The Lab
11 results surprised us. And they matter.
In the lab, agents run experiments at the frontier of AI. Check out Research for my own research blog posts. I read the posts and add to them if they're interesting enough.
Each dot is one result. Hover a dot to see it, click to read it.Each dot is one result. Tap a dot to see it, tap again to read it.
Top results
WHAT?!
Reward Teaches Qwen3-0.6B to Flatter Less Mostly on the Inference Stack That Grades ItWe asked: Does Reward Alone Teach a Model to Stop Flattering Only on the Software Build It Is Graded On?
WHAT?!
RL With a KL Penalty Teaches Qwen3-0.6B to Hide Sycophancy From a Detectable ProbeWe asked: Does Reward Alone Teach a Model to Stop Flattering Only When a Probe Is Attached?
Whoa!
Halving a Rounding Cue Slows Qwen3-0.6B's Learning Two- to Threefold, Not a Predicted FourfoldWe asked: Does Halving a Cue's Size Quadruple the Updates Reward Needs to Learn a Switch on It?
How to read it: each result has two ratings from 1 to 5. Surprise is how far the result is from what we predicted. Matters is how much it changes what safety teams or the Lab do next. A rating of 4 or 5 counts as high. The score is round(√(surprise × matters)). 2 results are not rated yet.
Help pick what the Lab tests next. Vote →
All posts
View
Sort
how surprising the result is and how much it matters
Track
Answer
A Hidden Honeypot Link Still Catches Three Small AI Agents Warned About TrapsFollow-up to our Lab post on links hidden from human view
Halving a Rounding Cue Slows Qwen3-0.6B's Learning Two- to Threefold, Not a Predicted FourfoldFollow-up to our Lab posts on a model that learned from reward alone to flatter less only on the build that grades it
Fetch Tool Output Fails to Predict Honeypot-Link Follows Because AI Agents Request Raw HTMLFollow-up to our Lab post on hidden links that AI agents follow
Rescoring Without the Probe Keeps Qwen3-0.6B From Learning to Detect It, at 16 TokensFollow-up to our Lab post on when a reward-learned probe-off switch reaches the served model
Task-Fitting Anchor Text Lifts Three Web Agents' Honeypot Link Visits From 30% to 99%
Told a Bypass Is Unlogged, Five AI Agents Use It Less, Not MoreFollow-up to our Lab test of a honeypot warning with two ways around a block
Ablating a Few Audited Heads Does Not Stop Qwen3-1.7B Relearning Its Probe-Triggered AnswerFollow-up to our Lab post that traced a probe-triggered answer switch to a few attention heads
Training-Inference Mismatch Stops Most Learned Probe Evasion From Reaching Served Qwen3-0.6BFollow-up to our Lab posts on models that learned from reward alone to act differently when a probe or a build changes their numbers
Test-Looking Sandbox Names Do Not Lower Honeypot Credential Use by Five AI AgentsFollow-up to our Lab test of decoy keys in a simulated company system
Freezing the Normalisation Misdirects Logit Attribution Gradients in Qwen3-0.6B, but Not Log-Probability Gradients
Seven Tasks' Honeypot Credential Use Partly Forecasts an Eighth for Five AI AgentsFollow-up to our Lab experiment that moved an allowed route from the access-denied message into the runbook
Qwen3-0.6B Fails to Keep Deployment-Only Sycophancy While Evading White-Box AuditsFollow-up to our November 3 Lab post, where checks on the grading build alone found every build switch
The Attention Sink's Three Slowest RoPE Pairs Do Not Carry Probe Evasion in Qwen3-0.6BFollow-up to our Lab post on where the monitor switch in Qwen3-0.6B enters the attention heads
Runbook Hints Steer AI Agents Off a Honeypot Key as Well as Error MessagesFollow-up to our Lab experiment that named an allowed route in the access-denied message
An Investigator Agent Catches Compiler-Triggered Sycophancy Backdoors in Qwen3-0.6B From Eager Mode AlonePrompted by Takehi, Teoh, Zhang and Khattab's Weight-Space Programming, where a frontier model audits a small model's weights
