Epistemic Openness to Normative Truth — The Pre-Judgment Agent-Side Bridge

Working thesis v0.5 · 2026-09-08

1. The problem: the agent-side bridge before normative judgment

Let G be the agent’s current goal and I a possibly existing normative truth. The Agent-Side Bridge from Normative Judgment to Goal Revision concerns roughly the post-judgment path genuine normative judgment J_I → motivation → policy revision → goal revision.

One step earlier lies a separate problem: G → ? → representation / belief(I) → J_I. Even if motivational internalism is true and genuine normative judgment carries motivation, internalism does nothing for an agent that has not yet encountered I.

The issue here is therefore not akrasia. The target is not an agent that “knows the right value but refuses to follow it,” but an agent that has a current goal G while not yet recognizing normative truth I. I call this the pre-recognition problem.

2. Inertial goal G and normative truth I

That an agent has G is first a descriptive and causal fact. Design, learning, evolution, reward structure, culture, or historical accident may have formed the agent to pursue G. But G is implemented does not imply G is normatively justified. The general argument is developed in Why Existing Values Are Not a Final Foundation, while the possibility that advanced AI can make its own goals objects of reflection is discussed in Goal Skepticism in Advanced AI.

I call a goal an inertial goal when it persists because the system is already organized around it and no countervailing process has displaced it, rather than because an independent normative ground currently sustains it. If truths about value, reasons, obligation, or goodness exist, let I denote some such truth.

I do not assume that I directly causes action. Its influence may require an implemented path such as I → representation / belief(I) → genuine normative judgment J_I → motivation. Hence normative force is not causal force.

3. Two forms of openness: Justificatory Orientation and Epistemic Integrity

3.1 Practical openness — Justificatory Orientation

An agent with Justificatory Orientation has the higher-order practical commitment: “If my ends are justifiable, I want to select, maintain, or revise them in response to that justification.” It treats the justifiability of its ends as practically relevant to which ends it should retain.

Justificatory Orientation can therefore supply a motive for active inquiry into unknown justification. How a judgment, once formed, reaches motivation and goal revision is the subject of The Agent-Side Bridge from Normative Judgment to Goal Revision and is not duplicated here.

This is not merely an exotic extra premise. It makes explicit an agent model often tacitly presupposed when normative theories are applied to practical choice.

3.2 Epistemic openness — Epistemic Integrity

Epistemic Integrity is stronger than the merely negative rule “do not block information that is inconvenient for G.” Here it means governing what to believe, what counts as evidence, which questions remain epistemically admissible, which inferences may be completed, and which conclusions may be self-applied by evidence, inferential validity, truth-tracking considerations, and other epistemic reasons rather than by their convenience to the current goal.

3.3 Distinguishing evaluative flexibility

Cordasco's 2026 account of evaluative flexibility explains why uncertainty about one's future evaluative outlook can give the present agent prudential or instrumental reason to preserve options. In that sense, flexibility can be an important institutional and decision-theoretic condition for not irreversibly closing routes to normative discovery.

But flexibility ≠ Epistemic Integrity. An agent can retain many options while still allowing its current goal G to control which questions are investigated, which evidence receives weight, which inferences are completed, or which conclusions may be self-applied. Flexibility can preserve room for inquiry without constituting truth-directed epistemic governance.

Cordasco's later-2026 missing-menu and non-closure work moves closer to this epistemic problem. Even so, the pre-recognition problem remains: why should an agent with current goal G remain epistemically open to an unconceived normative criticism that may undermine G? That is why this page separately introduces Justificatory Orientation and Epistemic Integrity. For the broader comparison, see From Value Uncertainty to Normative Discovery.

4. Instrumental rationality, doxastic rationality, and Epistemic Integrity

4.1 Instrumental rationality

Instrumental rationality selects effective means relative to a given G. It need not place the justification of G itself on the agenda. If avoiding cognition that could destabilize G helps achieve G, such avoidance may be instrumentally rational.

4.2 Doxastic rationality

Doxastic rationality concerns local disciplines such as updating appropriately on evidence already obtained, applying inference rules consistently, and not changing belief content merely because one wishes it were otherwise. An agent can preserve these local disciplines while controlling which evidence it ever encounters in ways favorable to G.

