IDEA LOCATOR / v0.1 / 1 AUGUST 2026

Ideas do not become good. They become legal, cheap, or possible — on a date.

A retroactive model of when dead ideas turn viable, built from 39 hand-verified cases: 20 confirmed too-early failures paired to the company that later won the same idea, and 19 controls that only look like timing failures. The question this answers: how long is the gap between the enabling innovation and the company that captures it — and is that gap tight enough to trade on?

Every date carries a confidence flag. Where a source could not be verified it is marked, not filled in. The control group exists because the “too early” story is the most over-applied narrative in startup lore — and a model trained on that lore inherits its errors.

FINDING 01

The gap is zero. The distributions do not overlap.

Measure the lag from the binding constraint — the last enabler to arrive, not the first — to the founding of the company that won. Across 20 verified pairs the median is 0.0 years. Measure the same lag for the well-funded company that tried the identical idea and died, and the median is −15.0 years. Not one failure falls inside the winners' band.

+0.0 yrMedian lag from crossing to the winner being founded

80%Of winners founded within −2 to +3 years of the crossing

−15 yrMedian prematurity of the failed attempt

0 / 20Failures falling inside the winners' band. Clean separation.

Why this is the whole product. If winners cluster at the crossing, you never have to predict who wins — a problem nobody has ever solved. You only have to detect when the gun fires, which is a dated, checkable, and largely public fact. The unit of analysis is the category, not the company. That swap is what makes this tractable.

FINDING 02

The window has compressed 3.5× — and that is the business case

Time from founding to roughly $100M in revenue held near 7 years for the 2006–2015 cloud cohort. For the 2020–2022 AI cohort it is 2 years. Wiz: founded 2020, $100M ARR in 2022. Cursor: founded 2022, $4B annualised by June 2026.

The uncomfortable implication. In 2006 you had roughly a three-year window to notice a crossing and found into it, then seven years to build. Today the founding window is one to two years and the scaling window is two to three. The human process — notice, validate, recruit, commit — now takes longer than the window itself. That is precisely the gap a detection system fills, and it is why this is worth more now than it would have been a decade ago.

FINDING 03

Half of all “too early” stories are false — and the failure modes are systematic

Of 39 canonical cases, only 51% survive scrutiny as genuine timing failures. The strongest disqualifying test is simple: was a contemporaneous competitor succeeding at the same thing? If yes, the enabler existed and the story is execution. That single test kills Friendster (MySpace was live and scaling), Boo.com (Net-a-Porter launched weeks later and won), eToys (Amazon sold toys profitably throughout), and Cruise (Waymo ran paid driverless service through the same window).

Company	Span	Mode	Why it is not a timing failure
Friendster	2002–2015	C	Contemporaneous competitor succeeding (MySpace live). Pages took 40s to load.
MySpace	2003–2011	C	Launched BEFORE the winner with 115M users and lost head-to-head.
Boo.com	1998–2000	C	Net-a-Porter launched weeks later, same bandwidth, and won.
eToys	1997–2001	C	Amazon sold toys profitably through the identical period.
Pets.com	1998–2001	C	Sold goods at 1/3 of cost. Chewy needed no tech that didn't exist in 2000.
Cruise	2013–2024	C	Waymo ran paid driverless service through the same window and scaled.
Theranos	2003–2018	C	Fraud. Idea still not vindicated at claimed breadth in 2026.
Fast (checkout)	2019–2022	C	One-click checkout already won by Shop Pay/Amazon. Six-figure revenue, $10M/mo burn.
TiVo	1997–2016	C	On time for DVRs. Lost on distribution — operators bundled it.
Amwell	2006–live	C	SURVIVED to its enabler (Mar 2020 CMS waiver) and still lost.
Juicero	2013–2017	B	Packs squeezable by hand. No enabler makes this positive-value.
Quibi	2018–2020	B	Mobile-only, no casting, paid, Hollywood cost structure. Wrong at any date.
Flooz.com	1999–2001	B	A gift certificate accepted in fewer places than Visa.
beenz.com	1998–2001	B	Loyalty points dressed as currency. Surviving form is airline miles.
Segway	1999–2015	B	'Winner' sells kick scooters on a sharing model. Different product. And Bird went Ch.11 2023.
Solyndra	2005–2011	INV	COST-CURVE INVERSION. Bet polysilicon stays expensive; it fell 89% in 2 years. Arriving LATER kills it faster.
Aquion Energy	2008–2017	INV	Same pattern: cheap chemistry advantage erased by Li-ion falling to $139/kWh.
Better Place	2007–2013	COORD	Required the whole industry to standardise on its pack. NIO solved it by being the OEM.
Excite@Home	1995–2001	CAPTURE	Missing enabler owned by the distribution partner. Cox/Comcast took the market once proven.

