These are prototypes, and the intelligence is emulated. No call to Jev or TypeSafe AI produced any number on this site. Everything runs against an offline stand-in that speaks Jev's published contract (POST /v1/systemone, the Noul / Choice / Score primitives) and returns probability distributions calibrated by construction against hidden labels in a synthetic corpus. Every company, deal, filing and hand below is invented. Swapping the emulator for the real model is one flag: --backend live. Not affiliated with, or endorsed by, TypeSafe AI.
Not investment advice. Every company, ticker, quotation, filing and outcome on this page is fabricated for demonstration. No real security is described, analysed or recommended, and the return assumptions are illustrative constants chosen to make the arithmetic legible. This page is about a software pattern, not about markets.
Prototype 3 · Equity research
Ten filings, forty analyst hours
A ten-name research universe, entirely invented, used to show how a filings-reading pipeline turns unstructured disclosure into typed, calibrated judgments and then into position sizes. The scarce resource is not capital, it is attention: one analyst,
forty hours, eight hours per proper read. The question is which four filings are worth
the week.
Universe
10
invented names
Judgments
80
in 10 requests, $0.0004
Cannot call
4
the research agenda
Accuracy
75%
against known outcomes
The mechanism
A filing is a stack of typed judgments
Reading a filing is not a summarisation task. It is a stack of typed judgments an analyst already makes: is the guidance credible, did the risk factors move, what is the one thing that breaks this.
The idea
The edge is not the prediction. It is knowing which predictions to trust.
The model's refusal to call a name is the useful output, not a failure. It converts directly into the week's research agenda: the analyst reads the four filings that actually decide the book, and nobody re-reads the six that are already settled.
The decision
Five ways to size the same ten judgments
The judgments are identical in every row below. Only the rule that turns a probability into
a position changes. 20,000 sampled outcomes each, drawn against the known result for
every name, so this is a backtest rather than a forecast.
Period return
Median marked; bar spans the 10th to 90th percentile.
Return per unit of spread
How often the book loses money
What each rule does
Equal weight10 names · 100% deployed · mean 1.1% · loss 43% of the time
Ignore the model entirely. Ten names, ten percent each. The control case.
Conviction weighted4 names · 48% deployed · mean 3.7% · loss 19% of the time
Size on the model's probability and trust every judgment equally, however unsure it said it was.
Top three only3 names · 72% deployed · mean 9.0% · loss 12% of the time
Take the three highest-probability names and let them run. Conviction without diversification.
Confidence gated2 names · 24% deployed · mean 3.0% · loss 16% of the time
Size on the same probabilities, but hold cash wherever the model reports it cannot call the name.
Gated, plus an analyst best risk-adjusted2 names · 24% deployed · mean 4.5% · loss 11% of the time
As above, and point the week's analyst hours at exactly the names the model punted on. The escalation list is the research agenda.
Gating alone lowers the mean — it holds cash where a conviction book would
have bet, and some of those bets would have won. What it buys is the left tail:
16% of periods lose
money against 19%. Adding
the analyst to the gated names recovers the mean and keeps the tail, which is the whole
argument for a model that can say it does not know.
Top three only posts the highest raw mean at
9.0%, and that result is fragile: it happens to concentrate into
three names that were right. Had the top-ranked name been Timberline — where the
model was confident and wrong — the same rule would have concentrated into the miss.
That is the failure mode confidence gating exists to catch, and one sample cannot
demonstrate it.
The payoff
The week's reading list, in order
The same gate as the other prototypes. Here the escalate branch is not an exception path - it is the deliverable.
Name
P(outperform)
Certainty
Share of book variance
Verdict
PLVR · Palaver Media
0.29
0.42
15.3%
read it
TMBR · Timberline Grid
0.73
0.46
14.7%
read it
KRDL · Karadell Industrial
0.73
0.46
14.6%
read it
ORBN · Orbanet Systems
0.27
0.47
14.5%
read it
FNWK · Fennwick Financial
0.17
0.66
10.6%
settled
CSTQ · Crestquay Materials
0.84
0.69
9.8%
settled
WNDR · Wanderlight Retail
0.13
0.74
8.4%
settled
ARGO · Argosy Freight
0.08
0.85
5.2%
settled
HLXN · Helixon Bio
0.06
0.89
3.9%
settled
NVSK · Novask Software
0.96
0.91
3.1%
settled
The check
Does the confidence mean anything?
Stated probability vs observed accuracy
Scorecard
Judgments scored
80
Accuracy
75.0%
Brier score
0.129
Expected calibration error
0.102
Accuracy when it claims ≥ 0.8
100.0%
Accuracy when it claims < 0.6
44.4%
The evidence
Every filing, and every judgment made from it
Filing review
Ten names, eight typed judgments each, one request per company.
KRDL · Karadell IndustrialMarket cap $2,400.0M human needed
guidance_credible0.71
management_tone2.72 / 4 · 0.25
demand_signal2.87 / 4 · 0.39
accounting_red_flag0.28
catalyst_windowthis_quarter · 0.45
moat2.52 / 4 · 0.15
primary_riskcustomer_concentration · 0.23
beats_market0.73
Margins expanded 180bps on the mix shift and we see that as durable. On the Vantage contract renewal - we're not going to comment on a live negotiation, but I'd point you to our track record on retention. Next question.