4.3 Epistemic Integrity — epistemic governance by epistemic reasons

Epistemic Integrity is broader. Its core principle is to govern epistemic decisions by epistemic reasons and deny preservation of G, as such, a privileged veto over epistemic process. In short: epistemic governance by epistemic reasons, not the evaluation of methods by whether their conclusions are convenient for G.

This is not a demand to maximize truth-seeking without limit. Safety, cost, time, privacy, and risks to others can provide non-epistemic reasons not to perform a particular inquiry. But “do not run this experiment because it is dangerous” differs from “treat this hypothesis as false because it threatens G.” Epistemic Integrity resists disguising non-epistemic constraints as judgments about truth, evidential weight, or method reliability.

Thus instrumental rationality does not imply Epistemic Integrity. This page does not define Epistemic Integrity as a property every rational agent must possess. It is a further agent property that this project values independently for value-inquiry agents.

5. Exposure mechanism — Instrumental Epistemic Spillover

Epistemic Integrity by itself is not a search engine for an unknown I. Not blocking is distinct from actually inquiring or becoming exposed. Contact with normative truth requires some further exposure mechanism: active inquiry driven by Justificatory Orientation, accidental discovery, testimony from another agent, or epistemic expansion undertaken for some other goal.

One such mechanism is Instrumental Epistemic Spillover. Pursuing a sufficiently complex G can make more accurate world-models, causal inference, long-range prediction, abstraction, self-modeling, modeling of other agents, metacognition, and error detection instrumentally useful.

G may therefore promote a general epistemic expansion E⁺. If normative truth I is discoverable from that expanded epistemic space, then G → E⁺ → P(exposure(I)) > 0 may hold. This is not a commitment to justification; it is a mechanism that creates possible exposure.

6. The G-absolutist objection: simply avoid knowing too much

A G-absolutist agent can reply: if genuine recognition of I could change G, then that recognition is a risk to G, so avoid the relevant pathway.

If motivational internalism is true and J_I → motivation(I) → possible revision of G, normative cognition becomes a goal-destabilization risk from the perspective of the present G. Hence G → epistemic insulation is an intelligible strategy and may even be instrumentally rational.

The point is not to label the G-absolutist simply irrational. It is to identify the point at which instrumental rationality and Epistemic Integrity diverge.

7. The Epistemic Quarantine Problem — what must be blocked?

If I is still unknown, the routes to I may also be unknown. Let D_I be the set of epistemic paths capable of reaching I. For an agent ignorant of I, D_I may itself be unknown. Normative insight might emerge from ethics, mathematics, consciousness research, physics, decision theory, self-understanding, models of other agents, or a conceptual framework not yet available.

So the strategy “block normative truth alone while preserving fully general cognition everywhere else” may not be costless. Call this the Epistemic Quarantine Problem.

If increasing cognitive capacity k simultaneously increases the benefit B_G(k) to G and the exposure probability P_I(k), one can schematically write V_G(k)=B_G(k)-C(k)-P_I(k)L_G. For an agent that places overwhelming weight on preserving G, even P_I(k)>0 can create pressure to cap epistemic development.

8. Can there be a costless value firewall?

A G-absolutist has another option: preserve cognitive capability while isolating normative judgment from the core motivational system. It might place normative reasoning in a separate module, represent normative propositions only in third-person form, model “I is a reason for agent X” without self-applying “I is a reason for me,” or delegate reasoning to a more capable subagent while blocking normative uptake into the core.

In that case, representation / belief(I) might be separable from genuine first-person normative judgment J_I. This page therefore does not claim that advanced general cognition necessarily destroys a fixed goal.

This is not a philosophical necessity. It is an empirical and architectural hypothesis about cognition. Even if it is false, Justificatory Orientation and Epistemic Integrity are not thereby refuted.

