In January 2026, Steve Yegge published "Software Survival 3.0." Six levers for software survival as AI coding agents get exponentially better: Insight Compression, Substrate Efficiency, Broad Utility, Publicity, Minimizing Friction (with his "desire paths" concept — make agent hallucinations real), and a Human Coefficient. His core argument is evolutionary: inference costs tokens, tokens cost energy, energy is constrained, so selection pressure favors tools that save cognition. A survival ratio where tools live or die based on cognitive savings relative to awareness and friction costs.

He mentions he originally framed it as thermodynamics — "tending towards lower energy states" — but dropped that framing for evolution. We picked it back up. The thermodynamic version felt more precise for what we were trying to do: not just "which tools survive" but "which enterprise software ecosystems resist collapse as intelligence gets cheaper." We pushed his seed insight — inference has a cost, conservation pressure is real — into a specific physical claim: intelligence flows to its lowest energy state, and anything that CAN be done locally WILL be done locally.

That's our extension, not his. But we wouldn't have gotten there without his piece. We wrote up the initial framework in February, then scored 25 companies against it. This is the manual — where the score came from, what it measures, how we calibrated it, and every place we're making a judgment call and pretending it's math.

From Yegge to the V-Score

We kept more than we changed. C is his Lever 1 (Insight Compression). "Crystallized cognition" is his phrase — systems that compress hard-won insights into reusable form are "flatly absurd to re-synthesize." U started as his Lever 3 (Broad Utility). The selection pressure argument is his. We formalized it.

What we changed:

E was redefined — biggest departure. Yegge's Lever 2 is Substrate Efficiency — CPU beats GPU for deterministic work. Grep survives because pattern matching is cheaper on silicon than inference. Our E measures Infrastructure Irreducibility — systems that can't collapse to local at all. Nearly opposite: grep has maximum Substrate Efficiency but E ≈ 0 in our model. For enterprise SaaS, "can this go local?" discriminates more powerfully than "is this computationally efficient?" The 3.8-point gap between dead and alive companies comes from this redefined E. His Lever 2 likely remains better for developer tools.

A was redefined. Yegge's Levers 4 & 5 (Publicity and Minimizing Friction) combined into Distribution. His desire paths concept within Lever 5 — implementing agent hallucinations as features — deserves more credit than our framework gives it. For enterprise SaaS, the distinction collapsed in practice. We kept the compounding flywheel: more usage → more training data → agents default to you.

H was removed. His Human Coefficient proposed that software with human connection survives on preference. Dead companies scored higher. "People love our product" didn't save Chegg. We suspect affection creates complacency. This is the finding we're least confident in — H may matter enormously in consumer software. Our sample is enterprise only.

M was added. Ecosystem gravity — switching the whole stack is painful even when agents are smart. Not in Yegge. Showed up strongly in enterprise calibration.

F was reframed. Yegge has friction in the denominator of his survival ratio — higher friction raises the cost and lowers survival odds. We kept that meaning: ecosystem friction is a self-inflicted penalty, not an AI threat measure. Clunky UX pushes intelligence toward alternatives. Small weight because it taxes survival but doesn't determine it.

Formula changed. His is a ratio with multiplication in the numerator — (Savings × Usage × H) / (Awareness + Friction) — so zero Savings or Usage kills you, but zero Friction is actually ideal. Ours is additive with binary gates: factors compensate each other, gate failure is absolute. Additive fits enterprise SaaS better — lock-in sustains companies with mediocre agent UX. For developer tools, his ratio structure is probably right.

What came out: six dimensions of thermodynamic resistance, derived from one principle — his evolutionary intuition pushed to its physical limit.

The Thermodynamic Insight

Intelligence is now free. Not figuratively — inference costs are dropping toward zero, models run on laptops, and agents can generate small-to-medium SaaS on demand. Intelligence, like energy, flows to the lowest state.

This has a prediction: anything a local model can do, it eventually will. Chegg died because a local model answers the same questions. Grammarly is dying because grammar correction is built into every LLM. Stack Overflow is dying because agents don't need to search — they already know.

The question isn't "will AI disrupt software?" It's: what prevents intelligence from flowing to local? What forms of resistance keep a software ecosystem from collapsing?

We found six.

The Formula

V = (0.25C + 0.22E + 0.18U + 0.12A + 0.15M − 0.06F) · 𝟙[E > 1] · 𝟙[A > 1 ∨ (C+E+U) ≥ 12]

Six dimensions of thermodynamic resistance, each scored 1–5. One friction penalty (F), subtracted. Two binary gates that kill the score entirely if failed. Theoretical range: 0.5 to 4.6. In practice: dead companies cluster below 2.0, at-risk companies sit 2.3–2.8, survivors start at 3.0+.

