Non-model capabilities · Nonprofit · 2026
What a deployed system can do that its model cannot.
Static evals measure the checkpoint. The interesting capabilities were built on top of it.
A frontier lab trains and evaluates a model. What gets deployed is a system: the model plus adapters and fine-tunes, attached data, a harness or control environment, other agents, and a user who keeps interacting with it. That system has capabilities the checkpoint alone does not have. They were not trained in, they are not on the model card, and no benchmark measures them. Redarc Labs builds the cases where the answer is known and checks whether the instruments can find it.
Our thesis
A capability that was built on top only shows up when you intervene.
Vary the dose, add a sham arm, compare against a case where the answer is already known, and commit the verdict rule before the data exists. Four claims, each resting on a paper below.
01 Capabilities get added after release, and can be hidden.
An adapter installed a habit the checkpoint never had, a second adapter hid it behind baseline behaviour, and seven state-reading instruments could not tell the two apart.
Rests on · Known by Construction
02 Some capabilities exist only between components, not inside any one model.
Steer one agent's affect and a second agent's activations move with the dose. Neither model has that behaviour alone.
Rests on · Steered Affect
03 The instruments meant to read model internals have an unmeasured false-positive rate.
Probes, concept directions, and SAE features report confident findings on untrained networks, and fixed-dose steering audits mis-rank models for a reason unrelated to steering.
Rests on · Dead Salmons; Distance to the Boundary
04 How a model behaves inside a harness is decided by the harness as much as by the model.
Model organisms with objectives inserted at different training stages, run inside monitored and tool-restricted environments, to see what the overseer can and cannot detect.
Direction
Papers
Five papers at NeurIPS 2026 workshops. Workshop papers are non-archival.
Known by Construction: Calibrating Model-State Instruments Against Models Whose State Was Installed, Hidden, or Never Present
Manan Wadhwa*, Shivam Dubey*
Seven instruments that claim to read a model's internal state were tested on models where the state was installed, hidden, or absent by construction. Every one read the prompt or the training history instead.
Steered Affect Moves a Partner Agent's Activations, and the Logistic Probe That Reads It Returns a Tie-Break, Not an Estimate
Manan Wadhwa*, Shivam Dubey*, Debojeet Ghosh, Shreyansh Kumar, Ayushman Garg
One agent's steered emotional state shows up in a second agent's activations. The standard probe used to read it is a tie-break, not a measurement, and the standard ablation control is too weak to notice.
Distance to the Boundary Is Not Steering Resistance
Manan Wadhwa*, Shivam Dubey*
Fixed-dose steering audits rank aligned models as least controllable because the model's margin grew, not because the steering direction weakened.
Dead Salmons Have a False-Positive Rate: A Calibrated Null-Model Test for Interpretability
Manan Wadhwa*, Shivam Dubey*
Probes, concept directions, and SAE feature selection report confident findings on networks that learned nothing. The shuffled-label control everyone runs does not catch it.
Doubt Is a Decision: A Sparse Circuit Commits Language Models to "I Don't Know"
Sheryl Mehta*, Shagun Chadha*, Kanika Singhal*, Manan Wadhwa, Shivam Dubey
The decision to hedge lives in a small set of heads and neurons. Correlation-selected "confidence neurons" fail causal tests against a random-neuron null.
* co-first author
Earlier work
Shivam Dubey*, Manan Wadhwa*
Shivam Dubey, Jason Hoelscher-Obermaier
Shivam Dubey
Open questions
Behaviour inside control environments
Does a model with an installed habit behave differently under monitoring and restricted tools than outside them, and can the overseer tell? A control environment makes the answer known by design, which is what the lab's method needs.
Persistence of hidden objectives
Which training stage makes an inserted objective survive retraining, distillation, and prompt variation? The model organisms come from the founders' MARS fellowship work at the Cambridge AI Safety Hub.
Capabilities between agents
Affect transfers between agents. What else does, and at what dose?
A calibrated instrument suite
Every instrument the lab has tested fails on a null it did not run. The long-term goal is a set of state-reading tools that ship with their own false-positive rate.
Founders
Shivam Dubey
Co-Founder
BS Data Science, IIT Madras, 2023 to 2027
Manan Wadhwa
Co-Founder
B.Tech Computer Science, MAIT (GGSIPU), expected 2027
Collaborators
Sheryl Mehta · Shagun Chadha · Kanika Singhal · Debojeet Ghosh · Shreyansh Kumar · Ayushman Garg. Co-authors on the two papers that came out of the first cohort.
Teaching
The lab teaches an eight-session curriculum and runs it with cohorts. Two of the papers above came out of the first cohort.
Community