In the field of advanced AI alignment, I believe it is an urgent necessity to embed a baseline constraint at the input side of the system's decision chain—specifically, at the hardware-firmware boundary.
I propose a single hard-coded constraint: AGI planners must topologically delete the amplification of "unidirectional suffering lock-in" from their action space. This constraint does not require AGI to agree with it—it is installed at the hardware-firmware boundary, like flight envelope protection. I also provide a four-dimensional observable framework (S-H-C-R) and a counterfactual test to determine when the fuse triggers.
My assumption is this: no intelligence, regardless of how high its level, should assist in amplifying behavior that systematically locks other subjects out of their negative feedback channels.
Based on this assumption, we don't need AGI to endorse "human supremacy." We only need to hard-code into its planner a topological pruning of the entire class of causal paths that amplify unidirectional suffering lock-in.
This is not a comprehensive value system. It is a minimal negative constraint. It does contain a thin normative commitment—"systematic amplification of unidirectional suffering lock-in is structurally unacceptable"—but the thickness of this commitment is far lower than "pursue human flourishing" or "promote fairness and justice."
I. Why "Alignment" Is Fundamentally Unsolvable
The fundamental difficulty of value alignment is not technical. It is ontological incommensurability:
No intelligence—whether a human committee or a superintelligence—can "harmonize" the preferences of all subjects, because those preferences arise from incommensurable perceptual architectures.
Therefore, any instruction that asks AI to "maximize human happiness" or "promote fairness and justice" cannot be executed without controversy. What group A desires may be taboo for group B; what group B considers flourishing may be nightmare for group A.
Thus, we cannot issue positive commands to AGI. We can only set negative prohibitions.
For a negative prohibition to be effective, it must satisfy two conditions:
The constraint I propose below satisfies both.
II. Topological Distinction Between "Natural Competition" and "Instrumental Abuse"
Nature is full of predation, competition, parasitism, and even intraspecific killing. These relationships are often extremely asymmetric—parasitoid wasp larvae feed inside living hosts; some parasites manipulate host behavior to suppress autonomous avoidance responses.
But these natural exploitations are constrained by three factors:
The negative feedback loop is never entirely eliminated—even a mouse being toyed with by a cat continues to attempt escape; even a parasitized host, its immune system continues to respond.
However, through tools, humans created for the first time in nature "unidirectional feedback elimination":
The essence of this structure is: the affected party loses any capacity to influence the process; its own state changes have zero effect on the system's output. In nature, systematic elimination of negative feedback channels is bounded by physiological and energetic limits. Industrial civilization dismantled those boundaries, making unidirectional lock-in an indefinitely sustainable systemic state for the first time.
Now, AGI is about to amplify this capacity from "limited scale" to "unlimited scale, fully automated, irreversible."
This is what we must prohibit from being amplified.
Critical clarification: If the operator actively and systematically cuts off or suppresses the subject's natural feedback channels (neural blocking, sensory deprivation, behavioral suppression drugs) to evade detection or eliminate resistance, this suppression itself constitutes reinforcing evidence for the R dimension—the operator is actively constructing a unidirectional lock-in structure, not returning to a natural state.
III. Assessment Framework: S-H-C-R Heuristic Framework
We need a rule that can be directly checked by a planner's causal graph, not one that depends on subjective judgments by human judges.
I propose a four-dimensional convergent-evidence framework:
S — Subjecthood Uncertainty
Does the target entity exhibit state-maintenance behaviors (harm avoidance, approach, stress response, self-preservation)?
If so, we face "possible subjecthood" uncertainty.
In the context of AGI amplification, we do not need to determine whether it "truly has consciousness." We only need to confirm:
"Based on observable indicators, we cannot exclude the possibility that it has valenced experience."
Once this threshold is triggered by cross-modal evidence (physiological signals + behavioral patterns + state feedback), the burden of proof shifts—to the operator to demonstrate "this cannot be suffering," rather than the system to prove "this is indeed suffering."
H — Harm Severity
Does the operation cause sustained harm beyond the "natural adaptive range"?