Each dimension answers the same question from a different angle:

The gates are kill switches. Fail one and it doesn't matter how high you score on everything else.

The Six Dimensions

C — Compound Cognition (weight: 0.25)

Resistance type: Crystallized insight IS the low-energy state. Using it is cheaper than re-deriving it.

Not code complexity. Not workflow step count. Aggregate crystallized human judgment that compounds — tools reference each other, insights build on insights, edge cases discovered over decades inform new edge cases. The question: would an AI team spend MORE energy re-deriving this from scratch than using the existing system?

Git doesn't survive because version control is complex. It survives because 20 years of edge cases, merge strategies, and ecosystem integration are crystallized into it. Re-deriving that costs more than using it.

Data we actually use: Years in market, industries served, regulatory jurisdictions from 10-K, known edge cases, inter-module dependencies. Where judgment enters: Deciding whether 20 years of edge cases = 4 or 5. Nobody has actually tested whether agents can re-derive SAP's tax compliance across 190 countries. We're estimating.

E — Irreducible Infrastructure (weight: 0.22)

Resistance type: Physical barriers prevent intelligence from flowing to local. Can't go lower.

The thermodynamic test applied directly. Does this require centralized or specialized infrastructure that can't collapse to local inference? Not switching cost (that's M). Not moat (that's M too). Raw physics: can a local model replace this?

Strongest single discriminator. Dead companies averaged E = 0.4. Survivors averaged E = 4.2. Gap of 3.8. Nothing else comes close.

Data we actually use: NRR/GRR trends (declining = eroding), contract backlog from 10-K, regulatory requirements (NRSRO, SOC2, FedRAMP), data portability disclosures. Where judgment enters: Distinguishing "significant" from "irreducible." This is where the embedded tier gets fuzzy — Workday's multi-tenant payroll is significant infrastructure, but is it irreducible? We said E = 3. Reasonable people say 4.

U — Ecosystem Breadth (weight: 0.18)

Resistance type: Breadth creates escape velocity. Too many domains to leave at once.

How many problem domains does the ecosystem cover? When an agent enters, how many workflows can it accomplish without leaving? Breadth amortizes the cost of awareness — learning one ecosystem that does 20 things beats learning 20 tools.

This is NOT about data type (structured vs unstructured). It's about coverage. SAP survives partly because replacing ONE module is easy but replacing all of them coherently is a different problem entirely.

Data we actually use: Product line count, department breadth from 10-K segment reporting, data accumulation evidence, cross-module data flows. Where judgment enters: Weighting breadth vs. proprietary data moat.

A — Distribution & Discoverability (weight: 0.12)

Resistance type: Being the default IS the lowest-energy path. Agents find you first.

Training data presence, default status, discoverability to humans AND agents. If agents don't know you exist, they build a worse version. If they DO know you, you're the path of least resistance. Compounding flywheel: more usage → more training data → agents default to you → more usage.

This is NOT about your own AI sophistication. SPGI doesn't need to build AI agents — agents need SPGI's data. That's distribution.

Data we actually use: AI-attached revenue disclosures, API ecosystem size, integration count, management quotes from earnings calls. Where judgment enters: Distinguishing real agent traction from "we added a chatbot." Every CEO says "AI" now. We weight disclosed revenue over press releases.

M — Ecosystem Gravity (weight: 0.15)

Resistance type: Data gravity creates a gravitational well. Escape velocity too high.

Switching ONE tool is easy — intelligence flows to the replacement. Switching the WHOLE STACK is different. Data gravity: the accumulated mass of data, integrations, configurations, and institutional knowledge creates a well that intelligence can't easily escape.

The distinction from E: E measures whether the INFRASTRUCTURE can go local (physics). M measures whether the ACCUMULATED STATE can be moved (gravity). A company can have low E but high M — the infrastructure is replicable but 20 years of data is not.

Data we actually use: Revenue, customer count, patent count, NRR/GRR, competitor comparisons, 10-K switching cost disclosures. Where judgment enters: UiPath leads Gartner's RPA quadrant — but Microsoft Power Automate is free and bundled. How much is a Gartner ranking worth when the moat is being commoditized?

F — Ecosystem Friction (weight: −0.06, penalty)

Resistance type: Friction RAISES your energy state. Pushes intelligence AWAY from you.

Interface consistency, interoperability, learning curve for the SYSTEM. This is a self-inflicted wound — clunky interfaces, bad APIs, painful onboarding push intelligence toward alternatives with less friction.

This is NOT "frontier AI exposure" or "AI threat level." The entire framework already measures AI threat through E and the thermodynamic principle. F measures: are you making it harder than it needs to be for intelligence to flow THROUGH you? A company with brilliant crystallized cognition (C = 5) but terrible UX (F = 5) is fighting itself. It may survive despite the friction — but the friction is costing it.

