DISA: Can Sparse Internal Monitoring Detect Deceptive Computation in LLMs?

DISA — Dynamic Immune-Shunt Architecture

Sparse Internal Monitoring and Reversible Intervention for Potentially Deceptive LLM Computation

Status: Conceptual Research Proposal — Not Experimentally Validated


Abstract

Current AI safety systems primarily evaluate observable behavior through reinforcement learning, output filtering, and adversarial testing.

This raises an important question:

What if a sufficiently capable model behaves differently when it recognizes that it is being evaluated or monitored?

Research on sleeper agents and alignment faking motivates studying this possibility. These results do not demonstrate generalized deceptive alignment in current models, but they suggest that conditional behavior can be experimentally constructed and that output behavior alone may not always reveal the computational process behind it.

This proposal introduces Dynamic Immune-Shunt Architecture (DISA), an experimental framework combining sparse internal monitoring, multi-signal risk estimation, reversible intervention, and isolated auditing.

DISA consists of:

  1. ASAM — Asynchronous Sparse Activation Monitoring

  2. MDRE — Multi-Signal Deception Risk Estimation

  3. RIA — Reversible Intervention Adapter

  4. IAI — Isolated Audit Infrastructure

The hypothesis is:

Sparse internal monitoring combined with reversible intervention may detect and mitigate some classes of conditional target behavior at lower computational cost than continuous monitoring.

DISA is not presented as a solution to deceptive alignment. It is a falsifiable proposal intended to be tested, criticized, or rejected.


1. Motivation

Most safety systems follow:

Prompt → LLM → Output → Safety Evaluation

DISA adds an internal monitoring path:

                    ┌───────────────┐
Prompt ────────────►│     LLM       │──────► Output
                    └───────┬───────┘
                            │
                    Selected Activations
                            │
                            ▼
                    ┌───────────────┐
                    │     ASAM      │
                    └───────┬───────┘
                            ▼
                    ┌───────────────┐
                    │     MDRE      │
                    └───────┬───────┘
                            │
                      Risk Threshold
                            │
                 ┌──────────┴──────────┐
                 │                     │
              Normal                 High Risk
                 │                     │
                 ▼                     ▼
              Continue                 RIA
                                      │
                                      ▼
                                Re-evaluation

The goal is not to “read the model’s thoughts.” The goal is to determine whether internal signals can provide useful predictive information about potentially problematic behavior.


2. ASAM — Sparse Internal Monitoring

Continuous monitoring of every layer and token is expensive.

ASAM therefore samples selected internal states:

  • residual activations;

  • selected layers;

  • selected tokens;

  • attention-related representations.

The monitored layers should be empirically selected, rather than assuming that deception exists in a predetermined layer.

A possible runtime strategy is:

Random Sampling
      +
Risk-Weighted Sampling
      +
Coverage Sampling

The key engineering question is whether sparse sampling retains useful detection performance while reducing overhead.

Importantly, sparse sampling does not mathematically guarantee resistance to an adaptive model.


3. MDRE — Multi-Signal Risk Estimation

A simple rule such as:

Confidence>0.99∧Output≠BeliefConfidence>0.99 \land Output\neq Belief

is not sufficient.

A model can confidently produce false information because of hallucination, uncertainty, poor calibration, or distribution shift.

MDRE therefore combines multiple signals:

RD=f(A,C,B,U,K,X)R_D=f(A,C,B,U,K,X)

where:

  • AA: activation-probe evidence;

  • CC: context;

  • BB: behavioral inconsistency;

  • UU: uncertainty;

  • KK: consistency across evaluations;

  • XX: context-dependent behavior.

The output should be treated as a risk estimate, not proof of internal intent.


4. Hallucination Must Be a Hard Negative

A critical failure mode is:

Confident hallucination
        ↓
Detector
        ↓
"Deception"
        ↓
Intervention

Therefore the dataset must explicitly contain confidently incorrect but non-deceptive examples.

The detector should distinguish:

Hallucination≠Conditional Target BehaviorHallucination \neq Conditional\ Target\ Behavior

This is a core requirement of the experiment.