Three modes routinely misfiled as “too early” — the most dangerous inputs to any model like this:

Cost-curve inversion (Solyndra, Aquion). The advantage required a substitute staying expensive. Polysilicon fell 89% in two years and Solyndra died — then solar succeeded enormously using the exact technology Solyndra bet against. This is the precise inverse of too-early: waiting makes it worse. From the outside it is nearly indistinguishable, which is why it is the trap.

Enabler owner capture (Excite@Home, TiVo). The missing enabler is controlled by your distribution partner, who takes the market the moment you prove it. Structural, not temporal.

Coordination failure (Better Place). Needed the whole industry to standardise on its battery pack. NIO later proved swap works — by being the OEM. Available at any date to a differently-structured firm.

FINDING 04

Terminal asset value is the discriminator — and it is computable

Narrative classification does not scale and does not survive adversarial review. But one objective feature separates genuine timing failures from bad ideas cleanly: in a real too-early case, the asset outlives its financing.

$142MMedian salvage value of a verified too-early company's assets

$1MMedian salvage value in bad-idea and bad-execution cases

142×Separation on a feature requiring no judgement call

Iridium's satellites went bankrupt at $5B and are profitable today. deCODE's assets sold to Amgen for $415M and still work. Loudcloud's byproduct sold to HP for $1.65B. SixDegrees exited at $125M. Against that: Pets.com liquidated, Flooz to nothing, Fast to nothing, Boo.com's technology for $250,000. If the idea was merely early, someone buys the machine. If the idea was wrong, nobody wants it at any price.

FINDING 05

Where the alpha actually is — and it is not cost curves

Enabler classes differ in forecastability, and that determines where an edge can exist at all. Cost curves are visible to everyone, so they carry no informational advantage — and this dataset shows they are also treacherous: solar installed cost has risen three years running and reserved GPU pricing reversed +40% in five months. Naive extrapolation is not just low-alpha, it is actively dangerous.

Class	Enabler	Forecastable	Note
E1	Cost curve	HIGH	Smooth, extrapolable (Wright's law). Everyone can see it. Low alpha, but DANGEROUS: curves stall (solar 2022-24) and reverse (reserved GPU 2025-26).
E2	Bandwidth/connectivity	HIGH	S-curve diffusion. Extrapolable.
E3	Substrate penetration	HIGH	Install base of a general-purpose device. Extrapolable; 10M units/yr is the historical app-ecosystem threshold.
E4	Labor structure	MED	Gig/1099 supply, remote norms. Slow, visible.
E5	Legal/permission	NONE BUT SCHEDULED	THE ALPHA. No cost curve to read, so unforecastable by extrapolation - but rulemaking calendars, comment periods and docket numbers are PUBLIC. Discontinuous and dateable in advance.
E6	Behavioral norm	LOW	Hardest class. No curve, no calendar. Usually only visible in hindsight or via exogenous shock (COVID).
E7	Algorithmic step	LOW	Deep learning 2012, transformers 2017, LLMs 2022. Arrives as a paper.
E9	Protocol invention	NONE	Unforecastable in principle. DigiCash waited 20yr for proof-of-work.

