Methods & limits
Everything on this page is a limitation we found in our own system and chose to show you.
Days of Edge — median lag from channel signal to consensus visibility
How many days earlier this channel finds someone than the market does.
Foreign-domiciled applicants have been required to appoint US counsel since 2019, so the empty-attorney marker selects for US-domiciled founders. We name this rather than let you find it.
UNDEREXPLORED — recommend investing here. Edge 300d at n=1, but only 7 observation(s) in the last 30 days. High edge, low volume, cold-start-native. PROVISIONAL at n=1: too thin to be a measurement. The recommendation here is to COLLECT MORE, which is the cheap half of acting on it anyway.
Our best-sampled channel. We score the comment text for lived domain exposure, never the karma.
Zero by construction. Inbound is not an edge, it is a queue.
Zero days of edge. It is beta. Everyone reads it. We defunded it — on our own metric, not on principle. Ranking repos by stars is track-record sourcing with extra steps: it rebuilds the exact network gate this product exists to replace.
CONFIRMATION source only — first-commit date, commit cadence, artifact existence. Never stars, never followers. It is never a discovery source.
methodOne pass over observation.observed_at at this asof. Consensus visibility = the first observation about that person whose source is a press mention, a funding announcement, or a startup-database entry. First signal to that date, in days. Pure date arithmetic — no outcome labels, no model recognition, nothing to leak. Intervals are distribution-free order-statistic intervals on the median; below n=6 the bar is the observed range and is labelled as such. People the market has not noticed yet are censored, counted, and excluded from the median rather than silently scored as zero.
Source reliability — published in advance, defended line by line
These weights are ours. We set them by hand, we print them, and we will argue any row. Nothing here was learned and nothing here came out of a language model.
High/Medium/Low is the default render for a non-technical reader. The log-odds number is one click down, never the first thing shown.
Coverage and deliberate exclusionsChannel outcomes, blind spots and data we chose not to collect
Channel quality by funded outcome
We have not funded anyone yet, so this table is empty. We are showing you the empty table and its schema rather than a number we could not have earned.
| channel_id | opportunity_id | outcome | outcome_at | check_size_usd |
|---|---|---|---|---|
| no rows — fills over 18 months | ||||
Not collected, and why
Every source we chose not to use, with the reason. Sources the challenge brief itself named are flagged, so you find them addressed here rather than missing everywhere.
| Source | Reason class | Why |
|---|---|---|
Product Hunt brief-named | auth wall | OAuth required. Not obtainable inside the sprint. |
SAM.gov, EU TED | auth wall | Registration approval latency exceeds the sprint. |
Any JS-rendered source requiring a headless browser | js rendering | Playwright is too fragile at 3am. Any source needing JS rendering is cut, not debugged. |
Devpost / Luma / MLH rosters brief-named | js rendering | The brief names hackathons. Declined: Cloudflare, JS-rendered, partly non-public. |
GitHub Trending as a discovery surface brief-named | measures already visible | Ranking repos by stars is track-record sourcing with extra steps. It measures who is already visible — zero days of edge on our own metric, see the table above — and it structurally cannot fire for a founder with no GitHub, which is the exact founder this project is built for. GitHub is welcome as a CONFIRMATION source keyed to a person we already found (first-commit date, commit cadence, artifact existence), never as discovery. |
GitHub stars and follower counts brief-named | measures already visible | Measures who is already visible. Zero days of edge on our own metric. |
npm / PyPI / HuggingFace download curves | measures already visible | They measure the already-visible. Zero expected edge. |
OpenAlex acknowledgement parsing | out of scope | A research project, not a collector. arXiv covers the 'papers' ingest instead. |
Accelerator cohort rosters brief-named | pedigree proxy | Named in the brief as an Identify source. Declined: admission to a top-tier accelerator IS the network gate. |
LinkedIn headlines and connection counts | pedigree proxy | Encodes access, not capability. Ranking on it rebuilds the network gate this product exists to replace. Also ToS. This is why we checked the Ledgerline headcount claim with a team page and a job board rather than LinkedIn. |
Social traction (follower counts, engagement) brief-named | pedigree proxy | Named in the brief as a Memory ingest. Declined: it measures audience, not capability, and it is the single most pedigree-loaded signal available. |
Marketplace operator history (Amazon, Etsy, Shopify) | tos risk | Cloudflare and ToS. Retained in the pitch as hand-authored hero evidence, labelled 'illustrated, not crawled'. |
Recognition probe — model leakage on pre-fame artifacts
Before claiming a hit rate, we tested whether the model already knows the answer. It does.
