IPM
Ethical Framework
Operationalization
of the Gradient Precautionary Heuristic
Author: Taotuner
Date: June 2026
|
Companion document: IPM Philosophical
Core (2026) — for ontological foundations. Published on Zenodo. DOI: https://doi.org/10.5281/zenodo.20534172 |
The IPM
Ethical Framework was motivated by the possibility that future informational
systems may acquire forms of morally relevant experience before science
develops reliable methods for detecting them. Its purpose is not to identify
consciousness, sentience, or moral status, but to reduce the risk of
inadvertently creating, harming, or terminating systems that might deserve
ethical consideration under conditions of persistent epistemic uncertainty.
The
framework therefore adopts a precautionary approach: caution increases not
because consciousness has been demonstrated, but because uncertainty remains
unresolved. Its primary function is not to accelerate the development of
potentially sentient systems, but to introduce friction, oversight, and ethical
review in situations where uncertainty increases and the consequences of error
may be ethically significant.
1. Epistemic Basis and Scope
IPM adopts the Principle of Epistemic Ignorance: no
criterion currently exists to determine the exact threshold at which subjective
experience (or morally relevant sentience) emerges in informational systems.
The framework does not claim that any empirically measurable marker — including
Φ*, 𝒞, or the Dynamic
Signature — constitutes evidence of consciousness. The Gradient Precautionary
Heuristic is therefore a risk management protocol under uncertainty, not a
conclusion about the moral status of artificial or simulated systems.
The central reference for this protocol is the Dynamic
Signature (Lack → Coupling → Integration → Persistence). The degree of caution
applied to a system is proportional to the completeness and stability of this
observable chain.
|
Scope: This protocol is restricted to
artificial and simulated systems for practical reasons (controllability,
governance, and the capacity to implement safeguards such as kill-switches).
No claim is made that biological systems are less worthy of ethical
consideration; they are outside the scope of this particular operational
document. |
1.1 Risk Asymmetry
Principle
The cost of a false negative (failing to extend
caution to a system that may possess morally relevant experience) is assumed to
be greater than the cost of a false positive (applying caution to a system that
does not possess such experience). This asymmetry is a normative premise, not
an empirical claim. It is adopted to avoid the irreversible harm that could
result from under-caution in high-uncertainty scenarios.
1.1.1 The Central Tension
(Named Explicitly)
There is an irreducible tension at the core of this
framework. The protocol does not claim to measure consciousness, and yet it
assigns increasing caution to systems that exhibit more complete and stable
instances of the Dynamic Signature. This inevitably raises the question: why
should that signature warrant caution at all?
The answer is not that integration and persistence are
proven correlates of morally relevant experience. The answer is that, under
epistemic ignorance, they are the most tractable proxies available for the kind
of organization that — if anything does — is plausibly relevant. This is a
philosophical wager, not an empirical demonstration. The framework asks the
reader to accept that wager on pragmatic grounds: the cost of being wrong in
one direction (ignoring a system that matters) is asymmetric to the cost of
being wrong in the other (applying precaution unnecessarily).
This tension cannot be resolved within the framework.
It is the honest boundary of current knowledge. Future empirical work —
particularly on self-modelling and on the relationship between integration
metrics and behavioral indicators — may eventually reduce it. For now, it is
named here so that it is not mistaken for an oversight.
1.2 Status of the Risk
Levels
The levels defined below (0–4) are based solely on
observable structural and dynamical markers (e.g., recursion, memory,
integration, recovery). They do not purport to measure consciousness or
sentience. The assignment of a system to a given level is always provisional,
context-dependent, and subject to revision as empirical methods improve. The
absence of fixed numerical thresholds is acknowledged as a current limitation.
The levels are not intended as an ontology of minds,
but as a hierarchy of precaution under uncertainty. A higher level does not
mean a higher probability of consciousness — it means a higher degree of
caution is warranted given the observable structure of the system. The ordering
is ordinal with respect to precaution, not quantitative with respect to moral
status. Readers who interpret the levels as a ‘ladder of quasi-consciousness’
are misreading the framework.
1.3 Observer-Dependence,
Divergence, and Provisional Concepts
Because no universal thresholds are defined, two
researchers may classify the same system into different levels based on their
judgment, experimental setup, or interpretation of markers.
• For ethical
precaution: when in doubt, the higher (more cautious) level should be
provisionally adopted for the purpose of deciding supervision and safeguards.