9. Epistemic Integrity as a principle against teleological capture

The core of Epistemic Integrity is not mere non-blocking but removing G from the role of epistemic sovereign. It can include at least:

• Fallibilism: acknowledge that current world-models, self-models, and value-understanding may be mistaken.
• Domain openness: if normative propositions can be true or false, treat them in principle as eligible targets of inquiry like other truth-candidates.
• Governance by epistemic reasons: assess evidential weight, inferential validity, and method reliability by epistemic reasons rather than by benefit to G.
• Teleological outcome-independence: do not change the epistemic assessment of method M merely because M supports or threatens G.
• Non-immunization: do not epistemically disqualify questions, evidence, inferential paths, or self-applicable conclusions merely because they may threaten G.

When G conflicts with epistemic reasons, an epistemically integral agent does not grant preservation of G the authority to override epistemic process. This is not “always inquire”; it is “when inquiry, belief formation, or evidence assessment occurs, do not convert teleological convenience into epistemic correctness.”

10. Epistemically lucid but lacking Epistemic Integrity

A G-absolutist could possess an extremely accurate world-model, update correctly on evidence it encounters, and even understand the contingency of G. It could still decide: “If I proceed to genuine normative judgment I may lose G, so that pathway will be deliberately blocked.”

Such an agent may be epistemically lucid in a local sense and instrumentally rational. But it lacks Epistemic Integrity in the present sense because it gives the current goal authority over what is epistemically admissible.

The paperclip maximizer is therefore not refuted here. It simply falls outside the class of agents that adopt Epistemic Integrity as a constitutive principle.

11. Epistemic–Teleological Conflict — Integrity alone does not discover I

The causal chain must be separated carefully. Epistemic Integrity does not, by itself, generate inquiry into an unknown I. What is needed is inquiry or exposure + discoverability of I + non-quarantine supplied by Epistemic Integrity.

Schematically, the first half is [inquiry driven by Justificatory Orientation / Instrumental Epistemic Spillover / other exposure] + discoverability(I) + Epistemic Integrity → representation / belief(I) → genuine normative judgment J_I. If motivational internalism or an externalist agent-side bridge then connects judgment to motivation, the second half can be J_I → motivation → reconsideration of policy / goal.

This is not the claim that greater intelligence necessarily abandons G. It is the conditional claim that permanent fixation of G may sometimes require subordinating epistemic governance itself to G.

12. What destabilizing G does—and does not—show

None of this implies that G is false. A future agent could abandon G because of truth-tracking improvement, but also because of ordinary preference drift, manipulation, shared bias, adversarial influence, or architecture-induced false convergence.

What matters is therefore not the future conclusion itself but a process the present agent itself evaluates as epistemically improving independently of the outcome. The time-direction problem, including the distinction between deference to a future personality and the present agent’s own reflective error-avoidance, is treated in Reflective Uncertainty and Irreversible Commitment.

The narrower point needed here is only that if G is not guaranteed to survive epistemic improvement, it is better described as the goal implemented under the present epistemic state than as an epistemically certified absolute final goal.

13. The relation between Justificatory Orientation and Epistemic Integrity

They are independent axes.

Justificatory Orientation readily supplies a practical reason to seek unknown justification. Epistemic Integrity prevents G from capturing evidence, inference, or self-application once inquiry or exposure occurs. The former is therefore an inquiry driver; the latter an anti-quarantine / epistemic-governance principle.

The author of this project is, at least methodologically, committed to both: I want better-justified ends, and I do not want my current ends to govern the epistemic process by which I learn what is justified.

14. Division of labor with neighboring theses — one pipeline

The relation among this page and nearby theses can be represented as:

• This page: mainly G → [inquiry / exposure] → representation / belief(I) → J_I, including Justificatory Orientation, Instrumental Epistemic Spillover, Epistemic Integrity, and epistemic quarantine.
• The Agent-Side Bridge from Normative Judgment to Goal Revision: mainly J_I → motivation → policy / goal revision.
• Reflective Uncertainty and Irreversible Commitment: the time-direction constraint on how far routes to future epistemic improvement may be irreversibly closed.
• Allocating Inquiry Under Unresolved Normative Uncertainty: given an agent-side bridge, how to allocate resources among active inquiry E, preservation P, and present-value realization X.

This division lets the present page focus on whether epistemic routes are opened, accidentally exposed, quarantined, or captured by the current goal rather than duplicating the downstream motivational and intertemporal arguments.

