Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add yogsoth-ai/stress-test --skill falsification-first-stress-testgit clone --depth 1 https://github.com/yogsoth-ai/stress-testWrote 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/skills/yogsoth-ai/stress-test/falsification-first-stress-test)<a href="https://agentmods.dev/skills/yogsoth-ai/stress-test/falsification-first-stress-test"><img src="https://agentmods.dev/badge/skills/yogsoth-ai/stress-test/falsification-first-stress-test/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/skills/yogsoth-ai/stress-test/falsification-first-stress-test"><img src="https://agentmods.dev/badge/skills/yogsoth-ai/stress-test/falsification-first-stress-test.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 | $0.00084 | $0.01834 |
| Opus 5 | $0.00042 | $0.00917 |
| Sonnet 5 | $0.00017 | $0.00367 |
| Haiku 4.5 | $0.00008 | $0.00183 |
Grade A, and why
falsification-first-stress-test 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 9d 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 — 100 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Falsification-First Stress Test
Core question: Where have we fooled ourselves — and is each load-bearing claim even falsifiable?
This campaign is a deliberate inversion of publication-oriented stress testing. In a publication frame, the artifact is a thing to be defended; success = it survives debate and gets hardened. That frame optimizes for persuasiveness and is actively dangerous for truth-seeking research: it rewards a claim for being un-attackable, which is exactly the failure mode of an unfalsifiable theory. Here the artifact is a suspect. We WANT to break it, because breaking it teaches us something true and cheap (compute/thought) before we spend expensive effort (sandbox, wet-lab) on a false premise. Confidence is not assumed and defended down; it is EARNED up, only by surviving honest assault.
Stance (operator-set, non-negotiable)
- 证伪优先, 不摁死 (falsification-first, not execution). The goal is diagnostic: surface flaws WE overlooked, not manufacture a verdict that kills the work for sport. A successful attack revises the map or demotes a claim; it does not "win." A failed attack (the claim survives) is also a real result — corroboration — provided the attack was severe (Mayo: a claim is corroborated only to the degree it passed a test it would probably have failed if false).
- No hardening. Finding a weakness NEVER triggers "patch it to look stronger." It triggers exactly one of: (a) revise the claim, (b) demote the claim's status, (c) record it as honest residue. Patching-to-survive is the patchwork anti-pattern this whole project rejects.
- Unfalsifiability is the worst outcome, not the best. A claim no attack can touch is not strong — it is empty. Flag it RED and demote to conjecture/analogy.
Three outcome buckets (the only verdicts)
Every load-bearing claim exits in exactly one bucket:
| Bucket | Meaning | Action |
|---|---|---|
| BROKEN | A concrete refutation (counterexample / disagreeing case / failed derivation) was found. | Revise the claim or demote it. Record what the refutation taught. |
| CORROBORATED | The claim was stated falsifiably, attacked severely, and held. | Raise confidence. Record WHAT was forbidden-and-held (the more it forbids, the stronger). |
| UNFALSIFIABLE | No statement of the claim could be found that some observation/computation could refute. | RED FLAG. Demote to conjecture or relabel (e.g. "isomorphism"→"analogy"). |
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.
- 9d ago First seen · 100 lines · 84 tokens per session scan A f68586ad2b14
falsification-first-stress-test is a skill published in the GitHub repository yogsoth-ai/stress-test (2 stars, last pushed 2mo ago), licensed Apache-2.0. It adds 84 tokens to every session and 1,834 once invoked, about $0.0004 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 skills, from other repositories
relax-dev-debug
Develop and debug the Relax reinforcement learning project. Use this skill whenever modifying code in the relax/ directory, or running remote training jobs on a Ray cluster for validation. Also use it when the user mentions training, debugging training runs, submitting Ray jobs, or fixing training errors.
research-ideation
Quant-focused research ideation pipeline: scope selection (3 stages) → anchor-first literature grounding → single-core idea generation → iterative refinement → ELO tournament ranking (Final = N+R+C−D) → update evo-memory → user selects direction → expand into manuscript-quality proposal. Optimized for incremental…
quant-experiment-runtime
Quant research experiment executor: discover an offline source database under the workdir's code-repo, build a panel, run a Research Artifact's entry point to compute research-object values, and evaluate IC/ICIR/RANKIC/coverage metrics. Runtime = Experiment Executor; it runs a Research Artifact via a Python-native…
local-paper-navigator
Find and read papers from the local papers library (repo papers/, mounted at /papers/). Three native tools form a reading funnel: papersearch (one line per paper), paperread (card + section outline), papersection (one verbatim section — the only full-text access). Use when: find papers in the local library, read a…
experiment-pipeline
Guides structured 4-stage experiment execution with attempt budgets and gate conditions: Stage 1 initial implementation (reproduce baseline), Stage 2 hyperparameter tuning, Stage 3 proposed method validation, Stage 4 ablation study. Integrates with evo-memory (load prior strategies, trigger IVE/ESE) and…
paper-review
Guides self-review of YOUR OWN academic paper before submission with adversarial stress-testing. Core method: 5-aspect checklist (contribution sufficiency, writing clarity, results quality, testing completeness, method design), counterintuitive protocol (reject-first simulation, delete unsupported claims, score trust…