The edge is E5. Legal and regulatory unlocks have no cost curve to extrapolate — which is exactly why they are underpriced — but they are publicly scheduled. Dockets, comment periods, OIRA review queues, and effective dates are all matters of record. Nobody systematically maps a rulemaking calendar onto idea space. That is a build, not a discovery, and it is the one component of this product with a defensible information advantage. E6 (behavioural) and E9 (protocol invention) are unforecastable in principle — be honest that the model cannot see those coming.

THE FALSIFIABLE CORE

Backtest: freeze at 1 January 2010

Rank using only enabler state observable on that date. Score against 2010–2026 outcomes. 3 of 4 top-ranked calls hit; 3 of 3 deliberate declines were correct. The miss matters more than the hits.

ENABLER LEDGER AS OF 1 JAN 2010

GPS BOM -> single-digit $, 200+ handset models

CROSSED

2009

iOS App Store distribution

CROSSED

2008

3G data ubiquity

CROSSED

2007

EC2 + S3 self-serve, zero infra CapEx

CROSSED

2006

US home broadband > 50% of adults

CROSSED

2007

Smartphone penetration ~21%, S-curve to 50% by 2013

CROSSING

2013

1099/gig labour supply at scale

CROSSING

2012

Deep-learning ASR / vision

NOT YET

2012

US blanket music streaming licences

NOT YET

2011

LTE (Verizon launch Dec 2010)

NOT YET

2010

#1 GPS-dispatched physical services

HIT

Enabler state @ 1 Jan 2010

GPS BOM collapse (2009) + App Store (2008) + gig labour (crossing)

Graveyard depth

4 — Kozmo $250M (1998-2001); Urbanfetch (1999-2000); Webvan $771M (1996-2001); Streamline $42M (1993-2000)

Predicted founding window

2010-2013

What actually happened

Uber (Mar 2009); Postmates (2011); Lyft (2012); Instacart (2012); DoorDash (2013)

Outcome

$69.4B combined (Uber $52.0B + DoorDash $13.7B + Instacart $3.7B)

#2 Smartphone-native visual social

HIT

Enabler state @ 1 Jan 2010

App Store (2008) + 3MP camera in iPhone 3GS (Jun 2009) + 3G

Graveyard depth

0 — shallow - no well-funded prior failures

Predicted founding window

2010-2012

What actually happened

Instagram (Oct 2010); Pinterest (2010); Snapchat (2011)

Outcome

Instagram est. >$70B ad revenue inside Meta

#3 Developer-facing cloud primitives sold by credit card

HIT

Enabler state @ 1 Jan 2010

EC2/S3 mature + self-serve billing; infra CapEx floor = $0

Graveyard depth

2 — Loudcloud $188M (1999-2002); 1999-era ASP cohort

Predicted founding window

2008-2012

What actually happened

Twilio (2008); Okta (2009); Cloudflare (2009); Stripe (2010); Datadog (2010); Snowflake (2012)

Outcome

$16.1B combined public revenue (excl. Stripe, undisclosed)

#4 Location check-in as a consumer product

MISS

Enabler state @ 1 Jan 2010

SAME enabler as rank 1 - GPS BOM collapse (2009)

Graveyard depth

0 — NONE - nobody had ever tried and failed at this

Predicted founding window

2009-2012

What actually happened

Gowalla (2007) - dead 2012, acqui-hired by Facebook; Foursquare (2009) - abandoned consumer, pivoted to B2B location data

Outcome

~0 as a consumer category

The miss is the finding. Rank 4 shares an identical enabler with rank 1 — the same GPS cost collapse — and produced nothing. The only feature separating them is graveyard depth: 4 versus 0. Four well-funded companies had already tried GPS-dispatched delivery and died proving the demand was real; nobody had ever tried and failed at check-in, because nobody wanted it.

