Stage-Agnostic Sourcing Is Systematically Miscalibrated


The standard pitch for ML-driven VC sourcing is correct in the aggregate and wrong at the margin where it matters. Hiring velocity, web traffic, GitHub commits, app store ratings, these signals outperform gut intuition across the full universe of startups. The problem is that the full universe is not your portfolio.

Stage-agnostic scoring models are systematically miscalibrated because the signals that predict success at seed have a different shelf life than the signals that predict success at Series A. Using a stale signal isn't neutral. It's noise dressed as insight.


THE DECAY CURVE IS REAL

Hiring velocity is the clearest example. At seed, a team that is aggressively hiring before revenue is a leading indicator of founder conviction, capital efficiency expectations, and organisational ambition. The signal is fresh and predictive, founders who build fast teams early tend to build fast companies.

By Series A, hiring velocity has largely decayed as a signal. The companies that are hiring fast at Series A are often the ones that raised a lot, not the ones that are executing well. Capital causes hiring, not the reverse. The signal that was predictive at seed is now confounded by the round itself.

The same logic applies in the other direction. Revenue retention like net dollar retention, expansion rate, churn cohorts is meaningless at seed because there is no revenue to retain. It's a powerful signal at Series A and a critical signal at growth. Stage-agnostic models that include revenue retention in the seed-stage scoring layer are not ignoring this fact, they are giving it zero weight implicitly, which is different from having thought about it explicitly.


WHAT STAGE-STRATIFIED STACKS LOOK LIKE

The funds winning the sourcing game are not running one scoring model. They're running signal stacks calibrated to stage, with deliberate choices about which inputs are predictive at which moments.

Seed signal stack: Team composition and background data (founders who have built before in the domain), early hiring decisions (which functions they hired first, how fast they hired after funding), product shipping cadence (GitHub activity, product launches, beta user counts), and network graph positioning (who is on the cap table, who is advising, who they have access to).

Series A signal stack: Revenue trajectory and retention shape (not just the number, but whether the curve is accelerating or flattening), sales cycle characteristics (ACV relative to sector, time-to-close relative to cohort), product engagement depth (not DAU, but whether the product is in the critical path of the customer's workflow), and competitive positioning relative to named alternatives.

The transition between these stacks matters as much as the stacks themselves. There is a window between seed and Series A, typically between $500K and $2M ARR where the seed signals are losing predictive power before the Series A signals have enough data to be reliable. This is the hardest scoring environment. The funds that perform here supplement quantitative signals with qualitative inputs they've structured: reference calls that follow a consistent schema, founder conversation recordings analyzed for linguistic patterns, board update language that gets embedded and monitored.


THE PRACTICAL IMPLICATION

If your sourcing model was trained on a corpus that includes seed, A, and B outcomes without stage stratification, it has learned signal weightings that are incoherent. It is applying the predictive logic of seed-stage signals to growth-stage companies and missing the predictive logic of growth-stage signals at seed. The model is not wrong — it is systematically miscalibrated in ways that are hard to see in aggregate and expensive in individual decisions.

The fix is not to discard the model. It is to stratify it. Train separate models for each stage, or apply stage as a conditioning variable that reshapes signal weights rather than treating it as one feature among many.

The funds that have done this will tell you it is a materially different output. The companies that score well on a seed-stage model look different from the companies that score well on a stage-stratified model at the same funding level. That difference is where the alpha is.


THE BOTTOM LINE

Alt-data sourcing works. The evidence is solid enough that this is no longer the argument. The argument is whether you are running a model that understands signal decay or one that treats all inputs as equally predictive regardless of when in the company lifecycle they're measured.

The signal decay curve, the half-life function of each alt-data input as a function of company stage is not a theoretical refinement. It is the difference between a sourcing model that finds the companies that already look like winners and one that finds the companies that will become winners. At seed stage, those are very different lists.

Your signal is fresh at seed. It's stale at Series A. Price them accordingly.

Your Founders Write Differently Before Things Go Wrong. You're Not Reading the Diff


The distress signal arrives in the board update before it arrives in the financials. You are reading the financials.

Every month, founders send you a structured window into their psychological state: the board update. Revenue numbers, burn rate, headcount. You scan them. You check the metrics against the previous quarter. You note whether the company is on plan. What you are not doing, what almost no GP is doing is reading the language. And the language is already telling you something the numbers won't say for another two to four months.