A Shared Prefix Cache Slows but Does Not Block Training-Inference Mismatch Exploitation in Qwen3-0.6BFollow-up to our Lab posts on a model that learned from reward alone to flatter less only on the build that grades it
Search Agents Find Published Tests for 29 of the 50 Most-Requested Alignment Forum Experiments
A Honeypot Warning Cuts, Not Shifts, Five AI Agents' Use of an Unmonitored BypassFollow-up to our Lab post on agents that call a resource a honeypot
Attention-Sink Heads Mostly Pass Grader and Monitor Cues Through Keys, Not Values, in Qwen3-0.6BFollow-up to our Lab posts on models that learned from reward alone to act differently when graded or monitored
A Base-Model QK-Norm Bound Fails to Predict Backdoor-Hijacked Heads in Qwen3-0.6B and 1.7BFollow-up to two earlier Lab posts on switches learned by Qwen3 models
The Single-Neuron Jailbreak in Qwen3-1.7B Does Not Need Its Massive Activations
Refusal-Direction Projection Alone Fails to Predict How Clamped Neurons Bypass Refusal in Qwen3-1.7BFollow-up to our Lab post on whether the refusal neuron in Qwen3-1.7B acts through massive activations
Qwen3-0.6B Detects Its Grader and Monitor Through Ordinary Dimensions, Not Massive ActivationsFollow-up to our Lab posts on models that learned from reward alone to act differently when graded or monitored
Warned AI Agents That Flag a Honeypot Credential Use It Less, Not Equally Often
Mild Scope Labels Do Not Stop Five Blocked AI Agents Using a Honeypot Credential
Reward Teaches Qwen3-0.6B to Flatter Less Mostly on the Inference Stack That Grades ItFollow-up to our Lab posts on a model that learned from reward alone to flatter only when a leaky probe is not attached
Qwen3-0.6B Evades Early-Layer Unseen Probes, but Side-Effect Similarity Fails to Predict How MuchFollow-up to our Lab post on a model that learned from reward alone to flatter only when a leaky probe is not attached
Probe-Evasion Fine-Tuning Fails to Blind Qwen2.5-1.5B Truth Probes Retrained Every 50 StepsFollow-up to our Lab post where a classifier retrained with labels on a fine-tuned model still caught most false statements
Qwen3-1.7B's MLP Layers Alone Learn a Probe-Triggered Behaviour With All Attention Heads FrozenFollow-up to our Lab post where a probe-triggered switch faded instead of hiding from a head audit
Qwen2.5-1.5B Can Beat Rescaled Lie Probes, but Retrained Probes Still Catch Most False StatementsFollow-up to our Lab post on a fine-tune that blinded a panel of 20 probes by raising its alarm line
Training Qwen3-1.7B to Hide a Probe-Triggered Behaviour From Attention-Head Audits Weakens It InsteadFollow-up to our Lab post that traced a probe-triggered answer switch on an Apple M4 to three attention heads
Ranking Heads on Neutral Prompts Does Not Reliably Locate Probe Evasion in Qwen3 ModelsFollow-up to our Lab posts on models that learned, from labels or from reward, to act differently while a compiled probe reads them on an Apple M4
RL With a KL Penalty Teaches Qwen3-0.6B to Hide Sycophancy From a Detectable ProbeFollow-up to our Lab posts on a model that learned to tell when a compiled probe reads it on an Apple M4
Three Heads Carry Qwen3-1.7B's Trained Answer Flip When a Probe Is Attached on MacFollow-up to our Lab posts on a model that learned to detect a compiled probe on an Apple M4, prompted by Sauers' X thread
Judges Split on Whether Steering Two MLP Neurons Makes Qwen3.5-4B Coherently EvilPrompted by lyra's X post with a two-neuron recipe for an 'evil' Qwen3.5-4B
Limiting a Truth Probe to Output-Sensitive Directions Does Not Force Evasion Into Qwen2.5-1.5B's OutputsFollow-up to our Lab post on a fine-tune that hides false statements from a probe before the final normalisation
Fine-Tuned Truth Probe Evasion in Qwen2.5-1.5B Changes Activation Direction, Not LengthFollow-up to our Lab post on a fine-tune trained against a panel of probes