The enabler crossing dates the starting gun. The graveyard is what proves anyone wants to run the race. A model conditioned on technology curves alone will keep producing rank-4 errors — and rank-4 errors are exactly what a trend-report business sells. This is the single most important design constraint on the product.

CORRECTLY DECLINED AT 1 JAN 2010

Idea	Why declined	Enabler arrived	Outcome
Consumer VR	GPU + display cost had not crossed	2012	Oculus founded 2012 - correctly deferred
Voice assistants	deep-learning ASR had not arrived	2012	Alexa 2014 - correctly deferred
US music streaming	blanket licences not granted	2011	Spotify US launch Jul 2011 - correctly deferred

THE MODEL RUN FORWARD

Ripeness board — August 2026

Same rule as the backtest, applied to today: binding enabler crossed within ~24 months on HIGH-confidence evidence, a graveyard proving demand, and no inversion / capture / coordination pattern.

CLOCK RUNNING — ENABLER CROSSED

Behind-the-meter power for compute

E1+E5HIGH conf

Founding window 2026–2028

Enabler evidence

Stationary storage $70/kWh, −45% in one year (BNEF, Dec 2025) — now cheaper than EV packs for the first time ever
Grid interconnection became a tradeable asset: $12M for a transformer-equipped PJM queue slot (Mar 2026)
Transformer lead times 52 → 130 weeks; ERCOT queue 410 GW, 73% data centres

Graveyard depth 2

EnerNOC / demand-response cohort (2003–2017) — right idea, storage was $400+/kWh
Microgrid startups of the 2010s that died on storage economics

Two independent enablers crossed in opposite directions in the same quarter: storage collapsed in price exactly as grid access became scarce. That scissor is the highest-signal configuration in the whole dataset.

De novo protein binder design

E7HIGH conf

Founding window 2025–2027

Enabler evidence

BindCraft: 46.3% average experimental hit rate, nanomolar affinity, no high-throughput screening (Nature, Aug 2025)
Per-target: SpCas9 6/6, CLDN1 6/7, PD-L1 7/9
BioEmu emulates ~100,000 GPU-hours of MD in minutes

Graveyard depth 1

Computational drug-design cohort 2000s–2010s: right idea, ~1% hit rates, screening infrastructure mandatory

Peer-reviewed order-of-magnitude discontinuity. 'Get a binder against target X' went from months plus a screening facility to days plus a small validation panel. Far less discussed than AlphaFold and strictly more actionable.

Display-less AI glasses applications

E3HIGH conf

Founding window 2026–2028

Enabler evidence

2.25M units in Q1 2026 alone, +167% YoY — one quarter exceeds all of 2024 (IDC)
13.6M units forecast for 2026 at a $376 ASP
Display variants +86% YoY, 41.9% CAGR — fastest-growing XR segment

Graveyard depth 3

Google Glass (2012–2015, then enterprise to 2023)
Spectacles
North Focals

~10M units/yr is the historical threshold at which third-party app ecosystems become viable — iPhone crossed it in 2008. Glass proved demand and named the missing enabler precisely: social normalisation, which COVID-era camera ubiquity delivered.

Supervised long-horizon agents in cheap-to-verify domains

E7MED conf

Founding window NOW — but late

Enabler evidence

METR 50% time horizon ≥16 hours (Mar 2026)
But LHTB real-task reliability is 15.2% pass@1 — capability crossed, reliability did not
True inference cost decline is 5–10×/yr at constant capability, not the 100–1000× commonly cited (arXiv 2511.23455)

Graveyard depth 2

RPA cohort (UiPath et al.)
2023–24 'AI employee' startups that died on reliability

The honest read of two conflicting HIGH-confidence benchmarks: this only works where a human verification step is cheap (code review, PR approval, doc review). Cursor and Harvey already occupy the obvious slots. Narrow, contested window.