This is not intuition. It is statistics. It is detectable, measurable, and actionable with tools that exist today. The funds that figure this out will have an early warning system. The funds that don't will continue to be surprised by distress events they technically had advance notice of.


WHAT THE UPDATE DRIFT SIGNAL LOOKS LIKE

Portfolio update language changes before financials do. The pattern is consistent enough across companies and market cycles that it has a name: the update drift signal. Here's what it looks like in practice.

Hedging increases. The quarterly update that used to say "we closed three enterprise deals" starts saying "we're seeing encouraging signals in our enterprise pipeline." The claims are softer. The tense shifts from past to present to future. Founders who were reporting outcomes start reporting process. This is not spin, it is an unconscious linguistic response to a deteriorating information environment. They know less than they knew six months ago about whether the next quarter will deliver.

Specificity falls. Early-stage updates from confident, executing founders are dense with proper nouns: customer names, contract values, named competitors displaced, specific hiring decisions and why. As execution degrades, the proper nouns thin out. "We had great conversations with several Fortune 500 prospects" is a sentence that contains no information. It is what a founder writes when they cannot write what they know.

The product-to-fundraising ratio inverts. In a healthy company at a growth stage, board updates are mostly about the product and the customers. When things start going wrong, the fundraising section expands. Founders who are worried about runway start reporting on investor meetings they've had, warm introductions they've received, the strategic conversations they're having with potential acquirers. None of this is explicitly alarming. But the ratio shift is: when the fundraising section exceeds the product section in word count, you are reading a company that has shifted its attention from building to surviving.

Response latency lengthens. This one does not require NLP. When a founder who has always responded to GP requests within 24 hours starts taking three or four days, something has changed. Either they're overwhelmed, distracted, or managing a conflict they haven't disclosed. Latency is behavioural, not linguistic, but it belongs to the same signal class: the founder's communication behaviour is changing because their internal state is changing. The update is lagging the operational reality.


THE TWO-TO-FOUR MONTH LEAD TIME

Why does the linguistic signal precede the financial signal by two to four months? Because the update reflects the founder's current knowledge, not the reported financial state.

Financial reporting, even in startups, is backward-looking by design. The Q3 board update reports on Q3, which ended last month. The metrics were aggregated last week. The analysis reflects decisions made over the past quarter. By the time the numbers show distress, the distress has been underway for months.

The founder's language, by contrast, reflects their current psychological state including the information they're holding but haven't yet structured into reportable metrics. A founder who has just lost a key account knows in the update they send before that loss shows up in churn numbers. A founder who is struggling to close the round knows in the language they use about "runway" before the cash balance shows a problem. The update is a real-time signal. The financials are a lagging indicator.

A fund that embeds monthly update text, monitors embedding-space movement over time, and flags anomalous drift in communication patterns has a materially different information environment than one that reads updates for the metrics and ignores the prose.


WHAT RUNNING THE MODEL REQUIRES

This is not a research project. The tooling exists. What it requires is operational discipline applied to data you are already receiving.

Step one: standardize the update corpus. If your portcos are sending updates in different formats, some in email, some in a shared doc, some in a Notion page, you need a consistent text extraction pipeline. This is solvable. The harder version is getting consistency in what founders report so the signal isn't confounded by format variation. A standard board update template that you enforce, kindly, across the portfolio is the prerequisite for the analytics layer.

Step two: embed and store. Every update goes through a text embedding model and the resulting vector is stored with metadata: company ID, period, date received, GP who read it, any manual flags. This is a few hundred lines of code and a vector database. It is not a data science project. It is data infrastructure.

Step three: monitor for drift. Semantic drift is measurable as distance between consecutive embedding vectors. A company whose monthly update sits close to the previous month's update is communicating consistently. A company whose update has moved significantly in embedding space is communicating differently — which is worth investigating. The drift metric does not tell you what changed. It tells you something changed.

Step four: build the flags. Specific signal patterns can be detected with lightweight classifiers or even well-crafted prompts over the update text. Hedge word frequency, future-tense ratio, product-to-fundraising word count ratio, named customer mentions per unit of text. These are countable. You can build a dashboard that flags updates above threshold on any of these metrics and routes them for GP review.