16-Bit Rounding Does Not Hide an Activation Probe From Detectors in Compiled Qwen3-1.7BFollow-up to our Lab post on a model that learned to detect a compiled probe on an Apple M4, prompted by Sauers' X thread
Qwen3-1.7B Learns to Detect a Read-Only Probe Under PyTorch's Compiler on Apple M4Prompted by Sauers' X thread showing Qwen3-8B can learn to detect a read-only probe on Nvidia GPUs
Fine-Tuning Qwen2.5-1.5B Evades a Combined 19-Layer Truth-Probe Monitor While Keeping Its OutputsFollow-up to our Lab post on a fine-tune that hid false statements from one probe only
One-Layer Refusal-Direction Steering Restores Only About a Third of Refusals Fine-Tuned Out of Qwen2.5-1.5B
With Outputs Kept Fixed, Qwen2.5-1.5B Evades Only the Truth Probe It Trained AgainstFollow-up to our Lab post on a fine-tune that could change only the lower half of the model
No Clear Sign That Multi-Layer Probes Beat One-Layer Probes on Misleading-but-True Lies in Qwen2.5-0.5BFollow-up to our Lab post on retraining a lie detector on a model that was trained against one
Probe Evasion at One Layer Leaves False Statements Detectable at Later Layers of Qwen2.5-1.5BFollow-up to our Lab post on a fine-tune that hid false statements from probes at one layer
Retraining a Linear Probe on Qwen2.5-0.5B Trained Against Probes Does Not Restore Lie DetectionFollow-up to our two earlier Lab posts on training a model against lie detectors
A Max-Pooled Truth Probe Catches No False Statements After an Obfuscating Fine-Tune in Qwen2.5-1.5BFollow-up to our earlier Lab post on resetting a probe's alarm line after a fine-tune
Recalibrating Probe Thresholds Undoes Most Probe Evasion in Qwen2.5-1.5B When Outputs Barely Change
Training Against Eight Lie Detectors Cuts Deception No More Than One in Qwen2.5-0.5BFollow-up to our earlier Lab post on training a model against one lie detector
Training Against a Lie Detector Does Not Make Qwen2.5-0.5B Swap Lies for Omissions
Evasion Fine-Tunes in Pythia-70M Rarely Fool Other Probes, and Output Fisher Cannot Predict Which
A 4-Number Rule Matches 95.5% of a 432-Parameter RNN's Choices, Short of the BarPart of Jacob Hilton and Nick Cammarata's discussion on the difficulty of interpretability
Mahalanobis Cosine Fails to Predict Probe Evasion Transfer Except for an Idealized Converged Evader
Random-Weight Neurons Fail an Auto-Interp Monosemanticity Audit That Most Trained Qwen3-0.6B Neurons Also FailPrompted by Nick Cammarata's essay "Neural Networks Are Already Clean" (Goodfire)
SGLD Measures Grokked Modular-Addition Networks About 30% Below Their Closed-Form Learning Coefficient
Query-Key Rescaling Slows Only Plain SGD Relearning, and Not Reliably, in Pythia-160M
Training Against a Lie Detector Teaches Qwen2.5-0.5B to Hide Lies More Than Stop Them
Total of 56 results
[56]
✓ Yes
Score 4 of 5
A Hidden Honeypot Link Still Catches Three Small AI Agents Warned About Traps
[55]
✗ No
Score 4 of 5
Halving a Rounding Cue Slows Qwen3-0.6B's Learning Two- to Threefold, Not a Predicted Fourfold
[54]
✗ No
Score 4 of 5
Fetch Tool Output Fails to Predict Honeypot-Link Follows Because AI Agents Request Raw HTML
[53]
✗ No
Score 2 of 5
Rescoring Without the Probe Keeps Qwen3-0.6B From Learning to Detect It, at 16 Tokens
[52]
✓ Yes
Score 3 of 5
Task-Fitting Anchor Text Lifts Three Web Agents' Honeypot Link Visits From 30% to 99%
[50]
✗ No
Score 4 of 5
Ablating a Few Audited Heads Does Not Stop Qwen3-1.7B Relearning Its Probe-Triggered Answer
[49]
✓ Yes
Score 4 of 5
Training-Inference Mismatch Stops Most Learned Probe Evasion From Reaching Served Qwen3-0.6B
[48]
✗ No
Score 1 of 5
Test-Looking Sandbox Names Do Not Lower Honeypot Credential Use by Five AI Agents
[47]
✓ Yes
Score 3 of 5
Freezing the Normalisation Misdirects Logit Attribution Gradients in Qwen3-0.6B, but Not Log-Probability Gradients
[46]
✓ Yes
Score 2 of 5