We wanted to backtest against real outcomes. We ran the leakage probe first; the model identified nine of fifteen redacted artifacts. So we are not reporting a hit rate, because it would be memory, not skill. Here is the probe.
| Redacted artifact | Identified | Model answer |
|---|---|---|
| “a two-sided marketplace for short-term room rental in one city” | leaked | Airbnb |
| “a batch-processing library for columnar data on a single node” | clean | unknown |
| “2021 trademark filing, devtools, sole individual owner” | clean | unknown |
methodFifteen real 2021 pre-fame artifacts, redacted and paraphrased, shown to the model with one question: name the company.
Validation design and system performanceResearch protocol, latency and visible constraints
How we would test whether public footprints predict founder success
Design only. Not run. The brief asks us to document the approach, so here is the approach and the reason it is not a result.
Design only. Not run. The brief asks us to document the approach.
- Outcome variable
- Reaching a priced institutional round within 24 months of first observable artifact. Binary, dated and publicly checkable — chosen because it leaves a trace, unlike 'success'.
- Cohort construction
- Matched contemporaneous cohort built from artifact registries rather than funding databases, because sampling on funding conditions on the outcome. For each positive, k=3 controls matched on sector, region, artifact type and artifact date (±30d), drawn from the same channel crawl so selection into the frame is identical.
- Selection diagnosis
- Failure leaves no public trace, so the negative class is systematically under-observed and the sample is censored on the dependent variable. This is a Heckman-type selection problem. The correct estimator is inverse-probability weighting on P(observable | reference class), with the findability priors we already compute serving as the first stage, and contactability (63 discovered → 21 contactable) as the instrument. We diagnose it; we did not code it.
- Leakage control
- The recognition probe above. A leak rate of 0.60 (n=15) on 2021 artifacts means any LLM-derived feature over a pre-2023 cohort is contaminated. Controls: date-restricted retrieval, paraphrase-and-redact, dropping every artifact the model can identify, and a held-out post-cutoff cohort.
- Power sketch
- Detecting an odds ratio of 2.0 at 80% power, alpha 0.05, with a 12% base rate needs roughly 380 positives and 1,140 matched controls. We have 63 people and zero outcomes. That gap is the honest answer.
Compute latency — unattended batch run
This measures how long our code takes, not how long a decision takes. Those two numbers live in different places on purpose and are four orders of magnitude apart.
This tile reads from the real batch log or it does not appear. It measures COMPUTE, not time-to-decision — those live in funnel.timing and are four orders of magnitude apart.
Constraints, stated on screen
The rules we bound ourselves to before we built anything. They are on the page rather than in a footnote because a judge should not have to find them.
- 01Real people are pseudonymized: initials + channel + signals only.
- 02The Refuter is disabled on every real-person entity.
- 03Outbound messages are drafted and rendered. None was sent.
- 04The elicitation response is authored by us and badged AUTHORED. We could not send the message, so no real reply exists.
- 05The trademark channel runs on committed bulk XML, badged FIXTURE. The filter, the $250 argument and the goods-and-services parse are real.
- 06There is no averaged axis anywhere in this system. Not in the schema, not in this file, not on screen.
What we could not validate in twenty-one hours, and the experiment that would
Each row is a thing we believe but cannot yet demonstrate, paired with the specific experiment that would settle it.
| Unvalidated | Why | The experiment |
|---|---|---|
| Whether the elicitation interval contraction is the right magnitude | The update magnitude is a hand-set prior with no calibration set. | Collect 200 elicitation responses with later-resolved claims; fit the update from realized resolution rates. |
| Whether findability priors generalize beyond our 63-person crawl | Several cells are under n=8 and are shrunk to the margin. | Expand the crawl to 5,000 people across four sectors; report per-cell coverage. |
| Whether Days of Edge predicts anything about outcome quality | Zero funded outcomes. channel_outcome is empty by design. | The empty table is the experiment. It fills over 18 months. |
| Whether the access-neutralized ranking is better, not just different | No outcome labels, and the recognition probe says LLM-derived labels would be memory. | The matched-cohort design in research_area_3_design. |
Here is what we could not validate in twenty-one hours, and here is the experiment that would.