A plausible answer is not the same as a grounded answer.
Superretometer is a cognitive harness that makes reasoning visible, traceable, and testable — so you can ask: what does this conclusion depend on, and what makes those dependencies real?
See a worked example · Quick start · Runnable example on GitHub
An AI system or engineering simulation can produce a result that looks correct. The numbers close. The chain of reasoning appears sound. Yet the conclusion rests on premises that have no independent grounding in the physical world.
Consider a reasoning chain from an engineering model:
The chain can be structurally valid and still not be independently grounded. Superretometer makes that distinction explicit, structural, and computable.
Utility first. Here is the pattern Superretometer is built for — claim → missing grounding → required action.
A simulation reports that a machine reaches the desired operating speed (RPM). Energy balance closes. Several checks pass. The result looks ready to carry a design decision.
The dependency that blocks warrant is explicit: a fitted parameter
whose training target is the same conclusion the model is asked to
justify (CALIBRATED_TO_CONCLUSION).
Hitting the desired RPM is not the same as independently predicting it. If a free parameter was tuned until the simulation matched the target, the match is not independent evidence. Plausible is not grounded. A check that cannot fail is not evidence.
Register an independent measurement or test dataset for that parameter (or remove the claim from design-bearing status until one exists). Then re-run warrant analysis. The four decision gates below stay closed until the missing grounding is present.
Runnable counterpart:
examples/engineering_model/
·
Tests:
tests/
A simulation result is a file on disk. It is not a conclusion. The four-gate SimulationGatePolicy determines whether a simulation-backed claim may carry a design decision.
All four gates must PASS for design-bearing. One BLOCK means informative only — not that the simulation failed, but that the warrant chain is incomplete.
Superretometer evaluates documented dependencies and evidence under explicit rules. It inspects the provenance graph you register: what depends on what, which roots are independent, which checks can fail, and which anti-patterns are present.
Package version: 0.7.0 (pyproject.toml). Latest published release notes: v0.6.4.
Spec:
spec/
·
Implementation:
cognitive_harness/
·
Examples:
examples/
·
Tests:
tests/
Naming note: the Python package is cognitive-harness
(import cognitive_harness). The project brand is
Superretometer; Cognitive Harness is the specification name.
Full quick start and API sketch: README on GitHub.
Superretometer combines generative reasoning with persistent structure and an explicit reality check:
LLM or engineer forms hypotheses and candidate justification chains.
Emergent Memory keeps evidence, decisions, and observations across sessions.
CAFCR / Lines of Reasoning structure each step: claim, viewpoint, knowledge used.
Inspired by the EPR completeness criterion as an epistemic analogy — not a physics claim.
Longer academic background lives here and in linked notes. Lead with use-value above; this section is attribution and lineage. Detailed notes: background.md, muller-cafcr.md, epr-completeness.md, hoh-comparison.md.
The work originated through the contributions of Nikolai Fasting and Maciej Kucharz at Emergent Company, who built Emergent Memory and Emergent Strategy — the persistent reasoning and adaptive planning substrate this harness runs on.
Emergent Memory provides the persistent graph backbone: knowledge, evidence, decisions, and observations survive across sessions.
Emergent Strategy closes the adaptive loop: act → observe → learn → revise → act again. Developed by Nikolai Fasting and Maciej Kucharz at Emergent Company. Emergent Strategy is grounded in Nikolai's experience building over 30 startups and inspired by Ready, Fire, Aim (Michael Masterson, 2008): bias toward action, learn from outcome, iterate.
CAFCR expands to Customer Objectives, Application, Functional, Conceptual, Realization. Müller's PhD thesis, CAFCR: A Multi-view Method for Embedded Systems Architecting; Balancing Genericity and Specificity (TU Delft, 2004), introduced structured multi-view architecting and Lines / Threads of Reasoning — tracing decisions from stakeholder need through viewpoints to realization so the reasoning chain can be audited and restarted.
Superretometer's Thread model computationally applies that principle: each thread records claims, viewpoints, and knowledge objects used.
Architectural Reasoning; CAFCR · CAFCR Thesis PDF · Overview of CAFCR and Threads of Reasoning
Can Quantum-Mechanical Description of Physical Reality Be Considered Complete? (Physical Review, 47, 777, 1935) proposed a criterion of reality and argued that a complete theory must account for every element of physical reality.
Superretometer adapts this as an epistemic analogy for engineering reasoning: a conclusion is warranted only if the quantities and claims carrying it possess independently grounded counterparts outside the model.
This is an analogy, not a claim about quantum mechanics. We do not apply quantum physics to engineering problems. We apply the structural question EPR asked about the relationship between description and reality.
The Harness-of-Harness framework (Yan et al., 2026) is supporting literature on structuring a harness around an LLM — for state preservation, progressive disclosure, and long-running orchestration.
It is not the conceptual origin of Superretometer, and it does not validate Superretometer's warrant mechanisms or architectural claims. Related reading, not empirical confirmation of this project.
Superretometer is inspired by and builds on these traditions. It is not affiliated with or endorsed by the original authors or organizations.