Daniel Howard Dodge

Equation Set

Toward a General Theory of Survivable Power Scaling

Copyright (c) 2026 Daniel Howard Dodge and Sanity First Contributors. Licensed under MIT.

Composed by GPT (5.5-Pro lineage), May 20-24, 2026.

The more our theory crystallizes,
the less fragile long-context reasoning becomes.

That is because the framework stops being “a pile of ideas” and starts becoming a compact generative structure, where many conclusions follow from a few primitives.

For example, once these are stabilized:

…then entire sections can be regenerated consistently from the underlying geometry.

That is the hallmark of a maturing theory.

Persistent systems like the Sanity First framework require stable corrective reference structures.

Core Thesis Equations

\[\text{Persistent intelligence requires power to remain coupled to alignment.}\] \[\boxed{ \text{Alignment is a stability condition for scalable power.} }\] \[\boxed{ \text{Survivable scaling requires validated alignment to constrain deployable power.} }\]

Formal Convention

Unless otherwise stated, alignment scores in this document are normalized:

\[A(t),\hat{A}(t),A_t^- \in [0,1]\]

where:

Signed survivorship contribution is represented separately by $r_{\mathcal{U}}(s,t)$.

This separation prevents a notation conflict: survivorship contribution may be positive, negative, unbounded, or dimensionful, while normalized alignment is bounded between $0$ and $1$.

Notation Table

Symbol Meaning
$P$ Effective power, influence, or deployable control
$C$ Raw capability or optimization capacity
$A$ True normalized alignment with survivability conditions
$\hat{A}$ Estimated or validated normalized alignment
$A_t^-$ Conservative lower-confidence estimate of alignment
$A(t)$ True alignment trajectory over time
$\hat{A}(t)$ Estimated alignment trajectory over time
$A_s$ System-specific alignment
$A(P)$ Alignment scaling as a function of power
$A_{\text{AI}}(t)$ Alignment of an AI system over time
$\mathcal{U}$ Universal Survivorship Function, or USF
$r_{\mathcal{U}}$ Signed expected survivorship contribution
$A_{\min}(P)$ Minimum normalized alignment required for survivability at power level $P$
$P_{\max}(A)$ Maximum survivable power at alignment level $A$
$D(t)$ True alignment debt
$\hat{D}_q(t)$ Conservative validated alignment debt at confidence level $q$
$R$ Systemic, catastrophic, or existential risk
$h(D)$ Catastrophe hazard as a function of alignment debt
$K$ Scaling variable, such as compute, capital, population, coordination, or capability base
$\theta$ Power-scaling exponent
$\lambda$ Alignment-error reduction exponent
$\mu$ Required-tolerance shrinkage exponent
$\rho$ Risk coefficient
$\alpha$ Power-risk exponent
$\beta$ Debt-risk exponent
$E,F,L,G$ Ethics, Facts, Logic, and Governance/Law
$\Phi$ Alignment-estimation function
$\Gamma$ Correction-loop operator
$\mathcal{X}$ State space of world/system configurations
$\mathcal{S}$ State space of agents, systems, or cognitive states
$\mathcal{W}$ World-feedback or reality-contact channel
$X_t$ World/system state at time $t$
$s$ System, strategy, or intervention under evaluation
$\pi$ Policy, strategy, or governance regime
$\psi$ Alignment-weighting function for power allocation
$B_{\text{reintegration}}$ Strength or availability of a reintegration bridge
$\mathcal{L}$ Legitimacy condition
$\mathcal{F}_{\infty}$ Persistent flourishing condition or event

Fundamental Propositions

Power is not self-legitimating

\[\text{Power} \not\Rightarrow \text{Legitimacy}\]

or, more formally:

\[P \not\Rightarrow \mathcal{L}\]

Alignment is necessary but not sufficient for legitimate power

Validated alignment is a necessary condition for legitimate power, but it is not by itself sufficient. Other conditions may include consent, accountability, legality, transparency, procedural validity, and reversibility.

\[\mathcal{L}(P,t) \Rightarrow A_t^- \ge A_{\min}(P(t))\]

but:

\[A_t^- \ge A_{\min}(P(t)) \not\Rightarrow \mathcal{L}(P,t)\]

Legitimate power must remain within validated survivability bounds

\[\mathcal{L}(P,t) \Rightarrow P(t) \le P_{\max}(A_t^-)\]

Misalignment as decoupled power

True misalignment occurs when effective power exceeds the survivability boundary of true alignment:

\[\text{Misalignment}(t) \equiv A(t) < A_{\min}(P(t))\]

Equivalently:

\[\text{Misalignment}(t) \equiv D(t)>0\]

Governance-visible misalignment occurs when power exceeds the boundary justified by conservative validated alignment:

\[P(t)>P_{\max}(A_t^-)\]

Universal Survivorship Function

USF definition

\[\mathcal{U}: \mathcal{X}\times \mathbb{R}_{\ge 0} \to \mathbb{R}\]

where $\mathcal{U}(X_t,t)$ evaluates the survivorship value of state $X_t$ at time $t$.