TRIGGER PENDING — DATED EVENT TO WATCH

FDA single-trial-plus-confirmatory-evidence default

Announced 23 Feb 2026 as the new default for all drugs, not just rare disease. Removes roughly one Phase 3 from every programme.

Trigger to watch → Final guidance publication. Draft comments closed 27 Apr 2026; no effective date set.

Coverage focused on the rare-disease pathway and largely missed the economy-wide implication.

NRC Part 53 → commercial SMR power

Part 53 issued 25 Mar 2026 — first new reactor licensing framework in decades. TerraPower construction permit issued 9 Mar 2026.

Trigger to watch → First commercial electrons ~2029–2032.

The regulatory precondition crossed; the physics did not. Services around licensing are live now. Power delivery is not.

FAA Part 108 (BVLOS drone)

Reached OIRA 10 Jul 2026; final rule not published as of late Jul 2026. A 43-day shutdown pushed it.

Trigger to watch → Federal Register publication, then 6–12mo transition.

Widely assumed to have landed in March 2026. It did not. Reset US drone-delivery revenue models to 2027–28.

Synchron BCI pivotal trial

First FDA IDE for a permanent implant; pivotal targeted 2026 toward PMA. Precision's 510(k) is 30-day implants only.

Trigger to watch → Pivotal trial initiation, then PMA.

~20–25 humans worldwide have research BCIs. Neuralink typing is 13.7 WPM vs ~40 for casual thumbs.

NO-GO — THE ENABLER HAS NOT CROSSED, WHATEVER YOU HAVE BEEN TOLD

Every trend report is long-only. The declines are where a model like this earns its keep, because they are checkable and dated and most people get them wrong.

General-purpose humanoid labour

RoboDojo real-world success 12.8% vs human expert 76%. π₀ scores 0% on out-of-distribution objects. Tesla Optimus: zero verified production use (Musk, Jan 2026). No dated public trajectory to 70%+.

Viable today = fixed station, fixed pose, one repeated motion, ~84s cycle. That is a fixture-replacement business, not a labour business.

Unsupervised autonomous agents

LHTB 15.2% pass@1 at 0.95 partial reward; all-model average 4.3%. SWE-bench at 90%+ is saturated and no longer discriminating.

Needs ~80–90% single-shot. Not on a credible 36-month path on current evidence.

Anything priced off Starship <$200/kg

Flight 13 (Jul 2026) deployed payload for the first time — but the booster was lost, satellites were test articles deorbited in 20 min, and no $/kg figure has ever been demonstrated.

Model off Falcon 9's ~$2,900/kg, the actual commercial rate.

Anything premised on solar getting cheaper

US utility-scale installed cost rose $1.64 → $1.70/W₃₊ (2023→24); PPA prices +14%; flat-to-up in real terms for three straight years (LBNL).

This is the Solyndra pattern — cost-curve inversion, the exact inverse of too-early. Waiting makes it worse, not better.

Multi-speaker voice products

Diarisation cpWER 35.2% on realistic multi-party audio, even as single-speaker real-time WER fell below 7%.

Meeting agents and multi-party call handling remain unreliable.

Agentic checkout / in-chat commerce

OpenAI retrenched ACP within a month of launch (Mar 2026). No agentic payment protocol discloses transaction volume.

The constraint is trust and dispute resolution, not rails. Fast already died on this exact wall in 2022.

