The Perturbation Network: Your Bottleneck Isn't Thinking — It's Access to Live Problems

Once you can compile problems into governed systems, the scarce input is high-quality labels of live reality — collected through honest, instrumented conversations, not disguised sales.

Scott Farrell · LeverageAI · Demand-side sensing for problem-to-system factories

There is a moment after the factory works when the old story stops fitting. You can take a difficult process — people, spreadsheets, documents, systems — compile its operating world, place deterministic software, AI and human judgment correctly, and ship a governed application that improves through use. The architecture is no longer the mystery. And yet the calendar is still thin.

The instinctive diagnosis is almost always the same: I need the commercial side — more leads, better outreach, mentoring that might become work, sales training for people who would rather build. That diagnosis is half true and wrong about the mechanism.

Your bottleneck has moved. It is no longer thinking capacity. It is qualified access to live problem shapes.

What you need is not a cold-outreach machine. It is a perturbation network: a deliberate human sensor array of domain-labelling conversations, instrumented like publishing probes, with a capture object and a trust protocol that keeps mentoring from turning into a con. After this piece you should be able to run those conversations Monday — full question set, signal card, outreach message, and a refusal rule for what not to build.

The factory without samples

A problem-to-system compiler does not starve for ideas the way pure services starves for hours. It starves for incoming samples: operating anomalies, domain frustrations, half-broken workflows, institutional workarounds, and comments from people close enough to reality to notice what outsiders cannot. Those samples are not product ideas — product ideas are cheap and usually wrong. They are labels of reality only someone inside the work can supply:

• This process is genuinely painful.
• This spreadsheet is acting as a hidden application.
• This decision still requires three senior people in a room.
• Nobody agrees what “correct” looks like.
• We reconstruct this answer every month.
• This is where scope and margin disappear.
• This output reaches the customer, but nobody can explain it.
• The system is officially successful, and everyone works around it.

Contacts label reality; your compiled IP supplies the collisions. That is the whole demand-side job once construction is no longer the hard part. Treat the network as a lead list and you either go quiet or go false (mentoring as camouflage). The first starves the factory; the second corrodes the trust that makes sensing possible.

Mentoring instinct, sharpened

The honest first move many builders make is: contact everybody I know — not for work, for mentoring — and ask what they would use this for. Indirectly it is still commercial. That squirm is useful; keep the instinct, change the epistemology.

Mentoring-as-frame drifts toward generic advice about you. What the factory needs is someone opening a window into their world:

Here is a strange, expensive thing we keep doing that makes no sense.

That is the reframe: domain-sensing conversations, not mentoring-as-soft-sales. You are collecting high-quality human discrimination about local operating reality — the same scarce input frontier labs still buy as labels — not cheerleaders. Some conversations become clients; many will not; a few hand you one sentence that expands into a different product. The network’s job is to keep that surface live, not convert every contact into a SOW.

One spark × dense canon (n=1)

I took one comment from a Tesla employee and built out a whole case study — a collision claim, n=1, treat it as such. He did not hand over a product design. He handed over one sharp anomaly:

I think the AI got that one wrong.

In a service context, a wrong AI triage is not a chatbot embarrassment. Remote pre-diagnosis and pre-ordered parts mean a model’s guess can already commit real inventory and shop time before a human re-opens the case.¹ The comment was a wedge into authority, receipts, exceptions, customer trust and organisational learning.

Because those thoughts already existed in compiled form — authority design, deterministic gates, governance, service memory, cognitive provenance — the anomaly expanded into the Invisible Foreman, Exception Sink, Friction Wedge, proposal cards, a micro-judgment DAG, and an architecture transferable to insurance, ITSM and lending. The anecdote was the spark; the canon was the fuel. Without the wiki it might have stayed a funny Tesla story; with it, one desk sentence became a system hypothesis.