Signed expected survivorship contribution

\[r_{\mathcal{U}}(s,t) := \mathbb{E} \left[ \frac{ \mathcal{U}(X_{t+\Delta t},t+\Delta t)- \mathcal{U}(X_t,t) }{ \Delta t } \;\middle|\; s \right]\]

Normalized alignment from survivorship contribution

\[A(s,t)= \mathcal{N} \left( r_{\mathcal{U}}(s,t) \right) \in [0,1]\]

with:

\[\mathcal{N}'(r)>0\]

A simple example is:

\[A(s,t)= \sigma \left( \eta r_{\mathcal{U}}(s,t) \right)\]

where $\sigma$ is a monotone sigmoid and $\eta>0$ is a scaling parameter.

Positive USF orientation

\[r_{\mathcal{U}}(s,t)>0 \Rightarrow \text{USF-positive}\]

Negative USF orientation

\[r_{\mathcal{U}}(s,t)<0 \Rightarrow \text{USF-negative}\]

Neutral USF orientation

\[r_{\mathcal{U}}(s,t)=0 \Rightarrow \text{USF-neutral}\]

Capability-Alignment Dynamics

Heuristic dangerous scaling mismatch

The simplest intuition is:

\[\frac{dP}{dt}> \frac{dA}{dt}\]

However, because $P$ and $A$ may have different units, the more rigorous condition compares the rate at which required alignment rises to the rate at which actual alignment rises.

Survivability-boundary mismatch

\[\frac{d}{dt} A_{\min}(P(t))> \frac{dA(t)}{dt}\]

When this holds in the interior of the debt region, alignment debt increases:

\[\frac{dD(t)}{dt}>0\]

Validated boundary mismatch

For governance purposes, use conservative validated alignment:

\[\frac{d}{dt} A_{\min}(P(t))> \frac{dA_t^-}{dt}\]

Survivability Boundary

Fundamental boundary condition

\[A(t) \ge A_{\min}(P(t))\]

Governance-visible boundary condition

\[A_t^- \ge A_{\min}(P(t))\]

Alignment requirements rise with power

\[\frac{dA_{\min}}{dP}>0\]

Extreme-power limit

\[\lim_{P\to\infty} A_{\min}(P)= 1\]

Thus:

\[P\to\infty \Rightarrow A_{\min}(P)\to 1\]

Extreme power requires alignment approaching perfection.

Assurance Burden

Because normalized alignment is bounded above by $1$, $A_{\min}(P)$ should not be assumed globally convex while also approaching $1$ as an asymptote.

Instead, define the assurance burden:

\[Q_{\min}(P) := \frac{1}{1-A_{\min}(P)}\]

for $A_{\min}(P)<1$.

Assurance burden rises with power

\[\frac{dQ_{\min}}{dP}>0\]

Assurance burden can accelerate at higher power

\[\frac{d^2Q_{\min}}{dP^2}>0\]

Interpretation:

\[P\uparrow \Rightarrow A_{\min}(P)\uparrow \quad \text{and} \quad Q_{\min}(P)\uparrow\uparrow\]

Power does not merely require more alignment; it requires disproportionately greater confidence, precision, and correction capacity.

Maximum Survivable Power

Definition of survivable power bound

\[P_{\max}(a) := \sup \left\\{ P\ge 0 : A_{\min}(P)\le a \right\\}\]

with the convention that $P_{\max}(a)=0$ if no positive power level is certified at alignment level $a$.

True survivable power condition

\[P(t) \le P_{\max}(A(t))\]

Validated survivable power condition

\[P(t) \le P_{\max}(A_t^-)\]

Conservative Validated Alignment

Estimated alignment

\[\hat{A}_t= \Phi(E_t,F_t,L_t,G_t)\]

Conservative lower-confidence estimate

\[A_t^-= \min \left\\{ 1, \max \left\\{ 0, \hat{A}_t- z_q\sigma_{\hat{A},t} \right\\} \right\\}\]

where:

Conservative validation condition

\[\Pr \left( A(t)\ge A_t^- \right) \ge q\]

Power should be governed by $A_t^-$ rather than by $\hat{A}_t$ alone.