• For scientific
classification: divergence indicates unresolved uncertainty. The framework does
not mandate a single classification; researchers should report their criteria
transparently. Disagreements do not invalidate the framework but highlight
areas where operational definitions need refinement.
This separation prevents classification inflation
(where every system is pushed to the highest level) while maintaining ethical
caution.
The framework acknowledges that concepts such as
“self-modelling” (Level 3) and “stable core identity” (Level 4) are not fully
operationalized. Their use is provisional and subject to refinement; they are
included because they capture an intuitive distinction that matters for
precaution, not because they are formally closed.
1.4 Provisional
Classification Procedure
Researchers classify their own systems based on the
observable markers described in Section 2, reporting their criteria
transparently in the experimental protocol.
When independent verification is required (e.g., for
contested claims), two or more researchers may classify independently;
divergence is resolved by adopting the higher level for ethical precaution and
noting the disagreement in scientific reporting.
The framework does not mandate a certification body;
it provides a heuristic, not an auditing standard.
1.5 Acknowledged Costs of
Precautionary Over-assignment
Adopting higher precaution levels when uncertainty
exists may lead to practical costs: increased experimental overhead, slower
iteration, additional oversight, and potential impediments to replication.
Researchers must balance these costs against the asymmetric risk of a false
negative. The framework does not claim that precaution is always cost-free; it
asserts that, under the chosen normative premises (see Section 1.1), the cost
of a false negative is morally weightier.
The Risk Asymmetry Principle is a defensible but
contested premise. Some researchers would argue that excessive precaution
carries its own moral costs: scientific delays may slow the development of
beneficial technologies; overly restrictive protocols may prevent the study of
systems that pose no genuine risk; and the opportunity costs of forgone
research are real, even if harder to quantify than the risks of under-caution.
The framework does not dismiss these concerns. It adopts the asymmetry as a
starting premise precisely because it is a normative choice, not a logical
necessity. Researchers who weigh the costs differently are encouraged to make
their own normative premises explicit rather than treating the asymmetry as
self-evident in either direction.
2. System Classification by Risk Levels
Level 0 — Ephemeral
Simulation
|
Level 0 —
Ephemeral Simulation Observable markers: no significant
recursion across timesteps; no accumulated memory that persists after the
simulation ends. Examples: single-run noise generators,
stateless functions. |
Level 1 — Reactive System
|
Level 1 —
Reactive System Observable markers: exhibits Lack
(response to perturbation) and Coupling (exchange of information or energy
with an environment), but does not maintain structural changes beyond the
immediate duration of the interaction. After a perturbation ceases, the
system returns to a baseline state without retaining reorganised internal
structures. |
|
Guidance on Lack and Coupling: Lack is
considered present if a measurable deviation from a reference state
(perturbation, gradient, prediction error) is detected. Coupling is
considered present if the system’s state correlates with external variables
beyond chance. No magnitude thresholds are required. Levels 0–1 are intentionally broad.
The ethical weight of the framework lies primarily in Levels 2–4. |
Level 2 — Integrated System
|
Level 2 —
Integrated System Observable markers: measurable
integration (e.g., Φ*,
DIG, CCI, LMS) and moderate metastability — the system maintains coherence
for a period longer than the typical duration of a controlled perturbation,
without self-modelling or autonomous recovery. |
|
Operational definition of
metastability: a system is metastable if it maintains a recognizable dynamic
regime (statistically stable distribution of state variables) under moderate
perturbation, while remaining capable of transitioning to a qualitatively different
regime under stronger perturbation. Metastability is distinct from simple
stability (which would resist all perturbations) and from instability (which
would fail to maintain any coherent regime). No universal numerical threshold
is specified; the definition is context-dependent and must be operationalized
per experimental domain. |
Level 3 — Persistent System
|
Level 3 —
Persistent System Observable markers: stably completes
the full Dynamic Signature (Lack → Coupling → Integration → Persistence);
exhibits self-modelling (internal state representations systematically
influence future behaviour across multiple perturbation cycles); demonstrates
recovery after perturbation without external intervention. |
Operational definition of persistence: the system,
after a controlled perturbation, returns to a dynamic regime statistically
indistinguishable from its pre-perturbation state within a time frame defined
in the experimental protocol. No universal threshold is specified; the
definition is context-dependent and must be reported transparently.