That is the commercial event IP-routing already names:

idea × person × capability × company × live problem × time

An idea alone is latent inventory until the right person introduces it to a live situation before the moment closes. A quieter version: a mid-sized data consultancy conversation that invented no new framework, only gave an existing one a destination. Before: inventory. After: a candidate. The missing object was the edge. Contacts are domain sensors, problem nominators, language sources, reality checks and access paths — not “prospects” until a signal card says so.

Reverse-engineered into a signal card, the layer separation is obvious: the desk comment is fact; the service-system expansion is inference. Confusing them produces mythical origin stories no future conversation can live up to.

The card does not claim “Tesla needs my product” or invent a conversion rate. It preserves a spark as structured evidence a dense canon can collide with — the difference between a war story and a sensing system.

Bad questions produce AI brainstorms

Most people cannot design your product from an architecture description. Ask “where could you use AI?” and you hear customer service, report writing, emails, summarisation, or “something with our data.” Feature categories, not problem shapes — they do not locate trapped expensive cognition, who funds the pain, or what would have to be true for a governed application to earn its keep.

Shape-exposing questions force labels of expensive reality. Use the full set:

Show one specimen — then their world

You need enough concrete structure for analogy, and not so much that the conversation becomes a pitch deck. The specimen sequence is owned elsewhere; here it is only the entry move for sensing.

I take a high-value process trapped across people, spreadsheets, documents and systems, compile its operating world, decide which steps need deterministic software, AI or human judgment, and build a governed application around it. This Power BI example is one specimen. Where do you see a similar problem shape in your world?

Enough for recognition; not enough to prescribe their solution. The FDE BI / Data Readiness Review specimen proves the method exists — it is not their roadmap.

The signal card

Do not rely on memory afterward — memory launders inference into fact. Capture each conversation as a repeatable object: the solo version of a consultancy opportunity card, with observed fact separate from system inference and recognition separate from commercial commitment.

A card with only the first four fields filled is still useful; a beautiful hypothesis with no observed fact is fiction. Discipline is separation, not completeness theatre.

Conversations as instrumented probes

Publishing can already run as an active sensor: emit one concept-sized probe, observe response or silence, feed the result back. The one-to-one conversation is the same loop at higher resolution. You emit a product sentence, visual, problem shape or provocative distinction; they return recognition, objection, analogy, a different buyer, “that would never work here,” “we have exactly that problem,” or silence. Each is telemetry.

The conversation is not merely networking. It is instrumented market sensing — provided you record what was tested and what came back.

End every conversation with:

Who sees a completely different part of this problem from you?

Not “do you know anyone who might buy?” — it creates the next perturbation and treats the network as a graph you walk, not a funnel you fill.

The trust boundary is infrastructure

One failure mode looks sophisticated and is corrosive: saying you only want mentoring while privately treating the person as a lead. When the conversation turns commercial — and good sensing often does — the gap between frame and intent becomes visible. People forgive commercial interest; they do not forgive being managed.

The false binary: a pure mentoring mask feels safe until it doesn’t and trains distrust of the frame; a pure sales frame makes them armour up into product brainstorms or procurement theatre, not operating labels. The membrane that holds both is an honest dual-role sentence — said early, not after awkwardness arrives:

I’m mapping where a particular AI delivery pattern genuinely has value. I’m not coming with a conventional sales pitch. I’d value your domain lens and one or two examples from your world. If something commercially meaningful appears, I’ll call it that rather than pretending it isn’t there.

You are asking for help and commercially active; both can be true. Honesty is not virtue signalling — it is the operating condition that keeps domain sensors willing to open the window next time, including when the answer is “nothing fits.”

Three objections that show up immediately

“This is just networking with extra jargon.” Without artefacts, yes. With a question set, signal card, telemetry tally and refusal rule, it is demand-side sensing. Drop the artefacts and it collapses into coffee chats.

