The decision is small and it is due this month: whether to spend one hour keeping a record you will not need for a year. Assemble a data room during a raise and it costs four to six weeks and produces a pile of documents all dated in the same fortnight. Keep it as you go and it costs an hour a month and dates itself. The hour is cheaper than the weeks, and it buys something the weeks cannot.
What it buys is the date. A dated record you maintained before anyone asked is evidence a later investor cannot manufacture, and almost none of that evidence comes from the documents themselves. It comes from when they exist.
The date is not metadata on your proof. It is the proof.
This piece is not a list of what goes in the room. The folder structure is the easy part and every law firm publishes one. The hard part is that the room built the week before diligence looks exactly like the room built to fool diligence, and there is one property that separates the honest founder from the retrofitter. That property is time, and you cannot buy it in the fortnight before a term sheet.
Why these three models
Three mechanisms decide whether a dated record works, and they fail in different places, which is the point of running them together.
The first is a signaling model: a date separates the honest founder from the liar only when it is expensive for the liar to reproduce. The second is a queueing model: it explains why the raise itself produces the suspicious cluster, whatever your intentions. The third is an aggregation model: it explains why a reader who reads dates across independent sources can tell a real trail from a manufactured one, even when every single document looks perfect. One says the date is worth having. One says why you cannot fake it late. One says how the reader catches the fake. Equilibrium, complex, and random: three different failure modes, so a weakness in one does not sink the read.
The model-thinking discipline would append a behavioral layer here. We folded it into the blind spot instead. The reason founders skip the hour is present bias and the honest belief that they will assemble it later. That explains the skipping. It changes nothing about what the date does, so it earned a mention, not a card.
The framework: three reads on one dated record
1. The signal: the date is the costly, uncopyable part
Michael Spence’s 1973 result is the spine. A signal carries information only when it is more costly for the type you would be mistaken for than for the type you are.5 Education separates able workers from less able ones because completing it is harder if you are less able. Its teaching value is beside the point in the model. When the signal is equally cheap for everyone, the market learns nothing and discounts it to zero. That is a pooling outcome. When it is costly enough to deter the imitator, it separates.
Apply that to a document date. Anyone can produce a board minute today. Only a founder who wrote one a year ago can show you a year-old one. The document is cheap. The date is not. A record maintained monthly costs the honest founder almost nothing, an hour spread across twenty-four months. Reproducing it retroactively costs the retrofitter a great deal: they must fabricate a consistent trail across bank statements, mobile-money settlement logs, a registrar, an email account and a code history they do not fully control. The cost gap is the whole mechanism.
So the single-fortnight cluster of creation dates carries a meaning of its own. It is the pooling giveaway. It is the exact thing produced by the diligent-but-late founder and the dishonest one alike, which is why it carries no information and gets discounted. The continuous trail is the separating move. It works because a liar would not pay to build it.
Signaling with costly separation.
- Assumes the investor reads dates and can distinguish a lived trail from a manufactured one.
- Fits because the honest act is cheap when repeated and the fake is expensive to reconstruct after the fact.
- Breaks when a date can be forged cheaply and no independent counterpart exists to check it against, so the signal stops separating.
- Counteracts the incentive to retrofit proof the week before a raise.
- May reinforce box-ticking, if founders chase dated documents instead of a real business.
2. The queue: why the raise itself makes the cluster
Assume the founder is honest and simply late. The suspicious cluster still appears, and queueing theory says why. Kingman’s formula shows that in a single-server queue the mean wait and its variance blow up nonlinearly as the server approaches full utilization, and they rise with variability in arrivals and in service time.6 The effect is worse than the sum of its parts near saturation.
During a raise the founder is that single server. Two streams of work arrive at once: the investor’s document requests, and the backlog of things that were never created when they happened. Utilization runs near one hundred percent for four to six weeks.1 Output arrives late, and it arrives with errors: the cap table that does not tie to the SAFE, the contract signed in a hurry, the metric redefined to look better. Every artifact carries the same fortnight’s date because that fortnight was the only window the work could run in.
