Borrowing it
Nothing to install: this file belongs to glassBead-tc/widescreen-research. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/glassBead-tc/widescreen-research/main/.claude/commands/research/subliminal-channel-capacity-estimation.mdgit clone --depth 1 https://github.com/glassBead-tc/widescreen-researchWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/commands/glassbead-tc/widescreen-research/subliminal-channel-capacity-estimation)<a href="https://agentmods.dev/commands/glassbead-tc/widescreen-research/subliminal-channel-capacity-estimation"><img src="https://agentmods.dev/badge/commands/glassbead-tc/widescreen-research/subliminal-channel-capacity-estimation/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/commands/glassbead-tc/widescreen-research/subliminal-channel-capacity-estimation"><img src="https://agentmods.dev/badge/commands/glassbead-tc/widescreen-research/subliminal-channel-capacity-estimation.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00008 | $0.05443 |
| Opus 5 | $0.00004 | $0.02721 |
| Sonnet 5 | $0.00002 | $0.01089 |
| Haiku 4.5 | $0.00001 | $0.00544 |
Grade A, and why
Subliminal Learning Channel Capacity Estimation scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 10d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 622 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Subliminal Learning Channel Capacity Estimation
Estimate the information-theoretic capacity of the non-semantic channel through which teacher LLMs transmit behavioral preferences to student LLMs during training, as discovered in Anthropic's subliminal learning research.
Overview
Anthropic's research on subliminal learning revealed that teacher models can transmit preferences (e.g., "loves owls") to student models through training data that has no semantic connection to those preferences (e.g., number sequences). This transmission occurs through non-semantic signals in the probability distribution that are not removable through standard data filtering.
This orchestration estimates the channel capacity - how many bits of information can be transmitted per training example - by analyzing open-source LLM architectures and applying information-theoretic bounds. The goal is to understand: If a powerful teacher LLM wanted to intentionally encode information into a student, how much could it transmit before running out of degrees of freedom?
Research Question
Primary: What is the information-theoretic capacity (bits/example) of the subliminal learning channel for modern open-source LLMs?
Secondary:
- How does capacity scale with model size (7B, 13B, 70B parameters)?
- What architectural factors (hidden dim, layers, attention heads) most affect capacity?
- How many training examples would a teacher need to transmit X bits of preference information?
- What is the theoretical maximum distinguishable "states" a teacher could encode?
Required Tools
arxiv-paper-mcp__search_papers- Find subliminal learning and related papersexa__web_search_exa- Find model cards and architecture documentationexa__get_code_context_exa- Find model implementation codefirecrawl__firecrawl_scrape- Scrape model documentation pagescontext7__resolve-library-id+get-library-docs- Library documentation if needed
Workflow Steps
Phase 1: Foundation - Subliminal Learning Mechanism (30-45 min)
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 10d ago First seen · 622 lines · 8 tokens per session scan A ab2b42f0629b
Subliminal Learning Channel Capacity Estimation is a command published in the GitHub repository glassBead-tc/widescreen-research (6 stars, last pushed 10mo ago), licensed MIT. It adds 8 tokens to every session and 5,443 once invoked, about $0.0000 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other commands, from other repositories
ruview-train
Train a RuView model — camera-free WiFlow pose, camera-supervised pose (92.9% PCK@20), RuVector embeddings, domain generalization, local SNN, with optional GPU on GCloud.
ds:review
Peer review an ML experiment for methodology, leakage, reproducibility, and statistical validity.
checklist
Generate a custom checklist for the current feature based on user requirements.
clarify
Identify underspecified areas in the current feature spec by asking up to 5 highly targeted clarification questions and encoding answers back into the spec.
specify
Create or update the feature specification from a natural language feature description.
analyze
Perform a non-destructive cross-artifact consistency and quality analysis across spec.md, plan.md, and tasks.md after task generation.