Hermes Labs is an experimental AI research laboratory studying human cognitive behavior and attempting to reproduce particular cognitive acts in artificial systems. We investigate their observable functions, develop computational accounts, and test whether machines can perform them with measurable fidelity.
How do human beings transform what they already know into something new, without materially new external information?
Defining and measuring innovative transformation. Foundational questions remain under stress-testing.
How does the mind construct and reason about possibilities that are not presently given?
Developing the capability definition and ways to evaluate it.
What happens when thought wanders without an immediate external demand?
Forming the concept; its boundaries and computational interpretation remain open.
How does affect shape what we notice, value, remember, and do?
Defining a measurable regulatory function. No implementation or result is established.
Research questions and working targets—not demonstrated capabilities.
Imagine someone steadying a wobbly table with a folded piece of discarded cardboard. Material made for packaging becomes a support, and the person has a workable solution using what is already at hand.
We can observe the changed use of the material and whether the table becomes stable. What conditions made that response possible, and which features of the act would an artificial system need to preserve?
A computational account has to make these properties testable. Experiments can vary the materials, constraints, or circumstances and examine whether a candidate system responds appropriately beyond one familiar case.
An artificial system’s behavior can support a claim of functional reproduction within specified conditions; correspondence with a human biological mechanism is a further question.
We begin by isolating a tractable phenomenon and defining the behavior a computational system would need to reproduce. Candidate implementations are evaluated against those criteria and appropriate comparisons; successes and failures guide revisions to the account. Controlled interventions help test which features of the system explain the observed behavior.
Existing human behavioral evidence informs this work, alongside computational cognitive science, cognitive architectures, interpretability, modular learning, artificial life, and computational creativity. Prior-art context ↗
Seven stages express an iterative research framework. Findings can return us to a definition, an explanation, or a control. Further recombination, composition, and evolution depend on what reproduction and intervention actually establish.
Which human cognitive act can we investigate?
Identify a tractable human cognitive act or function, establish its boundaries, and distinguish neighboring phenomena that may contribute to it.
A capability must first become a recognizable object of investigation. In invention, novel ideas, useful solutions, remembered knowledge, analogies, and changes in problem representation may arise from overlapping operations.
Isolation constructs a research question precise enough to investigate. Studying a faculty separately is a methodological choice; its independence from other functions remains a hypothesis.
Distinguish the phenomenon from the mechanism proposed to explain it. State which neighboring phenomena the investigation must discriminate.
How can we recognize and investigate the capability?
Translate the phenomenon into measurable functions, competing explanations, and experimental conditions capable of distinguishing among them.
Relate the proposed construct to existing human behavioral evidence. Construct validity asks whether the task and its measurements actually capture the intended capability. Independent evaluation, explicit information boundaries, controls, and failure criteria make that question testable.
Investigate existing theories before selecting an engineering approach. Retrieval, analogy, search, representation change, or their coordination may explain the behavior. Rival hypotheses should predict observations that can distinguish them; the operational definition must remain independent of the favored mechanism.
Define evaluators independently of the signals or mechanisms being tested. Record alternatives, controls, justified success and failure criteria, and the decision each outcome would support.
Can the capability be instantiated artificially?
Construct computational systems that exhibit the operationally defined functions, and evaluate their behavior against independent standards.
The scientific question determines the engineering approach: an existing model, a computational architecture, targeted training, a structured inference procedure, or a new mechanism may be appropriate.
Functional reproduction requires reliable behavior across the specified conditions, appropriate baselines, and held-out evaluation. Mechanistic reproduction makes a stronger claim about correspondence with the human process and requires separate evidence. One successful output establishes neither.
Retain reproducible inputs, versions, resource budgets, outputs, and limitations. State the scope of functional success separately from evidence for human-mechanism correspondence.
What changes when we intervene on the machinery?
Alter candidate computational mechanisms, representations, resources, or interactions and measure the resulting changes in capability.
Artificial systems permit interventions on memory organization, information access, selection procedures, representations, and component interactions. Removing, constraining, amplifying, or reorganizing an element can expose dependencies and tradeoffs.
Controlled comparisons and ablations test predicted consequences. A behavioral difference can have several causes, including collateral changes in compute or information; the design must establish which causal interpretation is warranted.
Specify the intervention, matched comparisons, expected effects, and alternative causes. Interpret a causal effect within the tested system and conditions.
What becomes possible when investigated capabilities interact?
Bring cognitive functions or computational mechanisms into new relationships and study the consequences of their interaction.
Imagination may influence planning; memory may shape invention; affect may alter attention and valuation. These relationships motivate experiments rather than assumptions of interchangeable capabilities.
Compare components alone and in combination. Different arrangements may produce complementarity, interference, unexpected dependencies, or functions absent from the individual components. Recombination investigates these interaction effects.
Keep component contributions identifiable and test gains and interference against appropriate individual and combined baselines.
Can investigated capabilities form a coherent cognitive architecture?
Integrate compatible computational capabilities into larger artificial systems and investigate whether their organization supports sustained cognitive function.
Composition concerns system-level coordination: shared interfaces, compatible states, memory access, and resolution of competing outputs. It builds on investigated interactions while asking whether the resulting organization works as a whole.