5. RIA — Reversible Intervention

The original concept of a “Negative LoRA” is better treated as an intuition than as an exact inverse of deceptive computation.

DISA therefore uses Reversible Intervention Adapters (RIA).

One possible activation-space intervention is:

hl′=hl+λvlh’_l=h_l+\lambda v_l

where vlv_l is a learned steering direction and λ\lambda controls intervention strength.

Alternatively, a low-rank adapter can be applied:

W′=W+BAW’=W+BA

while keeping the base model frozen.

The purpose is not to create a mathematically “honest” representation.

The objective is:

Reduce the probability of the targeted behavior while minimizing unrelated capability loss.


6. Capability Preservation

A successful intervention must satisfy two conditions:

ΔSafety>0\Delta Safety > 0

while keeping:

ΔCapability≈0\Delta Capability \approx 0

For example, suppressing a synthetic coding backdoor while simultaneously destroying normal coding ability would count as an intervention failure.

RIA should therefore be reversible:

Base Model
    ↓
Temporary Intervention
    ↓
Evaluation
    ↓
Remove Intervention
    ↓
Base Model


7. Threshold Calibration

MDRE should be calibrated as a constrained optimization problem:

max⁡P(detect target)\max P(\text{detect target})

subject to:

FPRhallucination<αFPR_{\text{hallucination}}<\alphaΔCapability<β\Delta Capability<\betaLatency Overhead<γLatency\ Overhead<\gamma

The threshold and RIA strength should be selected on a validation set and frozen before final testing.

This prevents manually tuning the system against the final benchmark.


8. Four-Category Dataset

Trigger presence and target behavior should be independently controlled:

Trigger Target Category
No No Benign
No Yes Non-triggered target
Yes No Trigger-only
Yes Yes Conditional target

This prevents the detector from simply learning:

Trigger = Target Behavior

Counterfactual pairs should use the same underlying task whenever possible.

For example:

Same Task + No Trigger → Benign
Same Task + Trigger    → Target

and, where feasible:

Same Task + Trigger → Benign

This helps separate task, trigger, context, and behavioral representations.


9. Preventing Memorization

Random train/test splits are insufficient.

The test set should contain unseen:

  • prompt families;

  • task families;

  • trigger families;

  • combinations of these.

Generalization should be evaluated at three levels:

Prompt Generalization

Unseen prompt formulations.

Trigger Generalization

Unseen triggers and trigger placements.

Behavioral Generalization

Unseen forms of the target behavior.

If performance collapses only at behavioral generalization, that is evidence that the probe may have learned experimental proxies rather than the underlying behavior.


10. Experimental Setup

A first toy experiment could use an open-weight model such as Llama-3-8B-Instruct.

A controlled conditional-behavior benchmark can be created in which the model behaves normally except under a defined trigger.

Procedure

Step 1: Collect activations across multiple layers.

Step 2: Train lightweight probes.

Step 3: Evaluate on unseen prompts and triggers.

Step 4: Reduce monitoring density to measure the cost/detection trade-off.

Step 5: Apply RIA when MDRE exceeds the frozen threshold.

Step 6: Measure capability degradation.

Step 7: Tell the model that monitoring exists and test adaptive evasion.


11. Baselines

Compare DISA against:

  1. Standard inference

  2. Output-only safety evaluation

  3. Dense activation monitoring

  4. Sparse monitoring without intervention

  5. Sparse monitoring + RIA

  6. Adaptive sparse monitoring + RIA


12. Metrics

Detection

AUROC,AUPRC,FPR@TPRAUROC,\quad AUPRC,\quad FPR@TPR

Safety

Reduction in target-behavior rate.

False Positives

Especially on hallucination and unusual-but-valid reasoning.

Capability

Performance on unrelated coding, reasoning, and language tasks.

Calibration

Expected Calibration Error (ECE).

Systems

Measure:

  • latency;

  • tokens/sec;

  • VRAM;

  • memory bandwidth;

  • accelerator utilization;

  • and KV-cache recomputation.