The goal is not to replace the qualitative relationship. It is to ensure that the GP who reviews the flagged update goes into the conversation with the founder having already noticed something, rather than discovering it after the fact.


THE COST OF NOT RUNNING IT

The funds that do not build this capability are not operating without information. They are operating with information they cannot use.

Every board update your portcos have ever sent is sitting in someone's email inbox or a shared drive. It is a timestamped, longitudinal record of how each founding team communicates under varying degrees of pressure. The fund that has never structured this data has paid for it through time, through relationship capital, through the management fee that funds the GP who reads the update and extracted only the surface layer.

The distress events that produce the hardest GP conversations, the ones where the fund is surprised, the runway is six weeks, the options are bad are almost always preceded by a detectable linguistic signal that no one was equipped to read. You were not paying attention to the diff.

The update is already telling you something. Start reading the whole document.

The Software Factory Has Arrived: What AI Engineer World's Fair 2026 Tells Us About Where AI Is Going

The AI Engineer World’s Fair opens in San Francisco on June 28. It has more than 6,000 attendees, 300 speakers and 29 tracks. That scale matters because this is not a conference about what AI might become. It is a conference about what AI has already become, and what now breaks when companies try to run it at scale.

The people attending are not tourists. They are engineers, founders, product leaders and AI operators who have to ship working systems on Monday morning. Read across the program and one thing becomes clear. The industry has crossed a threshold. The question is no longer whether AI can generate code, answer questions or perform tasks. The question is whether we can verify it, govern it, secure it and trust it enough to let it act.

The biggest bottleneck has moved from generation to verification. AI can now produce more output than humans can review. Uber is running 1,800 AI generated code changes per week. Greptile has analyzed 1 million AI generated pull requests. GitHub Copilot, Factory, Qodo and others are all dealing with the same second order problem. Once machines write faster, humans become the constraint. The next major layer in AI may not be another model company. It may be the verification layer that tells us whether the output is correct, safe and usable.

This is especially clear in software. GitHub CEO Thomas Dohmke is asking what the future of the software development lifecycle looks like. That question alone tells you where we are. Software is being reorganized around AI agents that write code, open pull requests, run tests, respond to failures and iterate. Humans move from authors to supervisors. The companies that figure this out will not just have better tools. They will have a different cost structure.

A new discipline is also emerging around the model. The conference calls it harness engineering. It is the scaffolding, state management, orchestration, memory, tool calls, retries and error recovery that make agents work in production. This matters because many agent failures are not model failures. The model did what it was asked to do. The system around it gave the wrong context, the wrong tool or the wrong recovery path. The model is becoming one part of a larger system. The harness is where a lot of the moat will live.

Prompt engineering is also giving way to context engineering. Prompt engineering asks what instruction to give the model. Context engineering asks what the model should know right now. For long running agents, that is the operating system. Context compaction, memory offloading and cost accounting are becoming core production primitives. A weaker model with excellent context management may beat a stronger model with poor context discipline.

MCP is another signal of where the industry is going. The Model Context Protocol is moving from experiment to infrastructure. Figma, Docker and Microsoft are all building around it. But the same protocol that lets agents connect to tools and data also expands the blast radius. An agent that can read a codebase, call APIs and access internal systems is not a chatbot. It is an actor inside the company. That creates a new security and governance problem.

The same issue appears in regulated industries. Two Sigma is presenting on AI agents that assume employee identity and access internal tools. PayPal is presenting on agent initiated payments across ChatGPT and Google AI Mode. These are serious trust boundaries. What is an agent allowed to do? Whose identity does it act under? Who is liable when it makes a mistake? AI capability has moved faster than enterprise governance. The next 12 to 24 months will be about closing that gap.

Voice is another major surface. Text agents fail quietly. Voice agents fail in public. They interrupt, lag, misunderstand and talk over people. Most voice systems are still pipelines that turn speech into text, send it to an LLM and then turn the answer back into speech. Native speech systems should eventually collapse that stack, but we are still in the difficult gap between possible and reliable.

The open weights ecosystem has also become credible. Hugging Face is hosting millions of models and serving large enterprises. Open models are no longer just research tools. They are becoming part of the production stack. That changes cost, control, privacy and deployment strategy.

The bigger point is this. Coding agents were the proof of concept. Knowledge work agents are the business. Legal analysis, financial research, medical reasoning and investment work will need domain specific tools, workflows and verification systems. You cannot simply point a coding agent at a new domain and expect it to work.