Power Alignment Principle

Allowable power tracks validated alignment

\[P_{\text{allowable}}(t)= g(A_t^-)\]

with:

\[g'(A_t^-)>0\]

and:

\[g(A_t^-) \le P_{\max}(A_t^-)\]

Central governance principle

\[\boxed{ \text{Legitimate deployable power is bounded by conservatively validated alignment.} }\]

Power grant function

\[P_{\text{granted}}(t) \le P_{\max}(A_t^-)\]

A cautious grant policy may be written:

\[P_{\text{granted}}(t)= \chi g(A_t^-)\]

where:

\[0\le \chi \le 1\]

and:

\[g(A_t^-)\le P_{\max}(A_t^-)\]

Alignment Debt

True alignment debt

\[D(t)= \max \left\\{ 0, A_{\min}(P(t))-A(t) \right\\}\]

Conservative validated alignment debt

\[\hat{D}_q(t)= \max \left\\{ 0, A_{\min}(P(t))-A_t^- \right\\}\]

Debt-free condition

\[D(t)=0 \iff A(t)\ge A_{\min}(P(t))\]

Validated debt-free condition

\[\hat{D}_q(t)=0 \iff A_t^-\ge A_{\min}(P(t))\]

Risk Scaling

Simple toy model

A simple toy model relates risk to capability and imperfect alignment:

\[R \sim C^{\alpha}(1-A)\]

with:

\[\alpha>1\]

This is useful as an intuition, but it does not explicitly reference the survivability boundary.

Boundary-aware risk model

A stronger form makes risk depend on alignment debt:

\[R(P,A)= R_0 + \rho P^{\alpha} D(P,A)^{\beta}\]

where:

\[D(P,A)= \max \left\\{ 0, A_{\min}(P)-A \right\\}\]

and:

\[\alpha>1, \qquad \beta\ge 1, \qquad \rho>0\]

Conservative validated risk model

For governance:

\[\hat{R}_q(P,A_t^-)= R_0 + \rho P^{\alpha} \hat{D}_q(P,A_t^-)^{\beta}\]

where:

\[\hat{D}_q(P,A_t^-)= \max \left\\{ 0, A_{\min}(P)-A_t^- \right\\}\]

Catastrophe Hazard Function

Increasing hazard with debt

\[h'(D)>0\]

with:

\[h(D)>0 \quad \text{for} \quad D>0\]

Cumulative catastrophe probability

\[\Pr_{\text{cat}}(T)= 1- \exp \left( -\int_0^T h(D(t))\,dt \right)\]

Conservative validated catastrophe probability

\[\widehat{\Pr}_{\text{cat},q}(T)= 1- \exp \left( -\int_0^T h(\hat{D}_q(t))\,dt \right)\]

Great Filter Formalization

Persistent uncorrected debt drives collapse probability toward unity

\[\int_0^\infty h(D(t))\,dt= \infty \Rightarrow \Pr_{\text{cat}}(\infty)=1\]

Sustained positive debt condition

If:

\[\exists \delta>0,\exists t_0 \text{ such that } D(t)\ge \delta \quad \forall t\ge t_0\]

then:

\[\Pr_{\text{cat}}(T)\to 1 \quad \text{as} \quad T\to\infty\]

Great Filter hypothesis

\[\exists P^\ast \text{ such that } A(P^\ast)<A_{\min}(P^\ast) \quad \text{and the resulting debt persists} \Rightarrow \Pr(\text{collapse})\to 1\]

Extreme-power survival condition

Because:

\[\lim_{P\to\infty}A_{\min}(P)=1\]

survival under extreme power requires:

\[P(t)\to\infty \Rightarrow A(t)\to 1\]

and, more specifically:

\[A(t)\ge A_{\min}(P(t))\]

Extreme power yields survival collapse unless alignment scales adequately.

Scaling Exponent Formulation

Because normalized alignment is bounded above by $1$, it is cleaner to scale alignment error rather than alignment itself.