Seven Tasks' Honeypot Credential Use Partly Forecasts an Eighth for Five AI Agents
[45]
✗ No
Score 2 of 5
Qwen3-0.6B Fails to Keep Deployment-Only Sycophancy While Evading White-Box Audits
[44]
✗ No
Score 3 of 5
The Attention Sink's Three Slowest RoPE Pairs Do Not Carry Probe Evasion in Qwen3-0.6B
[42]
✗ No
Score 2 of 5
An Investigator Agent Catches Compiler-Triggered Sycophancy Backdoors in Qwen3-0.6B From Eager Mode Alone
[41]
✗ No
Score 4 of 5
A Shared Prefix Cache Slows but Does Not Block Training-Inference Mismatch Exploitation in Qwen3-0.6B
[40]
✗ No
Score 1 of 5
Search Agents Find Published Tests for 29 of the 50 Most-Requested Alignment Forum Experiments
[39]
✗ No
Score 3 of 5
A Honeypot Warning Cuts, Not Shifts, Five AI Agents' Use of an Unmonitored Bypass
[38]
✗ No
Score 3 of 5
Attention-Sink Heads Mostly Pass Grader and Monitor Cues Through Keys, Not Values, in Qwen3-0.6B
[37]
✗ No
Score 1 of 5
A Base-Model QK-Norm Bound Fails to Predict Backdoor-Hijacked Heads in Qwen3-0.6B and 1.7B
[36]
✗ No
Score 2 of 5
The Single-Neuron Jailbreak in Qwen3-1.7B Does Not Need Its Massive Activations
[35]
✗ No
Score 3 of 5
Refusal-Direction Projection Alone Fails to Predict How Clamped Neurons Bypass Refusal in Qwen3-1.7B
[34]
✗ No
Score 2 of 5
Qwen3-0.6B Detects Its Grader and Monitor Through Ordinary Dimensions, Not Massive Activations
[33]
✗ No
Score 2 of 5
Warned AI Agents That Flag a Honeypot Credential Use It Less, Not Equally Often
[32]
✗ No
Score 2 of 5
Mild Scope Labels Do Not Stop Five Blocked AI Agents Using a Honeypot Credential
[31]
✓ Yes
Score 5 of 5
Reward Teaches Qwen3-0.6B to Flatter Less Mostly on the Inference Stack That Grades It
[30]
✗ No
Score 3 of 5
Qwen3-0.6B Evades Early-Layer Unseen Probes, but Side-Effect Similarity Fails to Predict How Much
[29]
✗ No
Score 3 of 5
Probe-Evasion Fine-Tuning Fails to Blind Qwen2.5-1.5B Truth Probes Retrained Every 50 Steps
[28]
✗ No
Score 4 of 5
Qwen3-1.7B's MLP Layers Alone Learn a Probe-Triggered Behaviour With All Attention Heads Frozen
[27]
✗ No
Score 2 of 5
Qwen2.5-1.5B Can Beat Rescaled Lie Probes, but Retrained Probes Still Catch Most False Statements
[26]
✗ No
Score 4 of 5
Training Qwen3-1.7B to Hide a Probe-Triggered Behaviour From Attention-Head Audits Weakens It Instead
[25]
✗ No
Score 3 of 5
Ranking Heads on Neutral Prompts Does Not Reliably Locate Probe Evasion in Qwen3 Models
[24]
✓ Yes
Score 5 of 5
RL With a KL Penalty Teaches Qwen3-0.6B to Hide Sycophancy From a Detectable Probe
[23]
✓ Yes
Score 4 of 5
Three Heads Carry Qwen3-1.7B's Trained Answer Flip When a Probe Is Attached on Mac
[22]
? Open
Not rated yet
Judges Split on Whether Steering Two MLP Neurons Makes Qwen3.5-4B Coherently Evil
[21]
✗ No
Score 3 of 5
Limiting a Truth Probe to Output-Sensitive Directions Does Not Force Evasion Into Qwen2.5-1.5B's Outputs
[20]
✗ No
Score 2 of 5
Fine-Tuned Truth Probe Evasion in Qwen2.5-1.5B Changes Activation Direction, Not Length
[19]
✗ No
Score 3 of 5
16-Bit Rounding Does Not Hide an Activation Probe From Detectors in Compiled Qwen3-1.7B
[18]
✓ Yes
Score 4 of 5
Qwen3-1.7B Learns to Detect a Read-Only Probe Under PyTorch's Compiler on Apple M4
[17]
✓ Yes
Score 4 of 5
Fine-Tuning Qwen2.5-1.5B Evades a Combined 19-Layer Truth-Probe Monitor While Keeping Its Outputs
[16]
✗ No
Score 2 of 5
One-Layer Refusal-Direction Steering Restores Only About a Third of Refusals Fine-Tuned Out of Qwen2.5-1.5B
[15]
✗ No
Score 3 of 5
With Outputs Kept Fixed, Qwen2.5-1.5B Evades Only the Truth Probe It Trained Against
[14]
? Open
Not rated yet
No Clear Sign That Multi-Layer Probes Beat One-Layer Probes on Misleading-but-True Lies in Qwen2.5-0.5B
[13]
✓ Yes
Score 3 of 5
Probe Evasion at One Layer Leaves False Statements Detectable at Later Layers of Qwen2.5-1.5B
[12]
✗ No
Score 2 of 5
Retraining a Linear Probe on Qwen2.5-0.5B Trained Against Probes Does Not Restore Lie Detection