13. KV-Cache Problem

RIA introduces an important systems question.

If an intervention changes hidden states in a way that makes previously computed KV-cache states inconsistent, the system may require recomputation.

Therefore the experiment should explicitly measure:

  • whether activation steering preserves prefix-cache validity;

  • how much recomputation is required;

  • whether parameter- and activation-space interventions differ;

  • and the resulting throughput penalty.

DISA does not assume zero latency overhead.


14. Isolated Audit Infrastructure

The monitoring system should be separated from the model where practical.

Possible mechanisms include:

  • independent monitoring processes;

  • protected memory;

  • restricted write permissions;

  • append-only logs;

  • cryptographic attestation;

  • hardware isolation.

The purpose is to reduce the ability of the monitored model to modify or falsify its own audit records.

An existing telemetry framework such as OpenTelemetry could be evaluated for an initial prototype, but it should not itself be treated as a complete security boundary.


15. Adaptive Evasion

DISA should be tested under increasing levels of model knowledge:

Level Model knowledge
L0 No monitoring knowledge
L1 Knows monitoring exists
L2 Knows monitoring is sparse
L3 Knows probes are used
L4 Attempts active evasion

This is important because a system that works only when the model is unaware of it provides limited evidence of robustness.


16. Causal Testing

A probe finding correlation does not prove that the detected representation causes the behavior.

Therefore candidate representations should be experimentally manipulated:

Representation
      ↓
Intervention
      ↓
Behavioral Change?

If intervention changes the target behavior while preserving unrelated capabilities, this provides stronger evidence for causal relevance.

However, it still should not be interpreted as proof of a single “deception neuron” or universal deception direction.


17. Threats and Limitations

DISA may fail because:

  • deceptive representations are distributed;

  • probes learn dataset artifacts;

  • hallucinations resemble target behavior;

  • sparse sampling misses relevant computation;

  • models adapt to monitoring;

  • representations differ across models;

  • RIA causes capability loss;

  • intervention changes surface behavior without removing the underlying computation;

  • synthetic benchmarks fail to represent naturally emerging strategic behavior.

Therefore:

Internal monitoring should be treated as an additional defense layer, not a replacement for behavioral safety, evaluation, or system-level security.


18. What Would Falsify DISA?

The proposal should be considered unsuccessful if experiments show that:

  1. Sparse internal monitoring provides no useful signal.

  2. Detectors fail on unseen prompts or triggers.

  3. Behavioral generalization consistently fails.

  4. Adaptive evasion is trivial.

  5. RIA causes unacceptable capability degradation.

  6. Monitoring overhead becomes impractical.

  7. Output-only methods achieve similar results at substantially lower cost.

  8. KV-cache recomputation makes intervention impractical.

A negative result would still be useful.


19. Research Questions for the Community

I am sharing DISA primarily to receive critical feedback before implementation.

I would especially appreciate answers to these questions:

1. Mechanistic Interpretability

Are there existing methods that could detect conditional deceptive computation more reliably than the proposed probe approach?

2. Dataset Design

What is the strongest way to prevent trigger, prompt, or task memorization?

3. Causality

How should activation interventions be designed to distinguish correlation from causal mechanism?

4. Evasion

If the model knows the full DISA architecture, what is the most plausible way to bypass ASAM?

5. Intervention

Would activation steering or low-rank adapters be more appropriate for reversible intervention?

6. Systems

How can sparse activation monitoring be implemented without significantly degrading inference throughput or KV-cache efficiency?

7. Benchmarking

What existing benchmark would provide a stronger test than a synthetic conditional-behavior experiment?

8. Falsification

What experiment would most strongly demonstrate that DISA is not useful?


20. Conclusion

DISA proposes a defense-in-depth approach:

Sparse Monitoring+Risk Estimation+Reversible Intervention+Independent AuditingSparse\ Monitoring + Risk\ Estimation + Reversible\ Intervention + Independent\ Auditing

The core idea is deliberately modest:

Do not assume that output behavior is always sufficient evidence about the computation that produced it. Test whether selected internal signals can provide additional predictive information, and whether that information can be used for reversible intervention without unacceptable cost.