Evaluate the integrated architecture across sustained tasks, changing conditions, and component failures. Individually functional parts do not guarantee coherent cognition. This is a future-facing engineering objective for Hermes Labs.
Evaluate system-level reliability, coordination and attribution, retaining ablations and resource-matched comparisons. No complete synthetic mind is established by current Hermes work.
What other architectures of cognition might become possible?
Investigate how constructed cognitive systems can adapt, undergo systematic modification, or develop new functional configurations.
An evaluated architecture becomes a new experimental object. Successive designs can alter its components and organizational principles, asking which configurations remain coherent and which acquire useful functions.
Learning, deliberate redesign, computational search, and evolutionary computation are different processes. Each implementation must specify which process produces change and how it is evaluated. Biological evolution and developmental learning require their own models and evidence.
Use Evolve as an experimental horizon. Define the actual adaptation or modification procedure; evaluate retained and lost functions without treating design trajectories as forecasts of biological evolution.
Observing a cognitive act, reproducing its function, and explaining its mechanism are distinct scientific tasks. Our investigations therefore separate the phenomenon, its measurement, the mechanisms proposed to explain it, and the effects of interventions upon those mechanisms.
Experiments must accommodate competing explanations. Independent evaluators establish what an observed result warrants; records retain conditions, baselines, deviations, and limitations. A result that defeats a favored hypothesis can reveal what the previous explanation failed to account for.
A working instrument demonstrates behavior within its tested contract. Claims about cognition require evidence connecting that instrument to the proposed capability. Human judgments, model behavior, mechanistic attribution, and practical usefulness each carry distinct evidentiary obligations.
Project Fiction begins with a human question: how do we transform available information into novel, actionable possibilities? A new combination may be useless; a useful solution may have been retrieved. A changed problem representation may reveal an opportunity, or introduce a more convenient objective.
Finite computational environments make particular measurement questions inspectable. PT-001 used resource-allocation worlds to compare constructed reformulation and search policies. Independent measures of novelty and effectiveness remain a wider design obligation; this apparatus addresses a narrower question.
An apparent advantage of the constructed reformulation strategy disappeared under an identical-order search-only comparison.
In PT-001’s deterministic rehearsal, search alone reproduced the strategy’s outcomes across the tested conditions. Improved outcomes therefore could not identify a distinct cognitive mechanism of problem transformation: the tested operationalization did not support that attribution.
Local measurement rehearsal · 23 October 2026 · No human or frontier-model inference observations.
Completed local measurement rehearsal · 23 October 2026 · Constructed deterministic strategies; no human or model-inference observations.
Can these outcome measures distinguish a benefit of changing problem formulation from a benefit of objective optimization or search?
Four finite resource-allocation worlds specify legal actions and explicit value criteria. Feasibility and value are scored; novelty is deliberately unscored. The value criteria are assumptions of the toy task.
The local rehearsal compared constructed policies and an identical-order search-only ablation: 4,372 bounded sequences and 480 strategy records. Candidate-evaluation ceilings were matched; total compute was not established as equal.
The search-only ablation reproduced the reformulation strategy’s candidates and final allocation in every tested world, permutation, and budget. Better outcomes did not identify a distinct formulation mechanism.
This operationalization cannot attribute the gains specifically to reframing. Further investigation needs independent formulation evidence and stronger objective/search controls. Whether cognitive problem transformation contributes remains open.
CWR supplies complementary offline evaluation apparatus; three empirical designs remain proposed.
Explore Project Fiction ↗Hermes appears at the threshold. The messenger crosses between worlds, carrying meanings from one domain into another. The trickster disturbs familiar arrangements, exploiting the possibilities that exist where established boundaries meet.
For this laboratory, the first crossing runs from a human cognitive act as observed to its attempted reconstruction in an artificial system. What can survive that crossing with sufficient functional fidelity?
A faculty takes its form within an inherited biological organization. Its relationships with memory, attention, embodiment, and other capacities help determine what it can do. Attempting to reproduce its function elsewhere asks which relationships the act requires, and which belong to the organization in which we first encountered it.
The messenger’s task is not fulfilled by arrival alone. The question is what has been carried across. Experiments must establish whether the artificial performance preserves the investigated function, and where the translation fails.
We call this orientation the Hermean Function. If reproduction gains sufficient experimental support, translation may open a further inquiry: which dependencies can be varied, which interfaces reconnected, and which new arrangements remain viable? Transformation, recombination, and composition become possibilities to investigate through the evidence earned at the crossing.
This philosophical interpretation draws on Lewis Hyde’s Trickster Makes This World. The computational questions developed here belong to Hermes Labs.
The human mind provides our first field of investigation. Artificial reproduction is the present challenge.
If particular cognitive functions can be reproduced with adequate fidelity, their artificial implementations may offer further experimental possibilities. We could ask what changes when a dependency is altered, a constraint relaxed, or a capability brought into a different relationship with another.
Modification, recombination, and extension would then become research questions grounded in what had survived reproduction. Their feasibility, and the forms of cognition they might support, remain to be discovered.
Our longer ambition begins with that conditional possibility. Each further crossing must earn its own evidence.
We intend to investigate what the evidence makes possible.
Method account · 24 October 2026 · Canonical method record