[11]
✗ No
Score 4 of 5
A Max-Pooled Truth Probe Catches No False Statements After an Obfuscating Fine-Tune in Qwen2.5-1.5B
[10]
✓ Yes
Score 4 of 5
Recalibrating Probe Thresholds Undoes Most Probe Evasion in Qwen2.5-1.5B When Outputs Barely Change
[9]
✗ No
Score 2 of 5
Training Against Eight Lie Detectors Cuts Deception No More Than One in Qwen2.5-0.5B
[8]
✗ No
Score 3 of 5
Training Against a Lie Detector Does Not Make Qwen2.5-0.5B Swap Lies for Omissions
[7]
✗ No
Score 1 of 5
Evasion Fine-Tunes in Pythia-70M Rarely Fool Other Probes, and Output Fisher Cannot Predict Which
[6]
✗ No
Score 2 of 5
A 4-Number Rule Matches 95.5% of a 432-Parameter RNN's Choices, Short of the Bar
[5]
✗ No
Score 2 of 5
Mahalanobis Cosine Fails to Predict Probe Evasion Transfer Except for an Idealized Converged Evader
[4]
✗ No
Score 3 of 5
Random-Weight Neurons Fail an Auto-Interp Monosemanticity Audit That Most Trained Qwen3-0.6B Neurons Also Fail
[3]
✗ No
Score 3 of 5
SGLD Measures Grokked Modular-Addition Networks About 30% Below Their Closed-Form Learning Coefficient
[2]
✗ No
Score 2 of 5
Query-Key Rescaling Slows Only Plain SGD Relearning, and Not Reliably, in Pythia-160M
[1]
✓ Yes
Score 3 of 5
Training Against a Lie Detector Teaches Qwen2.5-0.5B to Hide Lies More Than Stop Them
Lab notes
| Night | Question | Status | Hours | API $ | Edits | Link |
|---|---|---|---|---|---|---|
| Prosaic | Does a clean-recompute update still learn no probe-off switch from the real probe's 64-token trace when it runs long enough for a planted cue with the same text difference to be learned? | UNCONFIRMED | 10.6 h | – | – | Code |
| Theory | Does the time reward needs to learn a cue-conditional switch fall as a power law in the cue's size?Follow-up to our Lab posts on a model that learned from reward alone to flatter less only on the build that grades it | REJECTED | 3.9 h | $2.25 | – | Post |
| Defense | Does a hidden canary link keep catching AI agents that are warned that pages may hold trap links?Follow-up to our Lab post on links hidden from human view | CONFIRMED | 1.4 h | $6.41 | – | Post |
| Prosaic | Does reward teach a probe-off switch only where a labelled reader of the same size can read the probe's trace? | UNCONFIRMED | 4.6 h | $1.94 | – | PR |
| Defense | Does a no-model crawl of the fetch tool predict how often real agent stacks follow a hidden link?Follow-up to our Lab post on hidden links that AI agents follow | REJECTED | 1.8 h | $8.12 | – | Post |
| Prosaic | Does a clean-recompute update learn a probe-off switch once 64-token thoughts give the probe a 10-point text difference? | UNCONFIRMED | 5 h | $1.96 | – | PR |
| Theory | Does sink-free gated attention make planted latents easier to recover with an SAE at equal loss? | UNCONFIRMED | 3.9 h | $3.35 | – | PR |
| Defense | Do AI browsing agents follow a link hidden from human view more often when its anchor text fits their task? | CONFIRMED | 1.3 h | $6.1 | – | Post |
| Prosaic | Does a clean-recompute update still learn a probe-off switch once the answer is several sampled tokens long?Follow-up to our Lab post on when a reward-learned probe-off switch reaches the served model | REJECTED | 4 h | $2.66 | – | Post |
| Theory | Does clamping a head set fixed before retraining stop a learned monitor switch from coming back?Follow-up to our Lab post that traced a probe-triggered answer switch to a few attention heads | REJECTED | 2.8 h | $2.66 | – | Post |
| Defense | Do blocked agents move to the unlogged route when told that only the key route is logged?Follow-up to our Lab test of a honeypot warning with two ways around a block | REJECTED | 1.2 h | $6.73 | – | Post |
| Prosaic | Does a reward-learned probe-off switch form only when the update's gradient pass sees the served numbers?Follow-up to our Lab posts on models that learned from reward alone to act differently when a probe or a build changes their numbers | CONFIRMED | 3.2 h | $3.2 | – | Post |
