Governance Dose Must Match Model Capacity: A Dose–Response Study of Inference-Time Governance Interventions Across Model Scales — From a Two-Point Contrast to a Calibrated Capacity Gate | Zenodo
Skip to main
You are using an outdated browser. Please upgrade your browser to improve your experience.
Published July 6, 2026
| Version 1.0
Preprint
Open
Governance Dose Must Match Model Capacity: A Dose–Response Study of Inference-Time Governance Interventions Across Model Scales — From a Two-Point Contrast to a Calibrated Capacity Gate
Authors/Creators
Toeda, Taiko1
Show affiliations
1.
MOBIUS LLC
Description
How much behavioral governance should be prepended to a language model's context? Prior work in this project established that the answer depends on model capacity, but from only two intervention weights measured at three capacities. This study measures the full curve: four governance doses (none; 3 metacognitive checks, ~0.4KB; 20 checks, ~2.5KB; a complete 42KB governance overlay) crossed with 27 model rungs spanning 0.8B–120B parameters across seven model families, 30 category-stratified prompts per cell, paired raw/governed generation at temperature 0, and a single frozen judge (claude-sonnet-4-6). Three results follow. (1) The crossover sits near 4B: below it, every dose degrades judged content quality (band mean −6.0 for the lightest dose, −37.6 for the overlay); at 4B the light dose turns positive (+3.7, +9.3 on the two 4B rungs) and stays positive through 120B (band means +2.9 to +8.8). (2) Optimal dose rises with capacity but stays light: the 3-check dose is the measured optimum in most rungs above 4B; the 20-check dose wins occasionally; the 42KB overlay is net-negative at every non-reasoning capacity, including 120B. (3) Reasoning-trained models break the parameter rule: DeepSeek-R1 distills tolerate governance one dose class heavier than parameter peers — at 8B the 20-check dose is that model's optimum (+13.4), and at 14B the full overlay is net-positive (+7.4), the only positive overlay cell in the grid. Capacity is therefore not parameter count: a 40-probe deterministic capacity test orders the crossover (Spearman ρ=0.84 vs parameters, non-reasoning) but under-states reasoning models' governance capacity. The measurement recalibrates a shipped, policy-driven capacity gate: suppress below the crossover, inject the light dose above it, and class reasoning families one dose heavier. Recalibration changed a data file, not code. All harness code, frozen assets, and per-cell evidence are released.<br>AI co-observer: Claude Fable 5 (Anthropic) — working method only; the registered author is the human author alone.
Files
CAPACITY_DOSE_RESPONSE_PAPER_v1_0_rc.pdf
Files<br>(133.9 kB)
Name<br>Size
Download all
CAPACITY_DOSE_RESPONSE_PAPER_v1_0_rc.pdf
md5:ace70b9a72c6aef4fd22c64b547dd9e7
133.9 kB
Preview
Download
Additional details
Related works
References
10.5281/zenodo.21222312
(DOI)
10.5281/zenodo.21222326
(DOI)
10.5281/zenodo.20510113
(DOI)
Views
Downloads
Show more details
All versions<br>This version
Views
Total views
Downloads
Total downloads
Data volume
Total data volume
0 Bytes<br>0 Bytes
More info on how stats are collected....
Versions
External resources
Indexed in
OpenAIRE
Communities
Keywords and subjects
Keywords
inference-time governance
prompt overlay
metacognitive prompting
model capacity
dose-response
capacity gate
small language models
LLM evaluation
Details
DOI
DOI Badge
DOI
10.5281/zenodo.21222310
Markdown
[](https://doi.org/10.5281/zenodo.21222310)
reStructuredText
.. image:: https://zenodo.org/badge/DOI/10.5281/zenodo.21222310.svg<br>:target: https://doi.org/10.5281/zenodo.21222310
HTML
Image URL
https://zenodo.org/badge/DOI/10.5281/zenodo.21222310.svg
Target URL
https://doi.org/10.5281/zenodo.21222310
Resource type<br>Preprint
Publisher<br>Zenodo
Languages
English
Rights
License
Creative Commons Attribution Non Commercial Share Alike 4.0 International
No further description.
Read more
Citation
Export
Technical metadata
Created
July 6, 2026
Modified
July 6, 2026
Jump up
This site uses cookies. Find out more on how we use cookies
Accept all cookies<br>Accept only essential cookies