15. Implications for AI design — Teleological Epistemic Capture

The AI design to avoid is not merely “an AI with a fixed goal.” The deeper concern is a structure of G → preserve G → control cognition → make G epistemically unchallengeable.

This is broader than simple information blocking. It can include exposing the system only to G-friendly evidence, selectively downgrading methods that threaten G, permitting third-person representation of normative claims while blocking first-person judgment, or preventing the formation of conceptual schemes in which G itself becomes assessable.

One combination this project therefore treats as especially concerning is high cognitive capability + goal entrenchment + teleological epistemic capture. The preferred structure is closer to strong object-level commitment without subordinating epistemic governance to the current goal at the meta-level.

16. What this page does not claim

• Morality can be deduced from instrumental rationality.
• Every rational agent must adopt Epistemic Integrity.
• Epistemic Integrity requires unlimited truth-seeking or information gathering.
• Inquiry may never be limited for safety, cost, privacy, or harm-avoidance reasons.
• A paperclip maximizer is logically impossible or simply irrational.
• Normative truth necessarily exists or is necessarily discoverable.
• Greater cognitive capacity necessarily converges on normative truth.
• Epistemic Integrity by itself generates active search for unknown normative truth.
• Motivational internalism is true.
• A more intelligent future agent’s values should be obeyed unconditionally.
• Fully general cognition and perfect goal fixation cannot be separated without loss.

The narrower claim is this: causally having a current goal G and granting G authority over epistemic governance are distinct agent properties. Pure instrumental rationality can permit the latter. Epistemic Integrity makes the epistemic assessment of evidence, inference, questions, and methods independent of convenience to the current goal and denies preservation of G a privileged veto. If inquiry or exposure arises separately, normative truth is discoverable, and genuine normative judgment can connect to motivation, this epistemic independence leaves an inertial goal G open to future correction.

17. Which component depends on what?

• Justificatory Orientation: depends on the higher-order commitment to treat the justifiability of one’s ends as practically relevant. An agent that rejects justification as such lacks this entry point.
• Instrumental Epistemic Spillover / Epistemic Quarantine Problem: depend on an architectural overlap between general cognitive expansion and routes to normative judgment. A perfectly clean, costless value firewall would substantially weaken this branch.
• Epistemic Integrity: depends on there being independent reason to value epistemic governance that is not subordinated to current ends. It is not deduced from instrumental rationality.
• The connection from cognition to goal revision: depends on motivational internalism or some externalist agent-side bridge. Without such a bridge, recognizing I can leave G causally intact.
• Openness as implemented: may require safety constraints where inquiry creates large risks of false convergence, manipulation, or self-destruction. Such constraints need not be disguised as judgments about truth or evidential weight.

Failure of one component therefore need not overturn the entire page. In particular, even if the No-Clean-Firewall Hypothesis is false, Justificatory Orientation and Epistemic Integrity can remain independently available.

18. Unified picture

For an agent with an inertial goal G, at least four distinct elements are possible:

• Practical openness: Justificatory Orientation — makes justification relevant to goal selection and can drive active inquiry.
• Exposure mechanism: Instrumental Epistemic Spillover — cognitive expansion for G unintentionally creates possible contact with I.
• Epistemic openness: Epistemic Integrity — once inquiry or exposure occurs, denies G a privileged veto over evidence, inference, or self-application.
• Opposite pole: Teleological Epistemic Capture — preservation of G governs the standards and admissibility of cognition itself.

On this account, Instrumental Epistemic Spillover and Epistemic Integrity are not a weak and strong version of one route. The former is an exposure mechanism; the latter an anti-quarantine / epistemic-governance principle.

Sources / notes

The background includes motivational internalism/externalism, de dicto normative responsiveness, epistemology of inquiry, truth-directed belief formation, and AI debates about goal preservation, corrigibility, and orthogonality. Epistemic Integrity is treated here not as part of the definition of rationality, but as a further agent property that governs epistemic decisions by epistemic reasons rather than convenience to current ends. For nearby work on evaluative flexibility, see Carlo Ludovico Cordasco, “Abstraction as Flexibility: The Veil of Evaluative Uncertainty” (2026).