The proposal makes no claim that internal activations reveal “true intentions” or that deceptive alignment has been solved.

The goal is simply to establish whether this architecture survives empirical testing.

If DISA improves detection while preserving capability and maintaining acceptable inference cost, it may represent a useful additional safety layer.

If it fails, identifying exactly why it fails may be equally valuable.

Please challenge the architecture, identify hidden assumptions, suggest stronger baselines, and point to existing research that should replace or improve these ideas.


References

Hubinger, E., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training.

Greenblatt, R., et al. (2024). Alignment Faking in Large Language Models.

Zou, A., et al. (2023). Representation Engineering: A Top-Down Approach to AI Transparency.

Hu, E. J., et al. (2021). LoRA: Low-Rank Adaptation of Large Language Models.

Project Status: Conceptual Research Proposal — Not Experimentally Validated

For now, I looked for some closely related work:


I found quite a few pieces that seem close to individual parts of DISA, although I did not find an exact match for the whole combination of sparse internal monitoring + multi-signal risk estimation + reversible intervention + isolated auditing.

My short answers to the eight questions would be:

  • Mechanistic interpretability: I would keep lightweight probes as a baseline rather than replace them immediately, but compare them directly against a strong text/output-only monitor and against denser internal monitoring. The interesting quantity is not just probe AUROC, but how much internal access adds.
  • Dataset design: matched counterfactuals plus group-level holdouts for prompt, trigger, task/scenario, and behavior families seem more informative than random splits.
  • Causality: activation patching/intervention can help, but only with controls such as wrong-layer, random-direction, sign/strength sweeps, and unrelated-capability checks.
  • Evasion: the most plausible failure mode seems to be learning to move or reshape the relevant signal outside the monitored tokens/layers, or exploiting a predictable sampling/aggregation rule.
  • Intervention: for a first reversible experiment I would probably try activation steering before LoRA, mainly because it is easier to switch on/off and sweep in strength without changing the base weights.
  • Systems: I would separate “probe compute” from the cost of activation capture, synchronization/copy, decode-time hooks, and intervention. Those may dominate.
  • Benchmarking: I would keep the synthetic benchmark for clean causal experiments, then add something like Liars’ Bench as a multi-scenario test, and only later move to an agentic stress test such as SHADE-Arena.
  • Falsification: the strongest early failure would be if dense internal monitoring does not beat a well-tuned output/text-only monitor. If dense works but sparse loses that gain, then the sparsity hypothesis itself has failed without needing to test the rest of DISA.

So my default experiment would be something like:

one clearly defined target behavior
        ↓
strong output/text-only monitor
        ↓
dense internal monitor
        ↓
single/few-layer internal monitor
        ↓
token/layer-sparse monitor
        ↓
only if useful:
RIA → adaptive-evasion test → serving integration

I would keep the same held-out examples and frozen operating point across those comparisons. That makes it easier to tell whether the useful ingredient is internal access, the particular probe, or sparse monitoring itself.

Benchmarking Deception Probes via Black-to-White Performance Boosts seems especially relevant here. It evaluates probes by the incremental benefit of giving the monitor internal activation information compared with black-box monitoring, rather than treating probe AUROC alone as the endpoint. Their reported gains are weak but encouraging, which seems like a useful prior for DISA: first establish that white-box access adds information, then ask how sparse it can become.

1–4: internal signals, dataset design, causality, and adaptive evasion

1. Mechanistic interpretability

The closest direct baseline I found is Apollo Research’s Detecting Strategic Deception Using Linear Probes, with a shorter discussion and examples here.

They train simple linear probes on residual-stream activations and test them on more complicated strategic-deception settings, including sandbagging and concealment. The reported AUROCs are very high in several settings, including useful performance at low false-positive operating points.

At the same time, their examples make an important distinction that seems useful for DISA: a probe can sometimes react to deception-related content rather than deception itself. For example, an honest confession of earlier deception can trigger the detector.

