Bounded definition
The cited source supports treating neural feature superposition as a distinct concept within the stated mechanistic interpretability scope. Within this page, that proposition is limited to Limited to Definitions, toy models, geometry, sparsity, and feature-interference experiments. in “Toy Models of Superposition”; this candidate records the concept boundary and does not pool results from uncited systems or studies.
Definition and evidence boundary
A source-bounded concept record for neural feature superposition within mechanistic interpretability. The bounded proposition retained by the canonical record is: The cited source supports treating neural feature superposition as a distinct concept within the stated mechanistic interpretability scope.
The applicable scope is Limited to Definitions, toy models, geometry, sparsity, and feature-interference experiments. in “Toy Models of Superposition”; this candidate records the concept boundary and does not pool results from uncited systems or studies. This definition must not be generalized beyond the cited source and exact record boundary.
Claims: urn:maha:claim:mechanistic-interpretability-neural-feature-superposition
Mechanism and technical context
The work develops toy models in which neural networks represent more features than available dimensions under specified sparsity conditions. This is the source-bound technical context for the record; no uncited mechanism is added by the compiler.
Neural feature superposition does not by itself establish system-level performance, safety, manufacturability, scalability, economic advantage, clinical benefit, or deployment readiness. The mechanism or method is therefore presented as one component of a larger system, not as evidence for every downstream outcome.
Claims: urn:maha:claim:mechanistic-interpretability-neural-feature-superposition
How to interpret the evidence
No cross-source quantitative interval is asserted. Definitions, operating conditions, samples, instruments, and outcome measures must be checked against the exact cited locator during review. The evidence maturity recorded here is single study, and the claim kind is theoretical model.
Independent replication and cross-platform transfer have not been compiled for this candidate; the evidence maturity refers only to the bounded source contract. A toy-model mechanism does not establish that every feature in a production model has the same geometry or semantics. These qualifications travel with the claim whenever it is reused.
Claims: urn:maha:claim:mechanistic-interpretability-neural-feature-superposition
What the source supports and what remains unknown
The inspected source supports exactly this: The work develops toy models in which neural networks represent more features than available dimensions under specified sparsity conditions. It was read at Definitions, toy models, geometry, sparsity, and feature-interference experiments.
What remains unknown is everything outside that locator. Neural feature superposition does not by itself establish system-level performance, safety, manufacturability, scalability, economic advantage, clinical benefit, or deployment readiness. No quantity, comparison, or downstream outcome is established here unless a separately scoped record measures it.
Claims: urn:maha:claim:mechanistic-interpretability-neural-feature-superposition
Source identity, locator, and reuse boundary
The bound source is “Toy Models of Superposition” by Nelson Elhage, Tristan Hume, Catherine Olsson, et al., published by Transformer Circuits Thread on 2022-09-14; its declared stable identity is url:https://transformer-circuits.pub/2022/toy_model/index.html.
The inspected-content locator is Definitions, toy models, geometry, sparsity, and feature-interference experiments. Reuse is limited to citation-with-paraphrase. The candidate uses original boundary language and a short paraphrase linked to the cited source. No source passage, figure, or table is reproduced. This metadata establishes source identity and inspection scope, not the truth of claims outside the cited locator.
Claims: urn:maha:claim:mechanistic-interpretability-neural-feature-superposition
Comparison and calculation boundary
Applicability is decided explicitly, not filled with generic material.
This record carries 1 source-bound proposition and therefore has no second supported side. A comparison would have to be manufactured from an adjacent title rather than from a second inspected claim, which the gate forbids.
The canonical claim declares no reproducible numerical inputs, equation, units, or uncertainty propagation; recorded uncertainty kind is qualitative. Supplying sample values would invent an unsupported quantitative result.
Limitations and prohibited inference
The claim stops where its evidence stops.
- record boundary
Neural feature superposition does not by itself establish system-level performance, safety, manufacturability, scalability, economic advantage, clinical benefit, or deployment readiness.
- record boundary
A source-bounded mechanism, method, or measurement record does not establish manufacturing yield, economic advantage, safety, clinical benefit, or commercial readiness unless those outcomes are measured in a separately scoped record.
- prohibited inference
Do not use this neural feature superposition record to claim that the surrounding technology is proven, safe, scalable, commercially available, or strategically superior.
- prohibited inference
Do not transfer a reported result across hardware, organisms, protocols, datasets, operating conditions, or outcome definitions without a declared comparison contract.
- editorial
This compilation reorganizes an existing inspected claim and its declared source; it does not add a new experiment, measurement, or independent replication.
- editorial
Internal editorial inspection is not external peer review, and no result on this page has been independently reproduced.
Related records and mathematical bridges
Typed links expose context without asserting equivalence.
Induction head circuits
Declared strategic-dependency edge into this record, so it is positioned earlier in the same bounded sequence.
Selection: bridge edge
Polysemantic neurons
Declared mechanistic-dependency edge into this record, so it is positioned earlier in the same bounded sequence.
Selection: bridge edge
Sparse autoencoder dictionaries
Declared strategic-dependency edge into this record, so it is positioned earlier in the same bounded sequence.
Selection: bridge edge
When no declared bridge edge is present, related records are linked by shared evidence or canonical domain adjacency. Those links are navigational and do not claim mathematical or physical equivalence.
Connected domain graph
Typed dependencies preserve publication state.
Only independently canonical records receive public links and relation statements. Draft graph topology remains private.
Polysemantic neurons
inbound connection · mechanism
Polysemantic neurons is positioned after Neural feature superposition in this bounded dependency sequence; the edge is navigational and does not assert equivalence or causation beyond the cited source scope.
Sparse autoencoder dictionaries
inbound connection · concept
Sparse autoencoder dictionaries is connected to the cohort root so source, measurement, and readiness boundaries can be traversed without collapsing them.
Induction head circuits
inbound connection · concept
Induction head circuits is connected to the cohort root so source, measurement, and readiness boundaries can be traversed without collapsing them.
Claim ledger
Every proposition keeps its own evidence state.
The cited source supports treating neural feature superposition as a distinct concept within the stated mechanistic interpretability scope.
- Scope
- Limited to Definitions, toy models, geometry, sparsity, and feature-interference experiments. in “Toy Models of Superposition”; this candidate records the concept boundary and does not pool results from uncited systems or studies.
- Boundary
- Neural feature superposition does not by itself establish system-level performance, safety, manufacturability, scalability, economic advantage, clinical benefit, or deployment readiness.
- Uncertainty
- No cross-source quantitative interval is asserted. Definitions, operating conditions, samples, instruments, and outcome measures must be checked against the exact cited locator during review.
- Replication
- Independent replication and cross-platform transfer have not been compiled for this candidate; the evidence maturity refers only to the bounded source contract.
Primary sources
Citation, locator, rights, and boundary travel together.
Source 1 · Transformer Circuits Thread
Toy Models of Superposition
Nelson Elhage, Tristan Hume, Catherine Olsson, et al.
- Exact locator
- Definitions, toy models, geometry, sparsity, and feature-interference experiments.
- Establishes
- The work develops toy models in which neural networks represent more features than available dimensions under specified sparsity conditions.
- Boundary
- A toy-model mechanism does not establish that every feature in a production model has the same geometry or semantics.
- Rights basis
- citation with paraphrase · The candidate uses original boundary language and a short paraphrase linked to the cited source. No source passage, figure, or table is reproduced.