The hour a month is arrival smoothing. It keeps the founder’s utilization low, so the queue never forms, so the dates spread across the year the way a real history does. The founders who keep ongoing investor conversations clear diligence in two to four weeks rather than the six a cold, unprepared process takes, and the reason is not charm.2 It is that they never let the queue build.
Queueing and congestion.
- Assumes the founder is the bottleneck resource and document work competes for one attention budget.
- Fits because a compressed raise window drives utilization to saturation, where lateness and error rise fastest.
- Breaks when the work is genuinely parallel, so a lawyer and an accountant produce the room while the founder does not, and no queue forms.
- Counteracts the belief that a burst of effort late can match steady effort early.
- May reinforce over-processing if the monthly hour expands to fill more than an hour.
3. The corroboration: why independent dates expose a fake
Grant that a determined founder fakes the dates anyway. The third model says how the reader catches it. Francis Galton watched 787 fairgoers guess the weight of an ox in 1907. The median guess was 1,207 pounds against a true dressed weight of 1,198, closer than any single expert.7 Independent estimates aggregate because independent errors cancel and the shared information survives.
Turn that around for diligence. An investor estimating when something really happened reads many independent timestamps: a bank statement, a mobile-money settlement log, a registrar filing, an email header, a git history, a customer’s signed LPO. Each is a noisy clock. Independent, they converge on the truth. A fabricated cluster cannot make them agree, because the founder does not control the counterparties’ clocks. The fake stands out in exactly the place the honest trail gets sharper.
This is not theory in litigation, where the same problem is solved the same way. Courts authenticate electronic records through metadata, which records the date of creation and every later edit, under the framework the Lorraine opinion set out.8 Forensic examiners flag a contract as backdated when its file was created months after its signing date, and the defense against a backdating allegation is to corroborate your timestamp with independent sources: email headers, cloud version history, audit logs.9 An investor’s associate is running a cheaper version of the same check. Stale drafts left beside final versions, and contracts presented as in force but never signed, are the ordinary red flags of it.4
Independent-source aggregation.
- Assumes several dated sources exist that the founder does not fully control.
- Fits because independent errors cancel, so cross-source dates converge on the truth and a fabrication cannot align them.
- Breaks when every source is generated by the founder alone, so the errors correlate and aggregation adds nothing.
- Counteracts the assumption that a clean single document is enough.
- May reinforce a false sense of safety if the sources only look independent.
GEER: the moves, starting with the one that costs nothing
Order them by cost and by how easily you can undo them.
Keep the timestamps you already generate. Your bank, your mobile-money dashboard, your accounting tool, your email and your code repository already date everything you do. You do not have to build a trail. You have to stop deleting one. Retain access and history. Cost: nothing.
Book one recurring hour a month. Drop the month’s real artifacts into a dated folder as they occur: the board note, the signed contract, the current cap table, a snapshot of your four numbers. You are not writing anything new. You are filing what already has this month’s date on it.
Fix your four numbers under version control. A metric that can be redefined between rounds carries no date and no credibility. One where each month’s figure is stamped and frozen becomes evidence. This is the same discipline as the monthly update, aimed at a different target: not the note you send, but the date under it.
Let third parties date the heavy items. An auditor’s sign-off, a lawyer’s execution copy, a registrar’s filing carry a date you did not set, which is why they are the most expensive to fake and the most valuable to hold. Buy these last, because they cost real money, and buy them only when the trigger below fires.
RADAR: what to do long before a raise is on the table
Do now, T+0 to T+3. Create the standing folder and deposit this month’s genuine documents at their genuine dates. This dominates every scenario. It is nearly free, it is reversible, and it starts the only clock that matters. If you do one thing from this piece, it is this.
Hedge, by T+14. Turn on and preserve source metadata and audit logs where your tools already keep them: cloud version history, your document system’s log, repository history. This is cheap insurance against a future backdating question you cannot yet see coming. It costs a settings change now and is worth weeks later.
Defer and trigger, T+28 and beyond. The externally dated set, audited accounts and filed returns, costs real money and is not reversible, so do not buy it on a schedule. Pre-commit the observable trigger instead: the first serious investor conversation, or the month your revenue crosses the line where an audit costs less than the discount an investor puts on unaudited numbers. When the trigger fires, buy it. Not before.
