The photograph a fund is built on
An investing partner writes a thesis memo in March. It says: vertical SaaS for dental practice management is underpenetrated, incumbents are asleep, distribution through practice management associations is the wedge. The memo cites twelve customer calls, a market map, three comparable exits. It is accurate, carefully so, for the week it was written. It is also, structurally, a photograph. Heraclitus's river will not hold still for it, and the memo cannot tell you, eighteen months later, which of its twelve calls have since become wrong.
This is the oldest problem in the trade wearing a new name. Heraclitus, writing around 500 BC in Ephesus, said you cannot step twice into the same river, because other waters are always flowing on. He did not mean things change quickly. He meant identity is not a substance sitting still — it is a pattern maintained through turnover. A market is exactly this kind of thing. Dental SaaS in March is not dental SaaS in the following March: two new entrants have shipped, one incumbent got acquired and stopped sleeping, and the associations that were the wedge quietly signed an exclusive with someone else. The category name persists. Almost nothing underneath it does.
What arrives
Four streams matter, and none of them arrive on the schedule a memo review does.
Regulatory and corporate filings — Form D notices, UCC liens, state incorporation records, patent applications — surface new entrants and financing events, often before anyone issues a press release. Hiring signals — job postings, LinkedIn role changes, engineering headcount growth rate scraped from public profiles — reveal where a company is actually investing, which is a better predictor of strategy than anything in a deck. Product telemetry — app store review velocity, API changelogs, integration marketplace listings, pricing page changes captured by scraping — shows what a company shipped, not what it claims to have shipped. And market structure signals — customer concentration inferred from case studies, channel partner announcements, competitor pricing moves — describe the shape of the field the thesis is supposed to be about.
Each of these streams is noisy individually. A single job posting for a "Head of Payments" is not news. Six job postings for payments roles across four competitors inside one quarter, combined with two Form D filings citing "embedded fintech," is a market structure changing under a thesis that never mentioned payments.
What is held
The naive response is to read more, more often — a faster corpus. That fails the same way the slow one does, only sooner. What must be held instead is not a bigger snapshot but a set of beliefs, each with provenance and a timestamp: "incumbents have not shipped a mobile product" — confirmed three weeks ago, source: four app store listings and one customer call, decay rate estimated at one recheck per month, because incumbent product cycles in this category run slow. Compare "no well-funded new entrant has emerged" — confirmed yesterday, source: Form D scrape, decay rate one recheck per week, because financing events in a hot category cluster and compound fast.
The point of carrying decay rates per belief, rather than a single freshness date on the whole memo, is that the components of a thesis do not go stale at the same speed. Regulatory posture might be good for a year. Competitive headcount can flip in six weeks. A memo that is dated once, at the top, hides this. A belief store that timestamps each claim separately turns staleness into something you can query rather than something you discover in a board meeting.
What triggers revision
Revision should not wait for the partner to remember to look. It should be triggered by the stream itself crossing a threshold set when the belief was first written down.
- A competitor's engineering headcount growing faster than 15% quarter over quarter in a role cluster adjacent to the thesis's wedge.
- A Form D filing naming a use of proceeds that overlaps the thesis's claimed white space.
- App store review sentiment or velocity for an incumbent shifting outside its trailing twelve-month band, suggesting a shipped feature the market map did not anticipate.
- A channel partner previously described as exclusive to the thesis's target segment announcing an integration with a rival.
None of these events, alone, refutes a thesis. Together, and against a belief that was written down with an explicit confidence and an explicit half-life, they do the job a snapshot cannot: they tell the partner exactly which load-bearing assumption just cracked, and how long ago the crack started, rather than presenting a diffuse sense that "the market feels different now."
What the operator sees
What the partner should see, in practice, is not a dashboard of raw feeds — nobody wants to watch a job-posting scraper scroll — but a belief ledger attached to the thesis itself. Each claim in the original memo sits alongside its current status: confirmed, decaying, contradicted, or stale-unmonitored. A claim moves from confirmed to decaying automatically once its recheck interval elapses without a fresh signal. A claim moves to contradicted when an incoming stream crosses the threshold set for it. The partner is not asked to re-read the market. The partner is shown, in a line, that the wedge assumption is now contradicted, as of nine days ago, on the basis of two filings and a hiring pattern, and is invited to re-underwrite that one component rather than the whole thesis.
This is the operational meaning of the Heraclitean point applied to this domain. The frozen memo is the corpus: accurate at the instant of writing, silently wrong afterward, unable to say which parts have failed. Live diligence sprints — the partner doing a fresh round of ten calls before an IC meeting — are the Large World Model's episodic camera: a genuinely better look, but a look, taken once, closed again, blind to everything that moves between visits. The belief ledger, continuously reconciled against filings, hiring data, telemetry and market structure signals, is the instrument left in the water. It does not claim to know the market better at any single instant. It claims to know, at every instant, how old its knowledge is.
What it costs
None of this is free, and the cost is not primarily computational. It is attentional and organisational. Someone has to decide, for every load-bearing claim in a thesis, what stream would falsify it and what threshold on that stream counts as falsification. That is underwriting discipline that most memos never perform, because a memo is written to persuade an investment committee once, not to be monitored for two years. Building the ledger means writing the memo differently from the start: not "incumbents are asleep" but "incumbents have not shipped mobile, confirmed via four app store listings on this date, recheck monthly, contradicted if any incumbent ships a mobile release with over 4-star rating within sixty days."
That is more work up front and it exposes the partner to being wrong faster, in public, inside the fund's own systems. It is also the only design whose error is bounded by the market's own rate of change rather than by the partner's memory of when the memo was written. The characteristic failure of the trade — a thesis defended for a year after the market it assumed has dissolved — is not a failure of judgement at the moment of investment. It is a failure of instrumentation afterward: no gauge was left in the water, so the partner had no way to know the river had moved until a competitor's Series B round made it undeniable.
Two objections worth taking seriously
The first: most of what a thesis rests on barely changes. The size of the dental practice market, the basic economics of practice management software, the regulatory posture of state dental boards — these move on the scale of years, not weeks. A monitoring apparatus tuned to catch every hiring signal is chasing news, not insight, and most news has a half-life measured in days. This is true, and it is exactly why the ledger separates claims by decay rate rather than monitoring everything at the same tempo. The market-size claim gets rechecked annually. The competitive-headcount claim gets rechecked weekly. The failure mode being fixed is not "the memo didn't update fast enough." It is "the memo never said which of its claims were fast."
The second: continuous streams are useless without someone doing continuous underwriting on them, and underwriting time is the actual bottleneck, not access to filings and job boards. Correct. Intake is not the whole system. But intake sets the ceiling on what can ever be recovered — a partner who never tracked hiring signals cannot, no matter how sharp the eventual re-underwriting, reconstruct the fact that headcount tripled eight months before the round closed. Continuous intake with delayed reconciliation still lets the partner recover the correct, honestly timestamped belief late. Frozen intake cannot recover it at all. That is the difference the ledger is built to preserve.