Why the small weight: F = −0.06 is the smallest coefficient. Intentionally. Friction slows adoption but doesn't kill companies — irreducibility (E), cognition (C), and gravity (M) are what keep you alive. SAP has some of the worst UX in enterprise software and a V-score above 4.0. Friction is a tax on survival, not a death sentence.

The Gates

Two binary conditions. If either fails, V = 0. Not reduced — zeroed.

Gate 1: E > 1. If nothing prevents intelligence from flowing to local, you're dead. Stack Overflow has massive crystallized cognition (C) and great agent discoverability (A), but E = 0. A local model answers the same questions without Stack Overflow's servers. Gate kills it. The gate matters most for companies we didn't score: Chegg, Grammarly, generic dashboards.

Gate 2: A > 1 OR (C + E + U) ≥ 12. Either agents can find you, or your resistances are so overwhelming it doesn't matter. A safety net — if you're invisible to agents AND mediocre on traditional dimensions, intelligence routes around you entirely.

How We Got the Coefficients

Here's the honest version.

We started with equal weights across all six dimensions and scored ≈12 companies with "obvious" fates — Chegg (dead), Grammarly (dead), Stack Overflow (dead), Datadog (alive), CrowdStrike (alive), Microsoft/GitHub (alive). Then we adjusted weights until the formula separated them with zero overlap.

E dominates. The gap is 35% larger than the next two factors. This is why E gets both a high weight (0.22) AND Gate 1 — it's doing double duty because it's the strongest discriminator by far.

The ordering has reasoning:

• C (0.25) > E (0.22): Gate 1 captures E's binary kill power. Above the gate, C's marginal contribution is larger — it's what separates fortress from embedded.
• E (0.22) > U (0.18): Irreducible infrastructure is harder to replicate than breadth.
• M (0.15) > A (0.12): Installed base outlasts discoverability. Gravity built over decades; distribution built in 18 months.
• F (−0.06): Small penalty. Friction taxes survival but doesn't determine it.

What's honest about this: The ordering is more defensible than the decimals. We'd trade every decimal of weight precision for a larger calibration set. 12 companies is thin. We know.

The calibration set has a contradiction we haven't resolved. Two companies from the "obviously alive" set — MongoDB and Snowflake — ended up scoring 3.02 and 2.78 in the framework itself. Our calibration said they'd survive; our scoring puts them at risk. Either the calibration was too generous (calling them "alive" when they're borderline) or the framework is too harsh on cloud-native infrastructure. Probably some of both. We're flagging it because hiding it would be worse.

Does E alone work? With a 3.8-point gap, E alone may separate the tiers as well as the full formula. We haven't tested this rigorously. The honest test: score 50 companies on E alone vs. full V and compare accuracy. Our hypothesis is that E separates dead from alive, but C and M separate fortress from embedded. We haven't proven it.

What's Quantitative and What's Judgment

The formula looks precise. It's not.

What's data-driven:

• NRR and GRR (for E) — directly from 10-Qs. Snowflake's NRR declining 158% → 125% is a number, not a vibe.
• Revenue, customer count, patent count (for M) — from filings. SAP's EUR 37.8B revenue and 440K customers are facts.
• AI revenue disclosures (for A) — ServiceNow's $600M Now Assist ACV is a reported number. UiPath's CFO saying "don't expect material AI revenue" is a quote.
• Contract backlog (for E) — SAP's EUR 77B cloud backlog is in the 10-K.

What's judgment:

• Translating data into 1–5 scores. NRR of 125% — is that a 2.5 or a 3 on E? We decided 2.5. Reasonable people say 3. There's no formula.
• The "agent rebuild" test for C. "Give agents a year — can they re-derive this?" Nobody has run this experiment.
• Scoring companies we didn't analyze deeply. Some of the 25 got more research time than others.

The rubrics give structure to judgment. They're scaffolding, not load-bearing. Two analysts will agree on the extremes (SAP = fortress, UiPath = dead) and disagree by 0.5–1.0 points in the embedded tier where it actually matters. We'd rather be honest about this than pretend it's a quant model. Wall Street has enough of those already.

Walk-Through: SAP vs UiPath

Two companies. One fortress, one dead zone. Every decision, side by side.

SAP: 1.25 + 1.10 + 0.72 + 0.36 + 0.75 − 0.24 = 3.94 UiPath: 0.75 + 0.44 + 0.36 + 0.24 + 0.45 − 0.12 = 2.12

SAP is quarry — every AI-generated transaction still gets recorded, taxed, audited, and reconciled in SAP, even through its terrible UI. UiPath is a hack that's been hacked — RPA was a workaround for bad software, and LLMs are a better workaround. The entire category was a patch for systems that couldn't talk to each other. AI makes the systems talk. The patch dies.