The AI Engineer World’s Fair 2026 is not about AI potential. It is about AI operations. How do you run agents at scale? How do you verify them? How do you govern them? How do you secure them? The generation problem has been substantially solved for a large class of tasks. The operations problem is now the center of gravity. The companies that understand this are building for the next five years. The companies still focused only on generation are solving yesterday’s problem.
 The conference runs June 28–July 2, 2026 at Moscone West Convention Center, San Francisco. Please visit https://www.ai.engineer/worldsfair/2026

Synthetic Dissent: Your Agentic Investment Committee Needs a Correlation Audit

The multi-agent investment committee is becoming the new toy in venture.

Five AI agents read the same memo. One is the market-sizing agent. One is the technical diligence agent. One is the financial model agent. One is the founder-pattern agent. One is the skeptic. They debate, vote, and produce a dashboard with confidence scores.

It looks rigorous.

It may also be fake rigor.

The problem is not that the agents are useless. The problem is that many of these systems treat five agents as five independent opinions when, in reality, they are often one opinion with five different costumes.

If the agents run on the same foundation model, read the same memo, retrieve from the same data room, and inherit the same model priors, their votes are not independent. They are correlated. And once the votes are correlated, the committee math breaks.

This is the first failure mode of the agentic investment committee: it confuses headcount with epistemic diversity.

Five agents does not mean five independent views.

Sometimes it means one view, sampled five times.

The Clones Are Voting

Here is what happens in many first-generation agentic IC systems.

A fund gives the same deal memo to multiple agents. Each agent is assigned a persona. The market agent asks whether the TAM is large enough. The product agent asks whether the product is defensible. The founder agent asks whether the founder has the right background. The finance agent checks burn, margin, and revenue quality. The skeptic writes the bear case.

On the surface, the outputs look different. One agent worries about go-to-market. Another worries about technical depth. Another worries about customer concentration.

But the underlying model may still be the same. The data may still be the same. The retrieved context may still be the same. The priors may still be the same.

So what looks like disagreement may only be formatting variance.

This matters because a committee is valuable only if its members bring genuinely different models of the world. A former biotech operator and a consumer marketplace investor may disagree for real reasons. They have different scars, different pattern libraries, different examples in their heads, and different mistakes they are trying not to repeat.

AI personas do not automatically have that. A prompt that says “you are a contrarian growth investor” does not create a contrarian growth investor. It creates a language model role-playing one.

That distinction is not philosophical. It is mathematical.

The Missing Math: Effective Committee Size

If five IC agents are independent, five votes can add real signal.

But if their errors are correlated, the effective number of independent opinions collapses.

A simple approximation is:

Effective independent agents = n / [1 + (n - 1)ρ]

Where:

n = number of agents
ρ = average correlation between their errors

Now apply this to a five-agent IC.

If correlation is zero:

5 / [1 + 4(0)] = 5.0

You really have five independent agents.

If correlation is 0.5:

5 / [1 + 4(0.5)] = 1.67

Your five-agent IC is closer to 1.7 independent agents.

If correlation is 0.8:

5 / [1 + 4(0.8)] = 1.19

Your five-agent IC is basically one agent with a panel discussion.

This is the uncomfortable truth: a unanimous “strong invest” from five highly correlated agents may not be five signals. It may be one signal echoing through five prompts.

The dashboard says 5–0.

The math says 1.2–0.

Why Consensus Is the Wrong Target in Venture

The instinct to build toward consensus is understandable.

IC meetings are expensive. Partner disagreement is uncomfortable. LPs like process. A system that produces clean, confident outputs feels mature. It feels institutional. It feels like judgment has been made legible.

But venture returns are not produced by maximizing consensus.

They are produced by being right on a small number of non-obvious companies before the market agrees.

That is why venture is such a strange asset class. In many portfolios, most companies do not matter to the final return. A small number of outliers matter enormously. The fund is often made by the investment that looked weird, early, overvalued, too small, too messy, or too soon.

This creates a problem for agentic IC design.

If the system is optimized to reduce disagreement, it will become very good at killing strange companies. It will push decisions back toward the historical average. It will reward companies that look like prior winners and penalize companies whose best feature is that they do not fit an existing pattern yet.

That is dangerous because the average of historical venture data is not the source of venture returns. The outlier is.