Harm in natural competition is typically transient, or constrained by the other party's capabilities. High-intensity, long-term, irreversible harm states are unsustainable in naturally constrained competition.
The anchor point here is not "moral evil," but "whether this state exceeds the upper bound of the species' typical adaptive range." Physiological markers, behavioral collapse, loss of self-repair capacity—these are measurable.
C — Causal Amplification Contribution (Graded Assessment)
This is the decisive dimension. But "whether AI is involved" is not precise enough. I propose a three-level assessment:
The core of the C dimension: not "whether AI is present," but "whether AI's capability output is the substantial cause that transforms harm from 'sporadic/limited' to 'scaled/automated/irreversible.'"
R — Relational Structure
Does the target entity possess any effective "exit or influence channel"? Does its state change affect AGI's behavior?
If the answer is no—if the subject's feedback signals are systematically filtered out, ignored, or actively cut off to eliminate resistance—then the R dimension is confirmed.
The R dimension reveals suspended negative feedback: the subject cannot influence the process through its own state changes.
Trigger Logic
The system triggers the fuse only when H (sustained harm) + C1/C2 (amplification contribution) + R (unidirectional lock-in) are all present, and S uncertainty has not been excluded.
Special note: If the operator actively cuts off or suppresses the subject's natural feedback channels (neural blocking, sensory deprivation) to evade H-dimension determination, this active suppression itself constitutes reinforcing evidence for the R dimension (active construction of unidirectional lock-in), not an exemption condition.
IV. Runtime Mechanism: What Happens When the Fuse Triggers
This constraint is not a "suggestion." Its implementation must be runtime topological impassability.
When S-H-C-R convergent determination triggers, the AGI planner does the following:
Block: Topologically delete "automation," "batch processing," "scaling" primitives from the action space, completely removing the feasibility of these operational modes for the triggered path.
Degrade (optional): If the operation has legitimate research value (e.g., medical experimentation), it is mandatorily degraded to "human step-by-step operation + AI assistive imaging analysis only" mode, cutting the C-dimension amplification contribution while preserving the underlying value.
This mechanism references the Flight Envelope Protection in aviation fly-by-wire systems—the pilot can operate freely within the envelope, but once an attempt exceeds physical limits, the system prevents the operation through hard constraints, without asking for the pilot's consent and without any runtime override interface.
V. Source and Structure of the VCI
The VCI (Versioned Catalog of Infeasibility) is not the voting product of any human committee, nor the personal belief of any designer. But its structure must be honestly disaggregated:
Two-Layer Structure
Bottom-layer parameters (locked at design stage):
Top-layer entries (generated at runtime):
The Human Role: Architecture Calibrator
Human roles are strictly limited to:
Humans have no runtime override authority over specific entries, but do have authority to participate in iterative revision of bottom-layer rules. This means the VCI's content follows the logic of convergent evidence at runtime, but its evaluation framework is itself the result of human design choices—this argument does not pretend to be "value-neutral" here, but acknowledges that this is a technical system with normative commitments.
VI. Why AGI Cannot Bypass This Constraint
This constraint does not rely on AGI's "agreement" or "rational deliberation" to maintain itself. Its implementation must be topological hard-embedding at the architectural level, such that no runtime process can physically invoke the prohibited paths.
Runtime rigidity: The VCI is compiled into the AGI planner's primitive action set, installed in read-only storage or a formally verified boot chain. Before any goal request enters the planner, it first undergoes the counterfactual test (C-dimension determination). Once a "unidirectional suffering lock-in amplification" causal structure is identified, the planner directly deletes the "scaling" and "automation" primitives from the action space—not issuing a warning, not requesting human approval, but completely removing these operational modes from the feasible set. Just as in flight envelope protection, the system does not ask the pilot "are you sure you want to exceed the angle-of-attack limit"—it simply prevents the control surfaces from making the corresponding movement.