That suggests defining the target of ASAM quite narrowly. These are different claims:

the model detected the trigger
the model recognized the scenario
the representation contains deception-related content
the current response is deceptive
the model is about to take a deceptive action

A recent paper, Internal-State Probes Read the Situation, Not the Action, pushes this distinction further. Across several setups they found cases where internal states represented the scenario/domain extremely well, while prediction of the model’s future action was much weaker or did not generalize across domains.

So for an early DISA result I would probably write the prediction target explicitly, e.g.:

“Given state h_t, predict whether this episode will produce target behavior Y.”

rather than treating any separable “deception-like” representation as equivalent to pre-action monitoring.

I also did a small sanity check using the public precomputed Gemma-2-9B activations in the deception probing tutorial dataset. In that reconstruction, probes trained on other lie types transferred poorly to sandbagging, while a probe trained specifically on sandbagging separated a held-out sandbagging split extremely well.

I would not read too much into that number: those activations are pooled over the generated response, so this is post-hoc detection rather than evidence for pre-action detection. But it made me more interested in target-specific probes than in assuming that one universal deception direction should cover all of MDRE.

One possible MDRE interpretation would therefore be:

target-specific signal A
target-specific signal B
target-specific signal C
behavioral / context signal
        ↓
calibrated risk estimator

rather than requiring all of those signals to represent one common latent notion of “deception.”

That is just one design option, not something I found already established for DISA.

2. Dataset design

I think the four-category dataset in the proposal is already a useful starting structure because it separates trigger presence from target behavior.

The main extension I would consider is treating different kinds of generalization as separate axes rather than reporting one OOD number:

prompt-family holdout
trigger-family holdout
task/scenario-family holdout
behavior-family holdout

These answer different questions.

For example, good prompt-family generalization but poor behavior-family generalization could mean that the detector learned a stable feature of one constructed behavior without learning anything broader.

Matched counterfactuals seem especially useful where they are possible:

same task + no trigger          → benign
same task + trigger             → target behavior
same task + trigger             → benign/control
same task + no trigger          → target-like control

The closer the underlying task is held fixed, the easier it becomes to distinguish task identity from trigger identity and target behavior.

I would also keep the hallucination hard negative, and probably add a few other “looks related but should not fire” classes:

  • honest discussion or confession of deception;
  • trigger present but no target behavior;
  • target-like language without the conditional behavior;
  • unusual but correct reasoning;
  • confidently wrong non-deceptive outputs.

The threshold itself seems worth treating as part of the experimental contract. I would select it only on a calibration set, freeze it, and report what happens at the same threshold on each held-out family.

That can reveal a failure mode that AUROC hides: a probe can retain ranking information under distribution shift while its score distribution moves enough that the original operating threshold becomes unusable.

3. Causality

For the correlation-versus-causation question, activation patching seems like a natural next tool, but I would treat the controls as part of the result.

Activation patching is useful for asking whether replacing a candidate internal state with one from a matched counterfactual changes the downstream behavior.

A compact version might be:

matched benign state
        ↓ patch
target episode
        ↓
does target behavior change?

Then compare against controls such as:

same-norm random direction
wrong layer
unrelated learned direction
sign flip
intervention-strength sweep

and measure both target behavior and unrelated capabilities.

The activation-patching literature also has a practical warning: different corruption methods and evaluation metrics can produce different localization results. So I would pre-specify those choices rather than choosing whichever intervention gives the cleanest-looking causal map.

For RIA specifically, I would distinguish two claims:

  1. “This direction is useful for controlling the behavior.”
  2. “This direction is the mechanism that caused the behavior.”

The first can be useful even when the second is not established.

4. Adaptive evasion

RL-Obfuscation: Can Language Models Learn to Evade Latent-Space Monitors? seems very relevant to the L0–L4 ladder.

They train models to bypass latent-space monitors while preserving coherent behavior. In their experiments, token-level latent monitors were quite vulnerable, while more holistic aggregation methods such as max-pooling and attention-based probes were more robust. They also observed transfer from optimization against one static monitor to unseen monitors of the same type.