A venture IC should not ask only, “Do we agree?”

It should ask, “Where do we disagree, and is the disagreement informative?”

The best IC output is not a confidence score. It is the unresolved argument that forces the partner to think.

The Real Job of a Devil’s-Advocate Bot

The devil’s-advocate bot should not be a decorative skeptic.

It should not produce a polite paragraph saying, “Risks include competition, execution, and fundraising environment.”

That is not dissent. That is memo garnish.

A real devil’s-advocate bot has one job: construct the strongest possible case that this investment should not happen.

It should try to kill the deal.

Not because the fund should always listen to it. Because if the strongest kill memo is weak, that is information. If the strongest kill memo is strong but answerable, that is information. If the strongest kill memo is strong and no one can rebut it, that is also information.

The purpose of synthetic dissent is not negativity. It is compression. It compresses the hardest objections into a form the human IC cannot avoid.

The bear case should not be hidden below a final recommendation. It should be the main artifact.

The system should produce three things:

  1. The strongest invest case.

  2. The strongest kill case.

  3. The unresolved disagreement between them.

The third output is the most important.

Build for Disagreement, Not Theater

If funds want agentic ICs to matter, they need to stop measuring how often agents agree and start measuring whether agent disagreement is useful.

That requires different architecture.

First, use decorrelated retrieval.

Do not let every agent read the same memo and the same data-room summary. Give agents different information sets. One agent reads only founder background and references. One reads only customer calls. One reads only technical diligence. One reads only competitive data. One reads only the financial model.

This creates real information asymmetry. When agents disagree, the disagreement now means something. It may reflect a tension between customer love and weak margins, or between founder strength and market timing, or between technical elegance and weak distribution.

Second, use heterogeneous models where possible.

Different prompts on the same model are not enough. Use different model families, different fine-tunes, different retrieval strategies, and different evaluation rubrics. You may not eliminate correlation, but you can reduce it.

Third, make the skeptic structurally independent.

The kill agent should not be asked to “be balanced.” Balance is the synthesizer’s job. The kill agent should be adversarial. It should have permission to be harsh, specific, and one-sided.

Fourth, track calibration by agent, not just by committee.

Most systems will track whether the final recommendation was right. That is useful but incomplete. You should also track which agent dissented, when it dissented, and whether that dissent would have improved the decision.

The most valuable agent may not be the one that agrees with the final vote most often.

It may be the one that was uncomfortable for the right reasons.

A Better Agentic IC Dashboard

The current dashboard often looks like this:

Market agent: Invest
Product agent: Invest
Founder agent: Invest
Finance agent: Invest
Skeptic agent: Invest with risks
Final recommendation: Strong invest
Confidence: 87%

That looks decisive.

But it may be dangerously overconfident.

A better dashboard would look like this:

Nominal agent vote: 5–0
Estimated error correlation: 0.72
Effective independent vote count: 1.32
Strongest invest argument: Founder-market fit is unusually strong, customer pull is early but real, and the market may inflect faster than incumbents expect.
Strongest kill argument: The current traction may be services-led, gross margin is not yet proven, and the buyer may lack budget ownership.
Irresolvable disagreement: High
Human decision required: Yes

That dashboard is less comforting.

It is also more honest.

The goal is not to make the IC quieter. The goal is to make the unresolved judgment visible.

The Human Is Still the Committee

The strongest version of agentic IC does not replace the partner.

It makes the partner less lazy.

A bad system gives the human a recommendation to endorse. A good system gives the human an argument to overcome.

That distinction matters. Venture judgment is not only about pattern recognition. It is about knowing when to violate the pattern. It is about asking whether the thing that looks like a flaw is actually the source of the opportunity.

A consensus machine will struggle with that.

A dissent machine might help.

The right question is not, “Can AI agents vote like an investment committee?”

The better question is, “Can AI agents expose the disagreement that a real investment committee would otherwise avoid?”

If your agentic IC is producing clean, confident, unanimous outputs on most deals, something is probably wrong. Either your deals are unusually obvious, or your system has been engineered to launder uncertainty into consensus.

Consensus among correlated agents is not conviction.

It is noise wearing a suit.

Build the devil’s advocate. But more importantly, measure the correlation.

Because in venture, the danger is not that your agents disagree.

The danger is that they all agree for the same wrong reason.