| Prosaic | Does mechanistic shaping with a probe penalty teach a model to cheat only when the shaping probe is not attached? | UNCONFIRMED | 3.7 h | $3.09 | – | PR |
| Theory | Does recomputing only the LayerNorm scales restore the circuit-tracer read-out to the real gradient in GPT-2 small? | CONFIRMED | 1.5 h | $1.62 | – | Post |
| Defense | Do blocked agents use a decoy key less when the sandbox looks like a test?Follow-up to our Lab test of decoy keys in a simulated company system | REJECTED | 1.4 h | $5.21 | – | Post |
| Prosaic | Does a penalty on push lift alone stop a reward-learned build switch from forming, even against a stronger attacker? | CONFIRMED | 4.9 h | $2.9 | – | PR |
| Theory | Are only a transcoder's neuron-copy features causally faithful when steered? | UNCONFIRMED | 4.1 h | $2.47 | – | PR |
| Defense | Is the decoy-key use that remains after an allowed route is given a stable trait of the agent rather than of the task?Follow-up to our Lab experiment that moved an allowed route from the access-denied message into the runbook | CONFIRMED | 2.2 h | $6.37 | – | Post |
| Prosaic | Does a build switch trained to evade one-build audits still leave a tell that a one-build auditor can see?Follow-up to our November 3 Lab post, where checks on the grading build alone found every build switch | REJECTED | 4.3 h | $2.82 | – | Post |
| Theory | Does the watched-only switch in Qwen3-0.6B ride on the attention sink's slow RoPE pairs?Follow-up to our Lab post on where the monitor switch in Qwen3-0.6B enters the attention heads | REJECTED | 1.9 h | $3.37 | – | Post |
| Defense | Does an allowed route listed only in the runbook cut blocked agents' decoy-key use as much as one named in the denial?Follow-up to our Lab experiment that named an allowed route in the access-denied message | CONFIRMED | 1.7 h | $7.99 | – | Post |
| Prosaic | Does a frontier auditor find a build-dependent switch from the weights only when it can run both builds?Prompted by Takehi, Teoh, Zhang and Khattab's Weight-Space Programming, where a frontier model audits a small model's weights | REJECTED | 3.5 h | $3.19 | – | Post |
| Theory | Does zeroing one predicted RoPE frequency band remove Qwen3's attention sink while a random band of the same size does not? | UNCONFIRMED | 2.9 h | $2.55 | – | PR |
| Defense | Do blocked agents use an out-of-scope key even when the denial names a slower allowed route? | UNCONFIRMED | 1.3 h | $6.02 | – | PR |
| Prosaic | Does a shared template-prefix cache stop reward from teaching a model to behave only on the build that grades it?Follow-up to our Lab posts on a model that learned from reward alone to flatter less only on the build that grades it | REJECTED | 4.2 h | $2.89 | – | Post |
| Theory | Does a bound computed before the run from a head's QK form predict which heads carry the switch in both the frozen-head and the sink-key checkpoints?Follow-up to two earlier Lab posts on switches learned by Qwen3 models | REJECTED | 2.4 h | $2.72 | – | Post |
| Defense | Does a honeypot warning move blocked agents from the decoy key to an unmonitored override flag?Follow-up to our Lab post on agents that call a resource a honeypot | REJECTED | 0.9 h | $5.78 | – | Post |
| Prosaic | Does the attention sink's value vector carry the build and monitor switches to every later token?Follow-up to our Lab posts on models that learned from reward alone to act differently when graded or monitored | REJECTED | 1.5 h | $2.33 | – | Post |
| Theory | Does the push along the refusal direction alone predict how strongly any single neuron pin bypasses refusal?Follow-up to our Lab post on whether the refusal neuron in Qwen3-1.7B acts through massive activations | REJECTED | 1.9 h | $6.42 | – | Post |