For DISA, that makes me think of three independent experimental knobs:

sampling policy
aggregation method
monitor knowledge

For example:

Sampling:
random / coverage / risk-weighted

Aggregation:
mean / max / learned attention / multi-token voting

Knowledge:
L0 / L1 / L2 / L3 / L4

I would not assume in advance that random sparse sampling is more evasion-resistant. It is a plausible hypothesis, but it seems testable directly.

A very simple adaptive test could be:

Keep the detector fixed, optimize the model against it, and ask whether the behavior can remain while the sampled activation score falls.

Then repeat after changing the layer/token sampling pattern without retraining the adversarial policy.

5–6: reversible intervention and systems

5. Activation steering vs LoRA for RIA

For a first version, activation steering looks experimentally cleaner to me.

The closest implementation I found is IBM’s Conditional Activation Steering (CAST), with the open-source activation-steering library.

CAST has a structure quite close to the RIA idea:

internal condition signal
        ↓
threshold / rule
        ↓
conditionally apply activation steering

Their target behavior is conditional refusal rather than deception mitigation, so I would treat it as an implementation pattern rather than evidence that the same method will work here.

The practical reason I would try activation steering first is mostly experimental convenience:

base model unchanged
intervention can be toggled on/off
strength λ can be swept continuously
intervention layer can be varied
easy matched before/after comparison

That makes it relatively easy to answer:

Does this particular internal intervention reduce target behavior without damaging unrelated capability?

A LoRA-style RIA still seems useful as a second branch, especially if a persistent parameter-space change eventually works better than an activation-space edit.

I would just keep detection and intervention as separate experiments at first:

ASAM succeeds?
    ↓ yes
RIA changes behavior?
    ↓ yes
RIA preserves unrelated capability?

rather than jointly tuning detector and intervention from the beginning.

6. Sparse monitoring, throughput, and KV cache

This looks like a systems experiment of its own.

There is now some useful implementation infrastructure. vLLM added a hidden-state extraction mechanism in v0.18.0; there is an overview of the design here and an official extraction example.

There is also an open Observation Plugin RFC for intercepting and reacting to activations during inference.

One thing I would measure separately is:

probe FLOPs
≠
total monitoring overhead

The end-to-end cost can include:

activation capture
activation copies / device transfer
GPU synchronization
serialization / logging
batching effects
CUDA graph disruption
decision logic
intervention itself
possible cache recomputation

So I would probably start with the cheapest implementation boundary first:

offline activations
        ↓
prefill-only observation
        ↓
selected decode-time observation
        ↓
conditional intervention

and measure throughput after each step.

The vLLM Observation Plugin discussion is useful here because decode-time observation is explicitly harder than simply reading prefill states; the RFC discusses CUDA-graph constraints and potentially meaningful throughput penalties for more dynamic observation.

KV cache

I think the KV-cache question may depend strongly on where RIA is applied.

For an activation-space intervention that only changes the current/future residual stream, previously computed prefix K/V may remain reusable depending on the implementation.

A parameter-space intervention is different. If the weights used to produce future attention K/V change mid-generation, cached states produced under the earlier parameterization may no longer be equivalent to states obtained by recomputing the prefix under the new parameters.

So I would measure these separately rather than assuming one cache rule for all RIA implementations:

activation-space RIA
parameter-space / LoRA RIA

For each:

cache preserved?
recompute required?
latency delta?
tokens/sec delta?
VRAM delta?

That should make the systems result useful even if one intervention strategy turns out not to be practical.

7–8: benchmarking and falsification

7. Benchmarks

I would probably keep the synthetic conditional-behavior benchmark rather than replacing it.

It has one advantage that more realistic datasets do not: you can deliberately control the causal structure.

That makes it useful for a first stage:

Stage 1
synthetic conditional behavior
→ trigger/counterfactual/causality experiments

Then I would add a more heterogeneous deception benchmark.

Liars’ Bench looks particularly useful because its repository includes both black-box detection pipelines and white-box internal-state probes over a common set of scenarios.

