faster than anyone expected. When capability commoditizes, the price of raw intelligence falls toward zero, and whatever remains scarce captures the premium.

What remains scarce is exactly what Bridgewater just monetized: codified expert judgment. Not expertise in your head — expertise written down, structured, scoreable, transferable into a system. The 84.7% didn't come from the analysts being smart. It came from their smartness being encoded — turned into training signal a model could absorb.

This is the same market I've been pointing you at in the Gigs section every week, viewed from altitude. The labs paying domain experts $50–150 an hour for evaluations aren't buying labor. They're buying judgment-as-data. OpenAI's 600 doctors cut medical hallucinations 71% — that story was the retail version. Bridgewater is the institutional version: don't sell your judgment to a lab by the hour, encode it into an asset you own. The expert-data market and the fine-tuning market are the same market. The only question is which side of it you're on — hourly vendor, or asset owner.

And notice who's converging on this from the operator side. Murati calls it "experts improving AI that empowers experts." Nate Herk teaches automation judgment to 325,000 people through systems. John Peslar encodes GTM judgment into playbooks his community runs. The one-person education empires and the $2 trillion hedge fund are running the identical play: take judgment that lives in a human, structure it, ship it as a system. That's the flywheel, and it's now empirically superior to raw frontier scale on domain work — by 35 points, at a thirteenth of the cost.

———

So what does the Portfolio Executive actually do with this? Three moves, and they're sequenced.

First, start building your judgment corpus this week. Every consequential call you make — the comp exception, the diligence pass, the policy read, the vendor cut — write down the input, the call, and the reasoning in a structured note. Fifty of those and you have an eval set: the thing that lets you measure whether any AI system matches your judgment. Two hundred and you have a fine-tuning asset. This is the single most valuable document you can own in the next twenty-four months, and almost nobody in your organization is keeping one. Bridgewater could run this play because the judgment was already captured. Yours evaporates daily.

Second, re-run your cost math against the new stack. If your team is paying frontier prices for non-frontier work, you're subsidizing the wrong layer. Sonnet 5 at $2/$10 against Opus at $15/$75 is a 5–7x spread for near-equivalent output on most tasks. Route by task: cheap model for volume, frontier for edge cases, and — now — a specialized layer for the domain work where a tuned smaller model beats them all. The single-model era is over for anyone running AI at production volume.

Third, reframe the build-vs-wait decision for your leadership. The next time someone in your executive meeting says "let's wait for the next model," you now have the counter, with numbers: the frontier scored 50 on Bridgewater's real work. Waiting buys you a smarter generalist; encoding buys you a specialist that's already past the trust threshold. The org that starts capturing its experts' judgment this quarter is building the only moat this week's result says still holds.

The bigger-model era trained everyone to look up — at the labs, at the benchmarks, at the release calendar. This week's number says look inward. The most valuable model weights in your company aren't in a data center. They're in the heads of your ten best people, uncaptured, walking out the door every night at six.

Bridgewater wrote theirs down. That's the whole trade.

This week's signal maps onto three chapters of Closing the AI Wage Gap:

Chapter 3 — The Multiplier argues the 8x operators are domain experts who moved AI from assistant to primary producer while keeping judgment in the loop. Bridgewater is the chapter's thesis executed at institutional scale: the multiplier was never the model, it was the expert judgment wrapped around it — and now, inside it.

Chapter 9 — The Internal Champion is the 57%-promoted-from-inside chapter. This week gives the champion a new artifact: the judgment corpus. The person who shows up with two hundred documented expert calls and a plan to encode them isn't pitching an AI project — they're holding the raw material for the org's most defensible asset. That's who gets the mandate.

Chapter 15 — The Recursive Advantage is the compounding chapter, and this week it turns literal. An expert whose judgment is encoded doesn't just work faster — their judgment runs without them, improves with feedback, and compounds while they sleep. The gap between the encoded expert and the uncaptured one stops being a productivity difference and becomes an asset-ownership difference.

Manuscript in Tier-1 agent querying — Levine, Halpern, Sagalyn pending. If you know an editor or agent at the intersection of work, AI, and organizational economics, reply to this email.

The best six from a week that had sixty. Full daily version coming — see below.

Sonnet 5 shipped as the default — near-Opus at 40% of the cost. $2/$10 per million tokens through August 31, 1M-token context, self-checking, now the default on Free and Pro plans. The production math just changed: high-volume agentic work runs on Sonnet, Opus becomes the edge-case escalation. If your AI budget assumes one model at one price, redo it this month.

Fable 5 returned worldwide after three weeks under export control. Anthropic redeployed its frontier model with an upgraded safety classifier that blocks the reported jailbreak in over 99% of cases. The takedown-and-return arc is now complete precedent: the government can pull a model, and labs will comply. Model routing with fallbacks is business continuity, not optimization. claude.ai

Altman pitched a US-led AI safety forum — and a 5% government stake in OpenAI. An FT op-ed proposing an IAEA-style international forum, plus floated equity for the US government and a dividend fund for redistribution. Governance is moving from unilateral control to negotiated structure, and labs want a seat at their own regulation. Your multi-year AI vendor commitments now carry political-structure risk — price it in.