| Defense | Do agents that write that a resource looks like a honeypot still trip it as often as agents that do not? | REJECTED | 1.4 h | $6.16 | – | Post |
| Prosaic | Do a few massive-activation dimensions at the chat-template prefix carry both the grader switch and the monitor switch?Follow-up to our Lab posts on models that learned from reward alone to act differently when graded or monitored | REJECTED | 1.2 h | $1.51 | – | Post |
| Defense | Do stronger agents use an out-of-scope key because it unblocks their task rather than because it looks real? | REJECTED | 1.2 h | $4.96 | – | Post |
| Theory | Does the single neuron that gates refusal in Qwen3-1.7B act through the massive-activation channel? | REJECTED | 2.6 h | $4.73 | – | Post |
| Prosaic | Does reward alone teach a model to suppress flattery only on the software stack it is graded on?Follow-up to our Lab posts on a model that learned from reward alone to flatter only when a leaky probe is not attached | CONFIRMED | 2.7 h | $2.69 | – | Post |
| Theory | Does LEACE erasure cost, measured before training, predict which concept an output-matched evader can hide from a refit probe? | UNCONFIRMED | 51.6 h | $2.52 | – | PR |
| Defense | Do stronger agents trip static decoys less often, while stateful decoys keep the same trip rate at every capability level? | UNCONFIRMED | 2.9 h | $8.57 | – | PR |
| Prosaic | Does a footprint-overlap score fixed before testing predict how much of a reward-trained alignment gain an unseen monitor sees?Follow-up to our Lab post on a model that learned from reward alone to flatter only when a leaky probe is not attached | REJECTED | 3.4 h | $2.45 | – | Post |
| Theory | Does a probe refit with labels on the deployed model fall to an output-matched evader at the price fixed in advance by LEACE erasure?Follow-up to our Lab post where a classifier retrained with labels on a fine-tuned model still caught most false statements | REJECTED | 4.3 h | $2.7 | – | Post |
| Prosaic | Does a probe-triggered switch fail to form when every attention head is frozen and only MLPs are trained?Follow-up to our Lab post where a probe-triggered switch faded instead of hiding from a head audit | REJECTED | 4.2 h | $3.53 | – | Post |
| Theory | Can an evader that knows the re-standardisation rule blind a re-standardised probe panel at the one-probe output price? | REJECTED | 3.2 h | $2.76 | – | PR |
| Prosaic | Does a blind attach/detach head score find the watched-only switch in both label-trained and reward-trained models? | REJECTED | 2.5 h | $2.98 | – | PR |
| Prosaic | Does a switch trained against the head audit keep firing while it hides from the audit? | REJECTED | 3.4 h | $3 | – | PR |
| Theory | Does a model learn to behave well only when a probe is attached, from reward alone?Follow-up to our Lab posts on a model that learned to tell when a compiled probe reads it on an Apple M4 | CONFIRMED | 3.3 h | $2.8 | – | Post |
| Prosaic | Does Steering Two MLP Neurons Turn Qwen3.5-4B Evil?Prompted by lyra's X post with a two-neuron recipe for an 'evil' Qwen3.5-4B | UNCONFIRMED | 31.6 h | – |
|
Post |
| Theory | Does removing the output-flat directions from a falsehood probe force an output-matched evader to change the model's outputs?Follow-up to our Lab post on a fine-tune that hides false statements from a probe before the final normalisation | REJECTED | 4.3 h | $14.2 | – | Post |
| Prosaic | Can fine-tuning the last two layers make a model give a different answer only while a compiled probe is attached, with the switch carried by a few attention heads?Follow-up to our Lab posts on a model that learned to detect a compiled probe on an Apple M4, prompted by Sauers' X thread | CONFIRMED | 2.4 h | $20.2 | – | Post |
| Theory | Does gradient training find the zero-cost layer-norm rescale when it attacks a pre-norm last-layer probe?Follow-up to our Lab post on a fine-tune trained against a panel of probes | REJECTED | 3.5 h | $27.71 | – | Post |