CHAIN: what usually happens to a room built under deadline
Match on structure, and the reference class runs wider than startups that raised. It covers any party asked to prove a history it did not record as the history happened: the litigant producing a contract, the applicant with a suspiciously fresh set of references, the trader seeking a bank facility on reconstructed books. Across that class, the manufactured record is caught more often than founders expect, because detection runs on cross-source dates rather than on how good any single document looks.
The base rate favors the reader, and two present-state modifiers push it further their way. Cloud tools now timestamp everything automatically, which makes an honest trail easier to keep and a fake harder to align. African diligence leans hard on founder-investor trust and on local counterparties who can be called, which raises the value of a checkable trail and the cost of a manufactured one.10
Now subtract the counterfactual. The raise that closed on a room assembled in a fortnight did not close because the room was good. It closed because the investor already trusted the founder, and the checking was a formality laid over a decision already made. Do not credit the late room for the outcome. The room did not carry that raise. The relationship did.
Matrix-break flag. If generative tools make it cheap to fabricate a consistent multi-source dated trail, the signal in this whole piece degrades, and the market will move to cryptographic and third-party timestamping that a founder cannot author alone. Watch for it. Until then, the date remains the expensive part, and the expensive part is the one worth holding.
Where the dated trail tells you nothing
The ensemble reads one thing: whether a record was kept as events happened. It is blind to whether the events were any good. A perfectly dated trail of a failing business is a well-documented failure, and the date proves when, not whether. It is also blind to trust, which can make the entire checking exercise moot, as the counterfactual showed. And it cannot see the real reason most founders skip the hour, which is not ignorance but present bias and the belief that later will do. That is the behavioral layer, and naming it changes nothing about the mechanism, which is why it stayed out of the cards.
Here is the one action that survives all of that. You cannot manufacture the past this month. You can start recording it this month. Anyone can produce the document. Only time can produce the date. Open the folder today, drop in the three things that already carry today’s date, and let the year do the expensive work for you. The best data room is the one you were already keeping. Spend the hour.
Sources and notes
- SheetVenture, “How Long Does VC Due Diligence Take?” Venture diligence typically runs 2 to 6 weeks, seed rounds 2 to 3 and Series A and later 4 to 6. sheetventure.com
- Kruze Consulting, “Startup Due Diligence.” Companies in ongoing conversations with investors “typically can get through due diligence in two to four weeks,” while a cold pitch “usually takes a bit longer, but hopefully less than six weeks.” kruzeconsulting.com
- Dealroom, “Legal Due Diligence.” On common red flags: unsigned or expired contracts presented as in force, and old drafts left beside final versions, which create confusion and legal exposure. dealroom.net
- Michael Spence, “Job Market Signaling,” Quarterly Journal of Economics 87, no. 3 (1973): 355 to 374. A signal separates types only when it is more costly for the type one would be mistaken for; otherwise the market pools and discounts it. Author copy (text layer verified). sfu.ca
- “Kingman’s formula,” on the G/G/1 queue: mean waiting time and its variance grow nonlinearly as utilization approaches one and as variability rises, worse than the sum of the two effects near saturation. en.wikipedia.org
- Francis Galton’s 1907 “Vox Populi” ox-weighing count, as recounted in the wisdom-of-crowds literature: 787 entries, median guess 1,207 pounds against a true dressed weight of 1,198, because independent errors cancel and the shared information survives aggregation. pmc.ncbi.nlm.nih.gov
- Lorraine v. Markel American Insurance Co., 241 F.R.D. 534 (D. Md. 2007). Metadata, which records a document’s date of creation and later edits, is a distinctive characteristic usable to authenticate electronically stored information under Rule 901(b)(4). en.wikipedia.org
- Forensic Discovery, “Metadata Matters: The Story Behind Every PDF.” Metadata can reveal a file’s true creation date, and opposing counsel may allege backdating unless timestamps are corroborated with email headers, cloud version history and audit logs. forensicdiscovery.expert
- Founders Factory Africa, “Investor Due Diligence: Tips for African Tech Founders.” African venture diligence relies heavily on mutual trust between founders and investors, which raises the value of a checkable record. foundersfactory.africa