Scientific record
Purpose and scientific boundary
Hermes Labs is a data and AI research lab building specific models and synthetic cognitive capabilities. Its method asks whether capabilities that contributed to human cognitive power can be isolated, reproduced, varied, and composed, yielding knowledge about cognition as well as engineered systems. Better task scores alone do not establish the lab's thesis. The long-term question is what kinds of minds can exist; the immediate obligation is to make each narrower claim testable. [S33–35, S38]
Artificial systems serve as experimental organisms for cognitive science: environments in which researchers can change a capability or its coupling to others and examine consequences. This is a research framing, not evidence of biological equivalence, human-equivalent mechanisms, consciousness, feeling, or subjective experience. Results concern the specified artificial system and task conditions; transfer to human cognition requires an independent argument and evidence. Hermes seeks mechanisms that made human cognition evolutionarily powerful and possible combinations beyond biological constraints, rather than a wholesale replica of a human mind. [S34, S36–38]
Canonical research ladder
isolate -> operationalize -> reproduce -> perturb -> recombine -> compose -> evolve [S38]
| Stage | Research obligation |
|---|---|
| Isolate | Identify a sufficiently distinct cognitive function and explain why studying it independently is useful. Distinctness is a hypothesis, not an assumption. |
| Operationalize | Specify tasks, information boundaries, observable criteria, adjacent phenomena, alternatives, controls, and conditions under which the claim fails. |
| Reproduce | Demonstrate the operational capability in an artificial system against appropriate comparisons, with reproducible records and explicit limits. Similar output alone is insufficient. |
| Perturb | Manipulate a candidate cause—representation, memory, constraints, incentives, search, affect, or architecture—and measure predicted changes and tradeoffs. |
| Recombine | Combine capabilities and test interaction effects, including interference, rather than assuming independent gains add together. |
| Compose | Study increasingly integrated architectures while retaining attribution, ablations, and evidence about component contributions. |
| Evolve | Explore changed configurations and new forms of cognition; treat evolutionary trajectories as model-dependent possibilities, not forecasts of actual human evolution. |
These are program stages, not a rigid one-pass execution schedule. Operationalization, evaluation, reproduction, and perturbation iterate. The conversation first proposed decompose → replicate → measure → perturb → recombine → compose → evolve [S35], then made the seven-stage ladder explicit [S38]. Its later pilot workflow—capability → operational definition → controlled evaluation → baselines → failure taxonomy → mechanistic hypothesis → intervention → measured improvement [S42]—describes experimental execution. isolate → operationalize → evaluate → perturb → explain [S43] and the later operational sequence adding baselines [S58] remain historical workflow formulations, not replacement canonical ladders.
Functional mechanism, not anthropomorphic theater
Anthropomorphic theater means surface performance of a human cognitive state without evidence that its underlying function is present. “I feel anxious” does not establish affective regulation; imaginative prose does not establish possible-world construction; novel wording does not establish useful recombination; expressed uncertainty does not establish calibration. [S38–39]
Each project therefore needs behavioral criteria and a mechanistic hypothesis that can be challenged through manipulation. A behavioral effect supports the measured function within scope; causal intervention strengthens attribution but does not establish complete mechanistic identity with humans. Reports must say which claim is supported. Project Affect specifically targets synthetic mechanisms regulating attention, salience, valuation, memory, motivation, risk perception, and action selection; emotional self-description is not its success measure. [S36–39]
Evidence and experiment discipline
The existing workspace adds the practical standard: every proposal records hypothesis, operational definition, alternative explanation, control, success criterion, failure criterion, and next decision. Freeze justified numerical criteria and analysis before examining relevant outcome data. Record actual inputs, versions, prompts, conditions, resource use, seeds where applicable, outputs, deviations, and limitations. Preserve failures and inconclusive findings. No sample size, effect threshold, architecture, or success result is selected by this canon integration.
Retrieval is part of cognition's substrate. Fiction may distinguish operational retrieval, deduction, and recombination conditions without claiming mutually exclusive mental kinds or a simplistic construction-versus-retrieval ontology. Deduction can participate in constructing a recombination; useful transformation may operate on retrieved material. Test alternatives rather than naming task conditions as proof of mechanism.
Keep human construct/label validity, local signal, domain transfer, model capability, runtime usefulness, reward suitability, and human-to-model transfer as separate claims. Existing Fiction decision gates remain authoritative for their corresponding branches. Classifier accuracy does not establish a cognitive mechanism, runtime value, or reward suitability. Offline apparatus tests are engineering evidence, not human or model capability results.
What the first pilot must teach
Fiction Experiment 001 / Micro1 tests both Project Fiction and this method. Its required scientific artifacts are a defensible task family, failure taxonomy, evaluation protocol, intervention result, and methods memo explaining what was learned about studying cognition artificially. An honest collapse of the initial formulation can be useful science. Neither a hackathon win nor a persuasive demonstration substitutes for those artifacts. [S41–44]
2026-10-24 — Methodological and editorial clarification
Status: adopted descriptive orientation and methodological commitments; empirical validity remains unproven.
The user's Our Method brief adopts Experimental Cognitive Engineering as the name for Hermes Labs' methodological orientation. The prior October 14 candidate-name status is preserved in the chronology and decision history. This adoption describes the lab's approach; it establishes neither a new discipline nor scientific priority. Existing computational cognitive science, cognitive architectures, interpretability, modular learning, artificial life and computational creativity are relevant traditions. A specific contribution needs evidence and capability-specific comparison.
The seven stages remain unchanged. Their scientific obligations are clarified:
- Isolate: distinguish phenomenon and proposed mechanism. A separate program is not evidence of an independent cognitive module.
- Operationalize: examine existing explanations, construct validity and rival predictions before selecting an architecture. Independent evaluation, information boundaries, controls and failure criteria are essential.
- Reproduce: demonstrate reliable operational function under specified and held-out conditions, with baselines and reproducible records. Functional reproduction and correspondence with a human mechanism require different evidence. One successful output is insufficient.
- Perturb: manipulate candidate causes and test predicted effects with controls and ablations. Collateral resource/information changes can limit causal attribution.
- Recombine: investigate interactions among identifiable functions or mechanisms, including interference and complementarity.
- Compose: investigate integrated system coordination, interfaces and sustained behavior. Component success does not establish system coherence. This remains a future engineering objective.
- Evolve: investigate changed configurations; explicitly distinguish learning, deliberate redesign, search and evolutionary computation. Biological evolution and developmental learning are separate mechanisms, not synonyms for software iteration.
Current Fiction work supports conceptual investigation, measurement design and offline apparatus. PT-001 is a completed local measurement rehearsal: an identical-order search-only strategy reproduces the formulation strategy's outputs, so this implementation does not identify a reframing-specific benefit. Human/model cognitive reproduction, integrated minds and the entire ladder's completion are not established. The public trace uses revised measurement claim, rather than asserting that this rehearsal refuted the broader cognitive hypothesis. See PT-001.
The method iterates when evidence challenges a definition, mechanism or control. Independent evaluations separate phenomenon, measurement, mechanism, intervention and practical usefulness. No experiment is authorized or executed by this editorial clarification.
The lab's philosophical working description is the Hermean Function: transformation of existing capabilities through crossings between representations, mechanisms and systems. Its role is orientation; experimental work must establish whether a function survives translation and what changes in its organization. See philosophical interpretation (not included in this public selection: hermean-function.md) and DEC-0009.