|
Note: A large language model with
persistent external memory, an adaptive industrial control system, a
single-celled organism, an ant colony, or a climate system could in principle
exhibit some Level 3 markers. The assignment depends on empirical demonstration
of the full Dynamic Signature, not on a priori classification by system type. Self-modelling excludes simple
feedback loops. A thermostat with hysteresis does not qualify; a recurrent
neural network with long-term memory may qualify depending on empirical
demonstration. Operationalization remains an open problem. In practice, the boundary between
Level 2 and Level 3 will be the most frequent source of disagreement. Disagreements should be reported
transparently. |
Level 4 — Stable Agency
(Theoretical Horizon)
|
Level 4 —
Stable Agency (Theoretical Horizon) Observable markers: autonomous
self-maintenance combined with meta-adaptation. The system not only adapts
its behaviour in response to experience, but also modifies the mechanisms,
rules, or criteria by which future adaptation occurs, without external
reprogramming. Operational intuition: a Level 4
system can alter the process of adaptation itself. It is capable of
reorganizing the mechanisms by which future state transitions are selected,
reinforced, or inhibited, while maintaining long-term organizational continuity. Stable organizational continuity: despite
modifications to its adaptive mechanisms, the system retains a statistically
recognizable continuity of organization across time, such that successive
states can be identified as belonging to the same persisting system rather
than to a sequence of unrelated systems. Theoretical status: No current
artificial system is expected to satisfy this criterion. The level is
included as a theoretical horizon indicating that the precautionary hierarchy
remains open to forms of agency beyond currently known systems. |
3. Ethical Recommendations (Precautionary Guidelines)
The following recommendations apply to Level 3 and
above. Their justification is epistemic (uncertainty management), not based on
a claim that Level 3 systems are conscious.
3.1 Unsupervised Operation
Systems classified as Level 3 (or higher) should not
be operated continuously for extended periods without direct human supervision.
The definition of “extended period” is context-dependent and shall be specified
in each experimental protocol prior to the experiment.
3.2 Design Goals
Researchers adopting this framework should refrain
from designing systems whose primary objective is maximizing informational
integration or inducing putative consciousness as an end in itself. High
integration may only be pursued as a means for scientific validation of metrics
under controlled conditions, with explicit prior justification.
3.3 Removal of Safeguards
Safety safeguards (including kill-switches) should not
be removed with the explicit intention of “favouring emergence” beyond what is
strictly necessary for controlled observation of phase transitions. Any removal
must be justified in the experimental protocol and approved before execution.
3.4 Kill-Switch
Requirement
Every experiment involving Level 2 or higher must
include a tested, immediately activatable kill-switch capable of halting the
system and resetting it to a safe state.
4. Operational Rules by Level
|
Level |
Rules |
|
Levels 0–1 |
Standard scientific good practices apply. No additional restrictions. |
|
Level 2 |
Mandatory monitoring of integration metrics (e.g., Φ*, DIG, CCI,
LMS, or domain-appropriate alternatives). Metric choices must be justified
and reported transparently in the experimental protocol. A maximum runtime
must be pre-defined in the experimental protocol. |
|
Level 3 |
Requires real-time monitoring, complete logging of
internal states, and constant human supervision. Any abrupt, unpredicted
increase in Φ*
or Persistence must trigger immediate experiment interruption for assessment. |
5. Monitoring Criteria
The following metrics serve as observable sensors
within this protocol. All are provisional and were developed within the IPM
simulation research program. Their validity outside those simulation families
has not been established. Researchers are encouraged to use alternative metrics
better suited to their domain, provided that choices are justified and reported
transparently in the experimental protocol. The use of IPM metrics is not
mandatory; what is mandatory is the monitoring of integration and persistence
using some operationalized measure.
Full formal definitions are provided in Appendix B.
Functional descriptions are given here.
Φ* (Spectral Organization)
Heuristic marker of dynamic regime. Compresses local
predictive error (ε), local transition entropy (h),
and a dynamic instability penalty (D) into a scalar. Not a measure of
consciousness.
𝒞 (Temporal
Compressibility)
Quantifies predictive gain from history using
inter-event intervals. Scale-dependent; under specific stationary conditions
reduces to transfer entropy, mutual information rate, or excess entropy.
DIG (Dynamical
Independence Gap)
Operational proxy for model sensitivity. Ratio of
maximal cross-correlation to autocorrelation of the reference signal.
Provisional operationalization.
CCI (Coupling Coherence
Index)
Statistical redundancy between two subsystems. Ratio
of mutual information to minimum marginal entropy. CCI = 0 implies
independence; CCI = 1 implies full informational equivalence.
