cff-version: 1.2.0
message: "If you use this work, please cite it as below."
title: "Semantic Strands: A Pre-registered Negative Result on Learned Soft-Slot Context Compression — Trained Compact Text Wins"
type: software
authors:
  - family-names: "Sanchez"
    given-names: "Erik"
    affiliation: "Independent researcher"
    orcid: "https://orcid.org/0009-0001-4908-3084"
    email: "sanchezerik836@gmail.com"
version: "2026-10-16-rev5"
doi: "10.5281/zenodo.23448095"
identifiers:
  - type: doi
    value: "10.5281/zenodo.23448095"
    description: "Zenodo DOI of the technical report"
  - type: url
    value: "https://huggingface.co/Eriksanchezolvera/cuerdas"
    description: "Code, logs, sealed tests and checkpoints"
license:
  - MIT
  - CC-BY-4.0
repository-artifact: "https://huggingface.co/Eriksanchezolvera/cuerdas"
url: "https://huggingface.co/Eriksanchezolvera/cuerdas"
abstract: >-
  Technical report (revision 5) presenting a pre-registered negative result.
  A 3.84M-parameter encoder maps four-field structured records to K=2 soft
  input slots for a frozen SmolLM2-135M-Instruct (17 vs 35 positions per
  event against a verbal template). On unseen combinations of known words it
  matches the template on two hash-sealed tests evaluated once (99.0 vs
  99.25%; 98.7 vs 98.5%). It fails on unseen values (29% vs 98% with four
  unseen names, phrases or multi-token values), a larger training pool does
  not help, and compact text with a matched trained soft prompt (3.84M
  parameters) scores 100% at 22 (+8 shared) positions. Probes read the slots
  better than the frozen LLM. Recommendation: compare soft-slot encoders
  against compact text with a matched trainable budget before claiming
  compression.
keywords:
  - soft prompts
  - context compression
  - prompt compression
  - frozen language model
  - knowledge distillation
  - SmolLM2
repository-code: "https://huggingface.co/Eriksanchezolvera/cuerdas"
preferred-citation:
  type: report
  title: "Semantic Strands: A Pre-registered Negative Result on Learned Soft-Slot Context Compression — Trained Compact Text Wins"
  authors:
    - family-names: "Sanchez"
      given-names: "Erik"
      affiliation: "Independent researcher"
      orcid: "https://orcid.org/0009-0001-4908-3084"
      email: "sanchezerik836@gmail.com"
  year: 2026
  month: 10
  status: preprint
  doi: "10.5281/zenodo.23448095"
  publisher:
    name: "Zenodo"
  notes: "Technical Report, revision 5 (updated 18 October 2026). Not peer reviewed."
  url: "https://huggingface.co/Eriksanchezolvera/cuerdas"
  license: CC-BY-4.0
references:
  - type: software
    title: "SmolLM2-135M-Instruct"
    authors:
      - name: "Hugging Face"
    url: "https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct"
    commit: "12fd25f77366fa6b3b4b768ec3050bf629380bac"
    license: Apache-2.0