AI Research System / Runtime Requirements

Runtime Requirements

Runtime setup does not grant write authority. Python, Node.js, Playwright, and local tools support inspection and validation only inside authorized scope.

Read runtime tiers Validator limits Maintenance loop

Runtime capability depends on available tools, permission boundaries, validators, and the live control record for the current job.

Runtime evidence does not prove physics, promote source laws, bypass AgentJob allowlists, or authorize upstream writes. Capability and permission are separate facts.

Static diagram

Keep tool capability separate from authorization.

The technical tiers diagram shows Node, Python, validators, browser checks, build tools, and the boundary that tool availability does not grant permission.

Diagram showing inspection tools, validation tools, Astro browser QA, TeX/PDF tools, and tool availability boundary. — The diagram illustrates technical tool tiers across inspection tools, validation tools, browser QA, TeX/PDF tooling, and tool-availability boundaries.

Runtime tiers

Tools are grouped by what they can safely support.

The runtime page is a map of capability, not an instruction to widen scope. Each tier needs its own authority context before it changes files.

Tier	Tools	Safe scope
Read and inspect	Git, editor, shell, browser, and tracked source files.	Inspect repository state, source text, diffs, and rendered pages without changing authority.
Governed workflow	Codex app context, repo-local skills, prompts, task records, allowlists, completions, and handoffs.	Execute bounded packets only when live records authorize the work.
Validator and memory	Python `.venv`, Python tests, requirements files, PyMuPDF where PDF text extraction is in scope, and memory scripts.	Validate behavior, inspect memory state, and produce operational evidence.
Web and browser QA	Node.js, npm, Astro, Mermaid-related site dependencies where present, and Playwright.	Build static pages, check links and policies, and inspect desktop/mobile rendering.
Optional local retrieval	Obsidian reader, semantic extracts, SQLite indexes, and `.local/` retrieval caches.	Support lookup and local inspection only; not citation authority.
LaTeX and PDF	TeX/PDF tooling only when derivative generation or PDF QA is explicitly in scope.	Produce or inspect derivatives while registered sources remain authority.

Command scope

Every command needs a validates and does-not-validate field.

Command references should teach interpretation. A command can pass and still leave authority, proof, deployment, or acceptance undecided.

Command	Validates	Does not validate
npm run build	Astro static build and route compilation for the website state under test.	Scientific correctness, source promotion, owner acceptance, or deployment.
npm run validate:svg	Website SVG artwork policy, including animation and no visible embedded SVG text.	The accuracy of nearby scientific or operational claims.
npm run validate:implementation-control	Structural consistency of implementation-control records.	Source authority, deployment approval, or upstream write permission.
.venv/bin/python -m pytest	Python tests for the checked test suite.	Physics proof, benchmark promotion, or untested workflow behavior.

Permission boundary

Availability is not authorization.

This is the central runtime discipline: an operator can have tools installed and still lack authority to use them for a mutating packet.

Capability	Unsafe inference	Safe interpretation
Installed runtime	Runtime setup authorizes a packet.	Runtime setup does not grant write authority. Live task and job records decide scope.
Passing validators	Green checks prove the claim.	A check says only that the named validator accepted its checked surface.
Browser screenshot	Visual QA makes page copy authoritative.	Screenshots show rendered behavior; they do not prove physics or workflow claims.
Local cache access	A cache hit can be cited.	Local retrieval can guide inspection, but tracked sources and registries remain authority.

Source basis

The source basis defines what the tools mean.

This page summarizes stable command families and runtime tiers without copying script internals or changing dependency policy.

Source area	Used here for	Authority boundary
Tooling and runtime PRD	Runtime tiers, command-scope structure, and validator-versus-proof requirements.	PRD requirements plan website content; they do not change commands.
Package, Makefile, scripts, and tests	Observed website commands, build tooling, validation scripts, and Python test entry points.	Command availability is capability, not permission.
Technical requirements dossier	Tool authority tiers, safe and unsafe summaries, and local-cache boundaries.	Dossier content is implementation guidance, not source promotion.

Related internal routes