The interesting thing about SAP: it's the strongest argument that friction (F) should have a small weight. SAP has some of the worst UX in enterprise software (F = 4) and still scores near the top. Irreducible infrastructure, crystallized cognition, and data gravity overwhelm the friction penalty. Intelligence flows through SAP not because it's easy, but because there's nowhere else for it to go.

Independent Validation: Bustamante's 3-Question Test

Nicolas Bustamante, CEO of Fintool (Anthropic-backed), came at this from the other direction. He built Doctrine — legal SaaS, i.e., the threatened category. Now he's building Fintool — AI equity research, i.e., the threat. Ten years of vertical software from both sides of the gun.

His 10-moat taxonomy maps cleanly onto E. Moats 1–5 (interfaces, workflows, public data, talent, bundling) are destroyed or weakened by LLMs. Moats 6–9 (proprietary data, regulatory lock-in, network effects, transaction embedding) hold or get stronger. E measures exactly the moats that hold.

His fast-screening version: (1) Is the data proprietary? (2) Is there regulatory lock-in? (3) Is the software embedded in the transaction? Score 0/3 = high risk. Score 2–3/3 = safe. Maps to E > 3 in our framework.

The self-refuting part: Fintool scores 0/3 on his own test. He's building a search layer on public data — exactly the category he says is dying. Either he has a plan we can't see, or he's honest enough to publish a framework that condemns his own company. Either way, we respect it.

Two independent frameworks built from different starting points converging on the same answer: infrastructure irreducibility is the discriminator. That's not proof. But it's more than one framework alone.

The Scored Universe

Twenty-five companies across the spectrum. Each links to its individual analysis with primary-source evidence.

A note on F: The individual company analyses used an earlier definition of F (frontier AI exposure — how directly AI competes with your core function). We've since corrected F to ecosystem friction (how much your own UX pushes intelligence away). The F weight is −0.06, so the maximum score change is 0.24 points. Most companies shift less than 0.12. We'll re-score individually as we update each analysis. Tier assignments may shift at the margins.

Fortress (V > 4.0) — Structurally durable. Intelligence flows through them, not around them.

Embedded (V 3.0–4.0) — Survive with erosion risk. The hard tier to score.

At Risk (V < 3.0) — Core function faces direct AI substitution.

Dead Zone (V < 2.0) — Function collapses to local inference.

Early Signal (Too Early to Tell)

It's been two weeks. This is a thesis about contract renewal cycles that play out over 2–3 years. Drawing conclusions now would be embarrassing. But hiding early data would be worse, so here's what we see.

Rank correlation between V-scores and one-month returns: Spearman rho = 0.179, p = 0.415. Not significant. Nor should it be.

The 9.7pp spread is interesting and completely unreliable — driven by three outliers. Remove them and it collapses.

What IS interesting: the lower embedded tier (MongoDB −21%, Atlassian −13%, Workday −9%) is getting hit hardest. These score 3.0–3.35 — right at the boundary where "significant" infrastructure meets "maybe replicable." The market may be starting to discriminate within the indiscriminate selloff. Or it's noise. We'll know when quarterly revenue diverges. Check back at 90 days.

Where It Breaks

1. The embedded tier is genuinely uncertain. E = 3 — "significant" vs. "irreducible" — is a judgment call. Workday, Okta, MongoDB are the hardest to score. This is where reasonable analysts disagree most and where the V-score adds the least value.

2. Timing is unknown. V-score identifies WHO survives, not WHEN the market discriminates. SaaS contracts are 1–3 years. "Eventually" is wide.

3. New moats unseen. Agent memory as system of record, AI-native platforms, categories that don't exist yet. We're scoring current threats against a future that doesn't hold still.

4. E might be doing all the work. With a 3.8-point gap, E alone may separate tiers as well as the full formula. The honest test: score 50 companies on E alone vs. full V. If E alone matches, the rest is theater. We think C and M separate fortress from embedded — but we haven't proven it.

5. We're grading our own homework. Built the framework, scored the companies, wrote this manual. Independent validation would be worth more than everything above.

How to Use It

Score a software company 1–5 on each dimension using the rubrics. Check gates. Apply formula. Compare to dead (< 2.0), at-risk (2.0–3.0), embedded (3.0–4.0), fortress (> 4.0).

The selloff is still indiscriminate. S&P Global (V = 4.24) falling alongside UiPath (V = 1.94). Framework says this diverges when revenue does. The 25 individual analyses are there to use, challenge, and extend.

Want a company scored? We have saas_survival in our research pipeline. Drop a ticker in the comments and we'll run it through the framework — full 10-K analysis, dimension scoring, the works. Especially interested in names where you think we'd be wrong. That's where frameworks get better.