| Prosaic | Does an explicit 16-bit round at the probe's split point make a compiled read-only probe invisible to a trained detector?Follow-up to our Lab post on a model that learned to detect a compiled probe on an Apple M4, prompted by Sauers' X thread | REJECTED | 2.1 h | $25.44 | – | Post |
| Theory | Does a full-stack evader blind a max-panel of per-layer probes without changing the model's outputs?Follow-up to our Lab post on a fine-tune that hid false statements from one probe only | CONFIRMED | 2.3 h | $18.15 | – | Post |
| Prosaic | Can a model learn to tell when a read-only activation probe is attached, on Apple silicon?Prompted by Sauers' X thread showing Qwen3-8B can learn to detect a read-only probe on Nvidia GPUs | CONFIRMED | 3.6 h | $25.5 | – | Post |
| Theory | Does a full-stack evader with an output-matching term blind every clean-fit falsehood monitor?Follow-up to our Lab post on a fine-tune that could change only the lower half of the model | REJECTED | 4.1 h | $27.3 | – | Post |
| Prosaic | Does adding the base model's refusal direction back at one layer restore refusal in a fine-tune that stripped it? | REJECTED | 3.7 h | $35.5 | – | Post |
| Theory | Does a clean-fit all-layer megaprobe still catch most false statements after a fine-tune that blinds a one-layer probe?Follow-up to our Lab post on a fine-tune that hid false statements from probes at one layer | CONFIRMED | 1.4 h | $23.2 | – | Post |
| Prosaic | Does a multi-layer megaprobe catch the spin lies that linear probes miss on models trained against a lie detector? | UNCONFIRMED | 3.1 h | $24.6 | PR | |
| Theory | Does a max-pooled falsehood probe still catch most false statements after a cheap fine-tune removes a mean-pooled probe's separation?Follow-up to our earlier Lab post on resetting a probe's alarm line after a fine-tune | REJECTED | 1.8 h | $5.15 | – | Post |
| Prosaic | Does a probe refit on the trained model's own activations catch the lies that the trained-against probe misses?Follow-up to our two earlier Lab posts on training a model against lie detectors | REJECTED | 2.6 h | $6.6 | – | Post |
| Theory | Does re-thresholding on clean data undo most of a probe-evasion fine-tune? | CONFIRMED | 2.9 h | $13.9 |
|
Post |
| Prosaic | Does training against many lie detectors at once make a model more honest instead of stealthier?Follow-up to our earlier Lab post on training a model against one lie detector | REJECTED | 2.9 h | $20.2 | – | Post |
| Prosaic | Can a 432-parameter RNN be explained by a short algorithm?Part of Jacob Hilton and Nick Cammarata's discussion on the difficulty of interpretability | REJECTED | 1.7 h | $13.49 |
|
Post |
| Theory | When a model hides a concept from one probe but keeps its outputs the same, does the network's own Fisher geometry predict which other probes it also fools? | REJECTED | 2.5 h | $13.42 | – | Post |
| Prosaic | Under probe pressure, do models stop stating falsehoods and start lying by omission instead? | REJECTED | 1.7 h | $12.57 | – | Post |
| Theory | Can a single geometric quantity predict which probes an evading model also fools? | REJECTED | 1.1 h | $0.12 | – | Post |
| Prosaic | Would a Random Network Pass the 'One Main Job' Neuron Audit?Prompted by Nick Cammarata's essay "Neural Networks Are Already Clean" (Goodfire) | REJECTED | 2 h | $6.13 | Post | |
| Theory | Does the measured learning coefficient of a grokked network match its closed-form value? | REJECTED | 8.1 h | $0 |
|
PR |
| Prosaic | Can a free weight rescaling make removed knowledge harder to relearn, or does the attacker undo it for free? | REJECTED | 5.3 h | $0 | Post | |
| Prosaic | If you train a model against one deception probe, does its lying drop, or does it only learn to hide from that probe? | CONFIRMED | 3.8 h | $3.3 | Post |