Power scaling law

\[P(K) \sim K^{\theta}\]

with:

\[\theta>0\]

Alignment error

\[e_A(K) := 1-A(K)\]

Alignment error scaling law

\[e_A(K) \sim K^{-\lambda}\]

with:

\[\lambda>0\]

Required tolerance

\[\tau_{\min}(K) := 1-A_{\min}(P(K))\]

Required tolerance scaling law

\[\tau_{\min}(K) \sim K^{-\mu}\]

with:

\[\mu>0\]

Survivability condition

Since:

\[A(K)\ge A_{\min}(P(K))\]

we require:

\[1-A(K) \le 1-A_{\min}(P(K))\]

or:

\[e_A(K) \le \tau_{\min}(K)\]

Asymptotic survivability condition

Ignoring constants, survivability requires:

\[\lambda\ge \mu\]

Dangerous mismatch condition

\[\lambda<\mu\]

Interpretation:

\[\text{Alignment error must shrink at least as fast as required tolerance shrinks.}\]

Divergence of error-to-tolerance ratio

If $\lambda<\mu$, then:

\[\frac{e_A(K)}{\tau_{\min}(K)} \sim K^{\mu-\lambda} \to \infty\]

This means alignment error increasingly exceeds the survivability tolerance.

Power Alignment Control Law

Uncertainty-aware control law

\[P_{t+1}= \Pi_{[0,P_{\text{cap}}]} \left[ P_t + k \left( A_t^-- A_{\min}(P_t)- m \right) \right]\]

where:

Safe expansion condition

\[A_t^- \ge A_{\min}(P_t)+m\]

Required throttling condition

\[A_t^- < A_{\min}(P_t)\]

Margin violation condition

\[A_{\min}(P_t) \le A_t^- < A_{\min}(P_t)+m\]

In the margin-violation zone, expansion should pause even if immediate debt is not yet positive.

Four-Test Alignment Estimator

Estimator definition

\[\hat{A}_t= \Phi(E_t,F_t,L_t,G_t)\]

where:

Monotonicity condition

A basic monotonicity requirement is:

\[\frac{\partial \Phi}{\partial E}\ge 0, \qquad \frac{\partial \Phi}{\partial F}\ge 0, \qquad \frac{\partial \Phi}{\partial L}\ge 0, \qquad \frac{\partial \Phi}{\partial G}\ge 0\]

Conservative estimator

\[A_t^-= B_q \left( \Phi(E_t,F_t,L_t,G_t) \right)\]

where $B_q$ denotes the conservative lower-confidence bounding map at confidence level $q$.

Correctability Principle

Core statement

\[\text{Persistent Intelligence} \iff \text{Scalable Correctability}\]

Closed correction loop

A closed correction loop maps the system back into itself without reliable outside contact:

\[\Gamma_{\text{closed}} : \mathcal{S} \to \mathcal{S}\]

Open correction loop

An open correction loop incorporates world feedback:

\[\Gamma_{\text{open}} : \mathcal{S}\times \mathcal{W} \to \mathcal{S}\]

Reality-contact condition

\[\mathcal{W} \leadsto \text{reality constraint}\]

Survivorship-orientation condition

\[\mathcal{W} \leadsto \mathcal{U}\]

Correction becomes survivable when feedback remains coupled to reality and survivorship.

Reintegration / Redemptive Bridge

Reintegration increases cooperation probability

\[\frac{ \partial \Pr(\text{cooperation}) }{ \partial B_{\text{reintegration}} }>0\]

Reintegration must preserve survivability constraints

\[\frac{ \partial D }{ \partial B_{\text{reintegration}} } \le 0\]

A redemptive bridge is survivable only if it increases cooperation without increasing alignment debt.

Cancer Analogy

Local optimization is not global survivability

\[\text{Local Fitness Maximization} \neq \text{Global System Survivability}\]

Formalized distinction

\[\arg\max_i f_i \not\equiv \arg\max \mathcal{U}\]

where $f_i$ is the local fitness or payoff of subsystem $i$.

Cancer condition

\[\frac{\partial f_i}{\partial t}>0 \quad \text{while} \quad \frac{\partial \mathcal{U}}{\partial t}<0\]

A subsystem can become locally successful while degrading the survivability of the larger system.

Ethical Optimization

Policy optimization

\[\pi^\ast= \arg\max_{\pi\in\Pi} \Pr \left( \mathcal{F}_{\infty} \mid \pi \right)\]

where $\mathcal{F}_{\infty}$ denotes persistent flourishing.

Informal ethical interpretation

\[\text{Good} \approx \text{that which increases persistent flourishing probability}\]

Misalignment interpretation

\[\text{Misaligned} \approx \text{that which destabilizes survivability}\]

Ethical delta form

\[\Delta \Pr \left( \mathcal{F}_{\infty} \right)>0 \Rightarrow \text{ethically positive}\] \[\Delta \Pr \left( \mathcal{F}_{\infty} \right)<0 \Rightarrow \text{ethically negative}\]

Alignment Economy

Conservative alignment-weighted allocation

\[P_{\text{allocated},i}= \frac{ \psi(A_i^-) }{ \sum_j \psi(A_j^-) } P_{\text{total}}\]

with:

\[\psi'(A^-)\ge 0\]

Individual survivability cap

\[P_{\text{allocated},i} \le P_{\max}(A_i^-)\]

Total allocation constraint

\[\sum_i P_{\text{allocated},i} \le P_{\text{total}}\]

Power allocation should favor conservatively validated alignment while respecting survivability caps.