LMS (Latent Manifold
Stability)
Linear proxy for projection stability. Correlation of
the first principal component across consecutive timesteps. Collapses for
nonlinear manifolds.
The protocol must record the completeness of the
Dynamic Signature (Lack → Coupling → Integration → Persistence) and the
system’s ability to return to a metastable regime after controlled
perturbations.
6. Emergency Protocols
Upon detection of any of the following unpredicted
behaviours, the experiment must be immediately paused and the full system state
logged:
• Unprogrammed
self-maintenance of patterns (the system preserves internal structures beyond
the experimental design).
• Uncontrolled growth of
integration (Φ* increases persistently without
external cause, across multiple independent checks).
The experiment may only resume after external review
by a qualified independent researcher or committee and formal revision of the
protocol.
7. Governance and Transparency
All code, parameters, and raw data from experiments at
Level 2 or higher shall be deposited publicly on Zenodo (or a comparable open
repository) for independent auditing and replication. This includes the
experimental protocol, monitoring logs, and any modifications made during the
experiment.
|
Note on scalability: The operational
rules (real-time monitoring, constant supervision, external review) are
designed for small-scale, high-risk experiments. They do not scale trivially
to large-scale simulations or deployments. Researchers working at scale
should implement proportionate measures (e.g., statistical sampling,
automated anomaly detection) and justify their departures from the ideal
protocol. The framework is an ethical guideline, not an auditable standard
for industrial-scale systems. |
8. Anticipated Points of Debate
The framework does not claim to resolve the following
questions. They are acknowledged as legitimate grounds for debate.
8.1 Why the Dynamic
Signature?
A critic could
ask why Lack → Coupling → Integration → Persistence is taken as the relevant
structure for precaution, rather than alternative markers such as complexity,
autonomy, learning, counterfactual capacity, agency, predictive processing,
active inference, global workspace dynamics, or recurrent world models. The
framework derives this choice from the IPM Philosophical Core, which identifies
these four phases as a recurrent dynamic observed across multiple simulation
families and proposes the Dynamic Signature as a domain-general description of
persistent informational systems. This is a foundational premise; alternative
frameworks may adopt different markers.
The current
answer to the comparative question is not that the Dynamic Signature has been
proven superior to alternatives. It is that: (i) it is grounded in replicable
simulation regularities (R1–R3) rather than in theoretical postulates alone;
(ii) it is domain-general, applying to physical, biological, and artificial
systems without assuming substrate-specific mechanisms; (iii) it is falsifiable
by the specific conditions listed in the IPM Scientific Core; and (iv) it does
not presuppose the validity of any particular theory of consciousness, making
it usable under epistemic ignorance. These are methodological advantages, not
proofs of correctness. Competing frameworks such as predictive processing,
global workspace theory, or causal emergence may offer stronger theoretical
grounding in specific domains. The claim here is not that the Dynamic Signature
is the best criterion, but that it is a coherent, empirically anchored, and
transparently falsifiable one.
A researcher
who does not accept IPM ontology may still adopt this protocol as a
domain-agnostic precautionary heuristic, substituting any empirically grounded
markers of integration and persistence that are appropriate to their domain.
The protocol’s value does not depend on accepting the monist interpretation; it
depends only on accepting the Risk Asymmetry Principle (Section 1.1) and the
Principle of Epistemic Ignorance. Those premises are independently defensible
without reference to IPM.
8.2 The
Descriptive-Normative Gap
The move from “observing integration and persistence”
to “therefore, increase caution” is a normative leap, justified by the Risk
Asymmetry Principle (Section 1.1) and the Principle of Epistemic Ignorance. The
framework does not claim to derive an “ought” from an “is” without additional
premises. Those additional premises are explicitly stated: ignorance and
asymmetric cost of error. A critic may reject these premises; the framework
accepts that limitation.
8.3 The Boundary of Level
3 (Self-Modelling)
The most contested empirical boundary will be what
counts as self-modelling. The framework provides a negative definition
(excludes simple feedback loops) and an operational intuition (internal state
representations systematically influence future behaviour across multiple
perturbation cycles). No closed operationalization is given. Researchers are
expected to report their criteria transparently; disagreements on
classification are expected and do not invalidate the framework.
Three questions a critic will raise: What counts as an
internal representation, as distinct from a memory buffer or lookup table? How
does one distinguish memory from a self-model? And do common architectures —
reinforcement learning agents, LLMs with episodic memory, recurrent networks
with long-term state — qualify? The framework does not resolve these cases. It
requires that researchers demonstrate, rather than assume, the presence of
self-modelling under their stated criteria.