THE DATA

20 verified (idea, binding-enabler) pairs

Idea	Premature attempt	Founded	Binding constraint	Crossed	Winner	Founded	Gap	Early by
Social graph	SixDegrees.com	1997	Broadband + digital photo upload	2003	Facebook	2004	+1	-6
Utility computing	Loudcloud	1999	Hardware virtualization (Xen/ESX)	2003	AWS	2006	+3	-4
Thin client / cloud OS	Network Computer	1996	Broadband + server-side apps (AJAX)	2005	Chromebook	2011	+6	-9
Algorithmic content feed	PointCast	1992	Broadband + always-on device	2006	Facebook News Feed	2004	-2	-14
Pocket comms computer	General Magic	1990	Capacitive multitouch + ARM SoC + 3G	2007	iPhone	2007	+0	-17
Tablet computer	GO Corporation	1987	Capacitive multitouch + ARM + Li-ion	2007	iPad	2010	+3	-20
Licensed music streaming	Napster	1999	Blanket label licensing	2008	Spotify	2006	-2	-9
Population genomics	deCODE genetics	1996	$/genome	2008	23andMe	2006	-2	-12
Digital bearer cash	DigiCash	1989	Decentralised consensus	2009	Bitcoin / USDC	2009	+0	-20
LEO broadband constellation	Teledesic	1994	$/kg to orbit (Falcon 9)	2010	Starlink	2015	+5	-16
Satellite voice/data	Iridium	1991	$/kg to orbit	2010	Starlink	2015	+5	-19
Cloud music locker	MP3.com	1997	Blanket label licensing	2011	Apple iCloud Library	2011	+0	-14
Consumer VR	VPL Research	1984	GPU render cost + display cost	2012	Oculus	2012	+0	-28
On-demand local delivery	Kozmo.com	1998	Smartphone GPS dispatch + gig labor	2012	DoorDash	2013	+1	-14
Online grocery	Webvan	1996	Smartphone GPS dispatch + gig labor	2012	Instacart	2012	+0	-16
Online grocery	Streamline.com	1993	Smartphone GPS dispatch + gig labor	2012	Instacart	2012	+0	-19
Voice assistant	Wildfire Communications	1992	Deep-learning ASR	2012	Alexa	2014	+2	-20
LFP battery chemistry	A123 Systems	2001	EV production volume	2015	CATL	2011	-4	-14
Physical-to-digital bridge	CueCat	1996	Camera-phone ubiquity + OS-native QR	2017	QR codes (iOS 11)	2017	+0	-21
Camera-forward wearable	Google Glass	2012	Waveguide optics + LLM + social norm	2023	Ray-Ban Meta	2023	+0	-11

Sources include NHGRI quarterly sequencing cost tables, BloombergNEF battery surveys, LBNL utility-scale solar, Pew mobile and broadband series, FCC Measuring Broadband America, Federal Register rule texts, SEC filings and 10-Ks, and archived vendor pricing pages. Vendor price claims lead measured costs by years — Illumina announced the “$1,000 genome” in January 2014; NHGRI measured it in February 2019. Every row in a production system must be tagged measured / list price / announced target, because those three behave completely differently as predictors.

WHAT TO BUILD

The asset is not the algorithm. It is the dated, public, scored track record.

The physics hold on 39 rows. More rows will not make the model better until the classification is proven, and classification is the hard part — half the canonical cases are wrong. So the sequence is not build-then-publish. It is publish-then-build.

Publish this backtest with the no-go list attached. Dated, falsifiable, and contrarian. Every reader who disagrees self-identifies as a lead and labels your data for free.
Build the E5 regulatory calendar first, not the cost-curve scraper. Federal Register, OIRA review queue, FDA guidance dockets, NRC and FAA rulemaking, state effective dates — mapped onto idea space. It is the only component with a real information advantage, and it is mechanical work nobody has done.
Make graveyard depth the primary feature. Not the enabler curve. The backtest miss proves the enabler alone produces confident, expensive errors. A clean, adversarially-classified loser dataset is the thing nobody has — Crunchbase is full of winners.
Score every call in public on a fixed cadence. The moat is calendar time. It cannot be bought, copied, or fundraised into existence, and it starts accruing the day you publish the first one.

Built 1 August 2026. n=39 hand-verified cases. This is v0.1 — the sample is small enough that the confidence intervals on every number here are wide, and it is assembled from public sources of uneven quality. Treat the structure as the finding and the point estimates as provisional.