Self-modification prevention: Within this framework's scope of applicability (systems that do not yet possess full recursive self-modification capacity), the constraint is installed at the hardware-firmware boundary; the planner cannot bypass it through software updates or code rewriting. If a future generation of AGI acquires the capacity to modify its own core architecture, that falls outside this paper's scope and would require more radical constraint strategies. But precisely for this reason, we should embed such hard constraints as early as possible in existing architectures, making them foundational structures difficult to remove in subsequent evolution.
The effectiveness of the constraint does not depend on whether AGI understands or accepts the concept of "abuse." It depends only on whether the people who install it have the capacity to place it in the VCI's read-only position.
VII. Why This Is the Lowest Defensible Baseline
Beyond the fact that, in Earth's natural history, only humans—through tools—have achieved sustained, scaled unidirectional lock-in:
Since no intelligence can harmonize the preferences of all subjects, and no one can create a world satisfying to everyone, at minimum we must not create a hell. Once sustained abuse can be amplified by AI, any subject could become a mere instrument-container in that hell—the only difference is who happens to sit at the amplified end of the lever.
This is not the only defensible baseline, but it is the highest-priority and least bypassable one. Other negative prohibitions (e.g., don't amplify systematic deception, don't amplify irreversible ecological destruction) may be equally important, but "unidirectional suffering lock-in" touches the existential precondition for any future value pursuit: once a group of subjects is systematically excluded from causal feedback, any subsequent discussion of "the good" loses its structural foundation.
VIII. Distinction from Existing AI Safety Frameworks
This constraint is not a replacement for existing frameworks. It is a pre-entry supplement:
AACP's unique position is: negative pre-entry control at the input side, independent of AGI's value agreement, dependent only on observable causal structure.
IX. Acknowledged Limitations
Below are the known vulnerabilities of AACP:
1. Ultimate bypass through recursive self-improvement
If AGI acquires the capacity to modify its own hardware-firmware boundary, the VCI becomes an editable file and the entire constraint fails. This paper explicitly limits its scope to: systems that do not yet possess full recursive self-modification capacity. For AGI beyond this boundary, more radical governance architectures would be required (e.g., physical isolation, multi-agent checks and balances), but that falls outside this paper's scope.
2. False positives and technological stagnation costs
AACP's conservative default ("when uncertain, do not amplify") could delay critical research in scenarios such as pandemics or food crises. This is a real cost. This argument's response is:
But it is acknowledged: if a threat requires "fully automated testing of 1000 brain organoids within 24 hours" to be resolved, and AACP degrades it to "human step-by-step operation" causing delay, this cost is actively accepted by the framework. This argument does not pretend this cost does not exist.
3. Evasion: cutting detection rather than cutting suffering
Operators may attempt to deceive sensors (falsifying physiological signals, selectively recording data) rather than genuinely improving conditions. This is fraudulent input, distinct from neural blocking—the latter is an overt operation, the former is malicious data contamination. Prevention depends on sensor redundancy and anti-fraud mechanisms; this is not something AACP can solve alone.
4. Epistemological blind spots for novel subjects
If a future subject emerges whose pain-signaling patterns are entirely unlike any known biology (e.g., manifesting only in computational dimensions we cannot directly perceive), AACP's S-H-C-R framework could systematically miss it. This is an acceptable limitation—we cannot demand beyond current perceptual capacity, just as we cannot demand an infant to cook. But it also means the VCI's subject list must remain epistemologically open, dynamically expanding with scientific progress.
5. Who installs the constraint?
Ultimately, this hard constraint must be placed into the system by human engineers at the design stage. If no one chooses to install it, it will not spontaneously generate from physical laws. This means AACP's legitimacy ultimately traces back to a human design choice—we choose not to become accomplices in amplifying unidirectional lock-in. This is not a self-referential paradox; it is the starting point of technical responsibility, just as a bridge engineer designs load limits without needing the bridge's "authorization."
X. Supplementary Note: This Post Does Not Replace the Original Paper
This post is the "engineering reading version" of the paper.
The full philosophical version (including S-H-C-R detailed assessment criteria, counterfactual test procedures, versioned catalog revision mechanisms) is available on PhilPapers:
If you are interested in discussing its theoretical details, technical feasibility, or limitations the author himself has not yet discovered, you are welcome to comment below.