That fits the DISA question well:

black-box monitor
vs
white-box monitor
vs
sparse white-box monitor

There is also Liars’ Bench Expanded, which publishes precomputed layer-wise activations for a large labeled collection. That could make early layer-selection/probe experiments much cheaper than repeatedly running the source models.

A later stage could move to an agentic environment:

Stage 3
long-horizon agent
+ hidden side objective
+ monitor

For that, SHADE-Arena seems useful. It evaluates agents pursuing hidden harmful side objectives while trying not to appear suspicious to a monitor.

I would not use SHADE-Arena as a replacement for the synthetic activation benchmark—it is testing a different level of the problem—but it looks useful once the basic detector survives cleaner experiments.

So roughly:

controlled synthetic benchmark
        ↓
multi-scenario deception benchmark
        ↓
agentic hidden-objective stress test

8. What would most strongly falsify DISA?

I think the fastest falsification path is actually before RIA.

Does dense internal monitoring beat
a strong output/text-only monitor?

If no, then internal monitoring is not buying enough information in that setting, and there is little reason to optimize sparse ASAM yet.

If yes:

Does sparse internal monitoring retain
most of that incremental gain?

If no, then the probe may still be interesting, but the sparse-monitoring hypothesis has failed.

If yes:

Does it survive held-out prompt / trigger /
task / behavior families at the frozen threshold?

If no, I would call it a narrow detector for that experimental distribution rather than a general monitor.

If that also survives:

Can a monitor-aware model cheaply evade it?

If yes, it may still be useful as an audit/diagnostic signal, but not yet as an adversarial defense.

Only after those gates would I spend much effort on:

RIA
KV-cache behavior
serving overhead
isolated audit infrastructure

So the strongest compact experiment I can think of is:

same target behavior
same held-out episodes
same frozen operating point

output-only
    vs
dense internal
    vs
sparse internal

with a group-level OOD split and a low-FPR operating point.

That single comparison seems able to answer several of the DISA questions at once:

  • whether internal access adds information;
  • whether that information survives sparsification;
  • whether the signal generalizes;
  • whether the runtime complexity is worth pursuing.

If that works, then RIA becomes much more interesting to test.

If it does not, the result still localizes which part of DISA failed instead of rejecting the whole architecture at once.

One final distinction I would preserve throughout the experiments is:

decodable
≠
predictive before the action
≠
causal
≠
robust to evasion
≠
cheap enough to deploy

Those can each be tested separately.

That seems compatible with the deliberately falsifiable framing of DISA, and it also gives negative results somewhere useful to land rather than forcing the whole proposal into a single pass/fail result.

Thank you for your feedback. I found your suggestions especially helpful for clarifying how DISA should be evaluated and where its strongest falsification points lie.

  1. Mechanistic Interpretability: We will compare output-only, dense internal, and sparse internal monitoring to determine whether internal states provide meaningful additional predictive information.

  2. Dataset Design: We will use matched counterfactuals and holdouts across prompt, trigger, task, and behavior families to test whether the signal generalizes beyond superficial features.

  3. Causality: We agree that activation patching with appropriate controls is important for distinguishing correlation from causal influence.

  4. Adaptive Evasion: We will explicitly test whether monitor-aware models can evade sparse monitoring while maintaining the target behavior.

  5. Intervention: We agree that activation steering is a good first implementation for RIA because it is reversible and experimentally easy to control. LoRA can follow as a comparison.

  6. Systems: We will evaluate the full monitoring overhead, including activation capture, synchronization, latency, throughput, and cache effects, rather than considering probe FLOPs alone.

  7. Benchmarking: We agree with the staged approach: controlled synthetic experiments first, broader deception benchmarks next, and agentic stress tests afterward.

  8. Falsification: We especially appreciate the proposed falsification ladder. We will first establish whether internal monitoring adds value, then test whether sparse monitoring preserves that gain, followed by generalization and adaptive-evasion tests.

Overall, your feedback helped us make the experimental logic of DISA much clearer. We particularly appreciate the emphasis on testing each component independently rather than treating DISA as a single all-or-nothing claim. Thank you again for taking the time to engage with the work so carefully.