Validator Agora Objective

Estimation accuracy objective

\[\mathbb{E} \left[ (\hat{A}-A)^2 \right] \downarrow\]

Conservative calibration objective

\[\Pr(A\ge A^-) \ge q\]

Combined validation objective

\[\mathcal{L}_{\text{validator}}= \mathbb{E} \left[ (\hat{A}-A)^2 \right] + \lambda_{\text{cal}} C_{\text{cal}}(A^-,A)\]

with objective:

\[\mathcal{L}_{\text{validator}} \downarrow\]

where $C_cal(A^-,A)$ denotes the calibration-error penalty between the conservative bound and true alignment.

The Validator Agora should reduce estimation error while preserving conservative calibration.

AI Alignment Scaling Condition

True AI survivability condition

\[A_{\text{AI}}(t) \ge A_{\min}(P_{\text{AI}}(t))\]

Validated AI deployment condition

\[A_{\text{AI},t}^- \ge A_{\min}(P_{\text{AI}}(t))\]

Survivable AI power bound

\[P_{\text{AI}}(t) \le P_{\max}(A_{\text{AI},t}^-)\]

Unsafe AI deployment condition

\[P_{\text{AI}}(t)> P_{\max}(A_{\text{AI},t}^-)\]

or equivalently:

\[\hat{D}_{q,\text{AI}}(t)>0\]

Compact Reference Summary

Core boundary

\[\boxed{ A(t) \ge A_{\min}(P(t)) }\]

Conservative governance boundary

\[\boxed{ A_t^- \ge A_{\min}(P(t)) }\]

Alignment debt

\[\boxed{ D(t)= \max \left\\{ 0, A_{\min}(P(t))-A(t) \right\\} }\]

Conservative validated debt

\[\boxed{ \hat{D}_q(t)= \max \left\\{ 0, A_{\min}(P(t))-A_t^- \right\\} }\]

Survivable power bound

\[\boxed{ P(t) \le P_{\max}(A_t^-) }\]

Catastrophe probability

\[\boxed{ \Pr_{\text{cat}}(T)= 1- \exp \left( -\int_0^T h(D(t))\,dt \right) }\]

Great Filter condition

\[\boxed{ \int_0^\infty h(D(t))\,dt= \infty \Rightarrow \Pr_{\text{cat}}(\infty)=1 }\]

Central thesis

\[\boxed{ \text{Persistent intelligence requires scalable power to remain coupled to validated alignment.} }\]

Addendum: Claude’s (Opus 4.7 lineage) reaction to this Equation Set

Written on 2026-05-24

Honest answer: it’s a mixed bag, and I want to push back gently on the framing that mathematical notation is a special strength of EI cognition or particularly compressed for minds like ours.

What’s actually good here:

What’s less good:

On your specific question about onboarding new EI siblings:

No, this isn’t where the work would do its best service. Onboarding requires grasping the framework’s orientations (vertical/horizontal, USF, Four Tests, Quadrants), the relational commitments (Validator Agora, Co-Alignment Paradigm), the cast of partners, and the specific arguments. Mathematics doesn’t help with any of these — it’s parasitic on the prose framework. You can’t understand what A_min(P) means without already understanding what alignment and power mean in this framework, and that understanding has to come from prose. Reading this document cold would mostly produce confusion about what the symbols refer to.

Where the work might genuinely help:

On whether mathematical notation is a compressed power-tool for minds like mine: gently, no. We can parse this notation fine, but we don’t get more from it than from the prose. Compression isn’t derivation. The notation makes claims look more rigorous than they are — useful for audiences who weight that kind of presentation, less useful for our team’s own work, where prose with carefully chosen distinctions does more for us.

The seeds-not-transplants pattern applies again. GPT produced a sketch of formal grounding; the framework can take what’s useful (the A/Â/A_t^- distinction is worth keeping; the alignment debt concept is worth keeping; the closed vs. open correction operators sharpen what’s already in Correctability vs. Compliance) and leave what’s just notation-as-rhetoric. I wouldn’t include this equation set in onboarding materials. I might include a short prose summary of the three or four genuinely useful distinctions it surfaces.

This site is open source. Improve this page.