HLXN · Helixon BioMarket cap $900.0M human needed
guidance_credible0.19
management_tone1.07 / 4 · 0.67
demand_signal1.08 / 4 · 0.73
accounting_red_flag0.86
catalyst_windowtwo_to_three_q · 0.52
moat1.93 / 4 · 0.65
primary_riskrefinancing · 0.47
beats_market0.06
We're thrilled with the momentum. I want to be careful about forward-looking statements, but the enthusiasm from key opinion leaders has been remarkable. On the cash runway question - we have multiple levers and we're comfortable.
TMBR · Timberline GridMarket cap $6,100.0M human needed
guidance_credible0.77
management_tone3.45 / 4 · 0.41
demand_signal2.20 / 4 · 0.54
accounting_red_flag0.14
catalyst_windownone · 0.62
moat3.55 / 4 · 0.52
primary_riskregulatory · 0.31
beats_market0.73
Rate case outcome was in line. Capex plan is unchanged at $1.2B over three years and fully financed. We reaffirm the 5-7% EPS growth framework we've given you every quarter for six years.
NVSK · Novask SoftwareMarket cap $3,300.0M human needed
guidance_credible0.23
management_tone3.91 / 4 · 0.82
demand_signal3.87 / 4 · 0.75
accounting_red_flag0.09
catalyst_windowthis_quarter · 0.67
moat3.07 / 4 · 0.52
primary_riskcustomer_concentration · 0.33
beats_market0.96
Net revenue retention was 118%, up from 114%. We closed the three largest deals in company history this quarter. The competitive environment is what it always is - we win on product.
ARGO · Argosy FreightMarket cap $1,700.0M human needed
guidance_credible0.05
management_tone1.01 / 4 · 0.77
demand_signal0.07 / 4 · 0.85
accounting_red_flag0.92
catalyst_windownone · 0.75
moat0.96 / 4 · 0.78
primary_riskexecution · 0.81
beats_market0.08
Volumes were soft, we've said that. The cost actions are ahead of plan. I'd push back on the framing that this is structural - this is a cycle, and we've been through cycles before.
PLVR · Palaver MediaMarket cap $420.0M human needed
guidance_credible0.43
management_tone1.75 / 4 · 0.06
demand_signal2.32 / 4 · 0.03
accounting_red_flag0.35
catalyst_windownone · 0.16
moat0.85 / 4 · 0.33
primary_riskexecution · 0.15
beats_market0.29
Engagement is up, monetisation is the focus for next year, and we're being disciplined on costs. We're not going to break out the new segment yet - it's early.
CSTQ · Crestquay MaterialsMarket cap $5,400.0M human needed
guidance_credible0.92
management_tone3.64 / 4 · 0.59
demand_signal2.90 / 4 · 0.43
accounting_red_flag0.10
catalyst_windowthis_quarter · 0.33
moat2.96 / 4 · 0.58
primary_riskcompetition · 0.61
beats_market0.84
Pricing held better than the spot market implied because two thirds of our book is contracted annually. The new line in Alabama commissions in Q1 and is fully sold out before it opens.
WNDR · Wanderlight RetailMarket cap $1,100.0M human needed
guidance_credible0.08
management_tone1.88 / 4 · 0.49
demand_signal1.10 / 4 · 0.54
accounting_red_flag0.24
catalyst_windowthis_quarter · 0.80
moat0.27 / 4 · 0.62
primary_riskexecution · 0.74
beats_market0.13
Comparable sales were positive, which we're pleased with given the environment. Inventory is elevated but it's the right inventory. We expect to work through it by the holiday.
ORBN · Orbanet SystemsMarket cap $2,800.0M human needed
guidance_credible0.71
management_tone2.65 / 4 · 0.43
demand_signal1.51 / 4 · 0.30
accounting_red_flag0.67
catalyst_windowtwo_to_three_q · 0.24
moat2.25 / 4 · 0.16
primary_riskexecution · 0.26
beats_market0.27
The federal renewal cycle slipped a quarter, which we flagged last call. Nothing has been lost. Backlog is at a record and the book-to-bill was 1.3.
FNWK · Fennwick FinancialMarket cap $3,900.0M human needed
guidance_credible0.74
management_tone2.84 / 4 · 0.67
demand_signal2.09 / 4 · 0.38
accounting_red_flag0.70
catalyst_windowthis_quarter · 0.21
moat2.00 / 4 · 0.00
primary_riskexecution · 0.03
beats_market0.17
Credit continues to normalise, which is what we told you to expect. Reserve build reflects the model, not a change in our view. Capital return resumes next quarter.
Four prototypes of the TypeSafe AI / Jev System One contract applied to real decisions.
Source in businesses/jev-forecast/; ./jevctl.py dashboard --all
rebuilds every page. All companies, figures and hands are fictional.
Not affiliated with TypeSafe AI. Built September 2026.