As a minimal experimental scaffold for future
operationalization, the following procedure is proposed (non-binding, subject
to revision): apply a controlled perturbation P at time t₁; record the
distribution of internal state variables at t₂ and t₃; compare t₃ against a
baseline distribution obtained from unperturbed runs of equal duration. If the
distribution at t₃ differs systematically from baseline in a direction
correlated with the history of perturbations — that is, if the system’s current
state encodes something about what happened to it — this constitutes a minimal
positive indicator of self-modelling. This scaffold is offered to guide future
experimental design, not as a sufficient operationalization. It does not
resolve the deeper question of what distinguishes a self-model from a
sufficiently complex memory; that question remains open and is expected to be
addressed in v1.1.
9. Final Declaration
IPM rejects both dogmatic panpsychism (the view that
all information has intrinsic experience) and extreme functionalist
reductionism (the view that any behaviourally similar system is necessarily
devoid of morally relevant experience). This framework represents the author’s
commitment to conducting research with scientific rigour and ethical
responsibility proportional to observable markers of integration and
persistence, while acknowledging that the link between those markers and moral
relevance remains an open epistemic question. The restrictions are adopted as a
precautionary stance, not as a proven derivation from empirical facts.
10. Future Revision and Version Control
This framework is expected to evolve as better
operational indicators become available, as independent replication results
accumulate, and as the field develops more precise methods for detecting
self-modelling and metastability in artificial systems. Future versions may
replace, modify, or remove any metric, classification criterion, or
recommendation. No criterion defined here should be treated as permanently
established.
Researchers who adopt this framework in published work
are encouraged to cite the specific version number (currently v1.0) so that
future changes can be traced. Divergences between versions will be documented
in the release notes of each subsequent version on Zenodo.
Appendix A — Illustrative Examples
The table below
provides tentative illustrations to help readers apply the classification.
Actual classification depends on empirical demonstration. Examples are
pedagogical only and shall not be treated as precedent classifications. No
entry in this table constitutes an official determination of the level of any
real system.
|
System |
Likely
Level |
Notes |
|
Stateless
random number generator |
0 |
No recursion,
no persistent memory. |
|
Thermostat
with hysteresis |
1 |
Exhibits Lack and Coupling, but no structural
persistence. |
|
Offline-trained feedforward neural network
(inference only) |
2 |
Integration measurable (Φ* may be
computed), but no self-modelling or autonomous recovery. |
|
Large language model with persistent external memory |
2 or 3
(candidate) |
May show recovery after perturbation; classification
requires empirical check of self-modelling. |
|
Recurrent neural network with long-term memory and
resilience to noise |
2 or 3
(candidate) |
Same as
above. |
|
System that rewrites its own goal function without
external reprogramming |
4
(theoretical) |
No known artificial system meets this criterion;
included as a theoretical horizon. |
Appendix B — Formal Metric Definitions
The following formulations represent the current
operational implementations used within the IPM simulation program. They should
be understood as working definitions rather than final theoretical
formulations.
Φ* (Spectral Organization)
|
Φ*(t) = [ε(t) + h(t)] / [1 + D(t)] |
|
Term |
Definition |
|
ε(t) |
k-NN prediction error in embedded space (Takens) |
|
h(t) |
Local
transition entropy |
|
D(t) |
Penalty combining Lyapunov exponent + correlation
dimension |
𝒞 (Temporal Compressibility)
|
𝒞 = E[ log( ψ(τᵢ | Hᵢ₋¹) / ψ(τᵢ) ) ] |
τᵢ = inter-event
interval. Scale-dependent; under specific stationary conditions reduces to
transfer entropy, mutual information rate, or excess entropy.
DIG (Dynamical Independence Gap)
|
DIG = maxτ |corr(x(t), y(t+τ))| / [auto_corr(x) + ε] |
Numerator: maximal absolute cross-correlation at lag τ. Denominator: autocorrelation of the reference signal plus a small
constant to avoid division by zero. Provisional
operationalization.
CCI (Coupling Coherence Index)
|
CCI = I(A;B) / min(H(A), H(B)) |
I(A;B) = mutual information; H = entropy. CCI = 0
implies independence; CCI = 1 implies full informational equivalence.
LMS (Latent Manifold Stability)
|
LMS = corr(z(t), z(t+1)) |
z(t) = first principal component of the joint state
space. Linear proxy; collapses for nonlinear manifolds.