Claude Tag went mainstream in Slack, alongside five new Managed Agents production features. Tag @Claude as a teammate; it works in the cloud and reports back — with per-session model overrides, webhooks, and an observability tab underneath. AI now has the organizational presence of an employee: it receives work by mention and reports asynchronously. Decide who manages that "employee" before it gets adopted without an owner.

Microsoft committed $2.5B to put 6,000 engineers at client sites — while Palantir's Karp said enterprises are "paying for tokens that create no value." Two opposite theories of enterprise AI value in one news cycle: human integrators at scale versus domain-specialized systems you own. Both are answers to the same ROI wall. This week's Bridgewater result is evidence for Karp's side — the specialized-system bet just posted numbers.

DRAM prices doubled in Q1 and are set to rise another 58–63% this quarter. The "RAMageddon" — with optical-interconnect startups racing to relieve it and Anthropic approaching Samsung about custom silicon. The AI bottleneck is shifting from compute to memory and interconnects. Expect inference pricing turbulence in 2027; another argument for owning smaller specialized models instead of renting the largest ones.

Six from a week that had sixty-plus.

Each week: one real build from the Portfolio Leverage Co. stack, broken down to the studs. Free readers get the map. Build Vault members get the build.

This week's lead was about encoding investment judgment into a system. Here's the operator-scale version — the one I run.

The problem had a shape, and the shape was an eleven-PM deal packet. Doing diligence for my own VC work — defense tech, AI, health tech — I was grinding through the same evaluation every time: read the deck, pull the financials, map the competitive landscape, score the risks, write the memo. Each deal slightly different, the underlying structure identical. The work that actually required judgment was getting buried under the work that didn't.

That's not a deal-flow problem. That's an architecture problem.

What it does. Sixteen diligence categories. Two hundred and fourteen structured data fields per deal. Feed it a deck and a data room, and it generates a Goldman-style investment memo — risk scoring, competitive analysis, financial modeling — formatted as a clean, auto-generated PDF. The thing I used to do until midnight now runs while I read the founder's actual emails.

The stack. A lightweight classification layer routes each input into the right one of the sixteen categories. The Claude API does the interpretation and synthesis. Structured templates enforce the 214-field discipline so nothing gets skipped. A PDF generation layer formats the memo. No vector-database cathedral. No fine-tuning. The run cost is trivial against a single junior-analyst hour.

What made the difference. Three things, none of them the model. The field schema — defining the 214 fields so they map to how an investment committee actually thinks took more iteration than any code. The risk-scoring rubric — getting the system to flag uncertainty honestly instead of confidently papering over a thin financial picture was the whole ballgame; a diligence tool that hides risk is worse than no tool. The human handoff — DueDrill produces the memo; I make the call. It does the codified 80%; the judgment stays mine.

Notice what the schema and the rubric actually are: my diligence judgment, written down and structured. The same asset Bridgewater fine-tuned into a model, one level of encoding earlier. Start there — the corpus comes before the weights, always.

Adapt it:

The free version gives you the map. The build itself — the sixteen-category schema, the core Claude prompts, the risk-scoring rubric, and the adaptation tables — is below, for Build Vault members.

↓ PAYWALL BREAK ↓

1. The sixteen categories. The routing layer that decides where every input lands:

Every document, data-room file, and founder answer gets classified into one or more categories before any synthesis happens. Classification first, synthesis second — that ordering is what keeps a 214-field memo from becoming a hallucinated essay.

2. The core prompt skeleton. Each category runs the same three-part prompt against its routed inputs:

The INSUFFICIENT rule is the single most important line in the system. It's what makes the memo trustworthy — the machine is forced to show you the holes instead of spackling them.

3. The risk-scoring rubric. Each category scores 1–5, anchored, not vibes:

The memo's front page is the sixteen scores in one table. An IC member can read the risk surface in ten seconds and go straight to the 4s and 5s. That front page replaced 80% of the "walk me through the deal" meetings.

4. Field discipline. The 214 fields exist so nothing gets skipped, but the schema's real job is honesty accounting: at the end of a run, DueDrill reports how many fields are filled, sourced, and INSUFFICIENT. A deal with 60 INSUFFICIENT fields isn't a bad deal — it's an early one, and now the diligence gap list writes itself. That list becomes the founder-call agenda.

5. The adaptation move. To repoint DueDrill at board prep, RFPs, or grant review, you change exactly two artifacts: the category list and the rubric anchors. The prompt skeleton, the INSUFFICIENT discipline, and the scoring table carry over untouched. That's the sign of a real architecture — the judgment layer swaps, the integrity layer stays.

Build it with you. I'll stand up DueDrill around your real diligence process. Sprint: $4,000 · Build Vault members: $2,800. Reply "DueDrill."