“People will see through the dual-role sentence as sales.” Some will. Named commercial interest is less damaging than denied commercial interest. The sentence makes sales an allowed outcome, not a hidden one; negative results remain telemetry.

“One Tesla comment isn’t a method.” Correct. The comment is n=1 evidence that density multiplies signal. The method is sensors, questions, cards, instrumentation, trust protocol and refusal doctrine — the process that makes sparks usable.

What the other side gets

Domain-sensing fails if the contact is free labour for your pipeline. They should leave with a one-sentence architecture they can steal, one short specimen that proves the method without prescribing their answer, better questions for their own organisation, a concise synthesis afterward (promised — so send it), and permission for a negative result without social penalty. Forty-five minutes of labelling with nothing returned is extractive research under a collaborative costume. The synthesis is the receipt that the conversation was mutual.

A practical outreach message

Something you can send without rewriting your ethics each time:

Hi [Name] — I’ve spent the last six months developing and building a particular AI delivery pattern: taking a difficult business process spread across people, spreadsheets, documents and systems, compiling its real operating world, and turning it into a governed application with AI used for judgment and deterministic controls around authority and evidence.

I’m not approaching you with a conventional sales pitch. I’m trying to test where this pattern genuinely applies—and where it does not. You know [domain/company/function] far better than I do, and I’d value 30 minutes of your perspective.

I’ll show you one short working example, then I’d mainly like to hear about processes in your world where senior people repeatedly reconcile messy information to make a decision, produce a scope, or defend an answer. I’ll send you a concise synthesis afterward. If nothing fits, that is useful evidence too.

Broaden the sensor surface. Narrow the construction doctrine.

The risk in contacting everyone is a hundred interesting problems and scatter across all of them. A useful structural lesson appears in reported coverage of a leaked DeepSeek investor conversation. Independent reporting notes that a closed-door May investor meeting transcript circulated widely and that DeepSeek did not officially confirm authenticity — treat the specifics as reported, not first-party doctrine.² Secondary transcript packaging attributes a prioritisation metaphor: consumer and app-layer opportunities as “sesame seeds,” the AGI main line as the “watermelon” — pick seeds on the way, do not stop the car for them; leave whole product categories that are good businesses but off the intelligence main line to others.³

You do not need their roadmap. You need the couplet:

Broaden the sensor surface. Narrow the construction doctrine.
More perturbation at the edge; more refusal at the core.

Your main line is not “build anything with AI.” It is closer to:

Find high-value organisational cognition trapped in fragmented evidence and expensive human reconciliation; compile the world into a language-based learning layer; place deterministic software, AI and human judgment correctly; and deliver a governed application that improves through use.

That is the watermelon. An opportunity that does not advance that line may be a sesame seed: record it, route it, introduce somebody else, publish the insight — do not necessarily build it. Sensors discriminate at the edge; the factory refuses at the core. Confuse those jobs and strong compilers become exhausted generalists.

Do not optimise for “interesting conversations.” Interesting is cheap. The tally that matters: filled economic consequence, next edge, hypothesis still on the main line. A week of delightful chats with zero cards is social life with a professional alibi. If the pattern hypothesis cannot be stated as a sentence that advances the watermelon line, the build decision defaults to no — publish or route instead.

The loop you were missing

You do not need thousands of cold leads. You need a network that continuously supplies:

real problem
  → sharp observation
  → wiki collision
  → pattern hypothesis
  → small specimen
  → buyer recognition
  → paid discovery
  → governed application
  → learning write-back

Domain-sensing owns the first half. Downstream commercial protocol and system design are sibling territory — hand off when the signal card earns it. The next move is not “network more.” It is:

Deliberately build a human sensor network around your compiled intelligence factory — then install enough discipline that every conversation can become evidence, an idea, an introduction, a product candidate or an honest rejection.

Your bottleneck has moved. The thinking is already good enough to compile systems. What the factory needs now is live labels — and a network run as sensors, not as half-denied sales calls.