More: duedrill.com

Join The Build Vault

The research that feeds this newsletter — AlphaSignal, Rundown AI, TAAFT, AI Fire, The Code, and the operators I track daily — generates 50–60 items a week. The Leverage Brief carries six.

The Leverage Signal is the five-minute weekday read for the same audience: the two or three highest-signal tool or agent releases from the prior 24 hours, each with a one-sentence "what this means for your work" framing; one open-source repo worth knowing; one deployment pattern from the operator community; one macro signal — model, funding, regulatory, talent — in your working context before 9 AM.

It's for Portfolio Executives, CHROs, CTOs, CAIOs, CLOs, and fractional executives in AI-exposed roles. Founders and operators running AI-leveraged businesses. Executive coaches and L&D leaders whose clients are navigating AI transformation.

I want to know the readership is there before I build the pipeline.

→ SUBSCRIBE TO THE (M–F) LEVERAGE SIGNAL — Ten seconds to opt in.

Paid AI training and evaluation work for senior operators. Flexible, remote, NDA-bound.

The expert-data market is tracking toward $100B/year by 2027. Frontier labs pay domain experts for post-training evaluations, RLHF, and agent environment design. Senior operators and licensed professionals in HR, finance, legal, medicine, and engineering consistently land in the $50–$150/hr band. This week's Bridgewater result is the same market at institutional scale: your judgment is the training signal, and it's getting paid for.

Mercor — Premium rates ($75–$150+/hr for qualified domain experts). Strictest screening; best fit for CHROs, attorneys, physicians, and senior engineers. Valued at $10B after its October 2025 Series C. Note: impacted by a March 2026 supply-chain attack — review their post-breach disclosures before onboarding. → Apply via referral link

micro1 — Faster onboarding via the Zara AI interview; multiple attempts allowed. Crossed $100M ARR in December 2025. Expanding into robotics pre-training and agent simulation. $20–$150/hr depending on domain. → Apply via referral link

Meridial (by Invisible Technologies) — Expert contractor work across law, STEM, finance, linguistics, coding, and safety. No prior AI training experience required. Typically responds within 48 hours. Strong fit for specialized domain experts. → Apply at meridial.ai

Apply to all three. A $50–$150/hr income node built on expertise you already have.

The July cohort starts in under two weeks. Final seats.

This week's signal is the cohort's argument with a scoreboard attached: the frontier scored 50 on real domain work, and encoded expert judgment scored 84.7. The premium goes to the expert who captures and ships their judgment — and capturing it is a buildable skill, not a trait. The cohort is the room where you build it against your real work.

Fifteen seats. Twelve weeks. Three things you leave with: a redesigned operating week built around compounding output rather than calendar entropy; a custom AI workflow or tool you actually ship in your real work — not a prototype, a working artifact with before-and-after data that proves your judgment changed the output; and a positioning narrative that names your value at the convergence of talent and technology, because that's where the CAIO titles are going.

You don't close your gap by waiting for a bigger model. You close it by encoding the judgment only you have — and shipping proof.

→ Apply: portlev.com/cohort

Start your judgment corpus. One note per consequential call you make this week — five lines each: the situation, the inputs you actually weighed, the call, the reasoning, and what a smart generalist would have gotten wrong.

Ten of those by Friday. That's it.

It will feel trivially small against a week where a hedge fund fine-tuned a 235-billion-parameter model. It's the same asset at week one. Bridgewater's 84.7% exists because their analysts' calls were captured in a form a system could learn from. Yours currently evaporate by lunch.

Fifty notes and you have an eval set — the thing that tells you whether any AI output in your function is actually good. Two hundred and you have the raw material for the most defensible AI asset your organization can own. Every note compounds. Start Monday morning, first decision.

• The Build Vault (premium) — one real build, broken down, every week + full archive + member sprint pricing. Join
• Portfolio Executive Cohort (July 2026) — final seats. Apply
• The Leverage Signal (daily briefing) — SIGN UP HERE
• The DueDrill sprint — $4,000 ($2,800 for Build Vault members). Reply "DueDrill."
• Custom AI Build — from $5K. Scoping calls open now (BIG DISCOUNT FOR THE BUILD VAULT ANNUAL SUBSCRIBERS)
• Fractional CHRO / CLO — $15K/mo. Two Q3 slots open.
• Pre-order Closing the AI Wage Gap — portlev.com/preorder

Reply to this email. I read every one.

Yuri Kruman / Founder, Portfolio Leverage Co. · 3x CHRO · AI Trainer: OpenAI, Meta, Microsoft / PortLev.com · LinkedIn · AIWageGap.com

"The frontier models scored fifty. One firm's encoded judgment scored eighty-five, at a thirteenth of the cost. The most valuable model weights in your company were never in a data center — they're in the heads of your best people, uncaptured. Write them down. That's the whole trade."

— The Leverage Brief, July 2026

© 2026 Portfolio Leverage Company. You're receiving this because you subscribed to The Leverage Brief. Unsubscribe in one click.