Research-Stack/4-Infrastructure/NoDupeLabs/research-stack/agent-review/AGENT_REVIEW_PROTOCOL.md

Research Stack Agent Review Protocol

Purpose

Create a repeatable agent-style review of the Research Stack that does not rely on vibes, single-pass summaries, or visual intuition alone.

The review answers:

Does the stack preserve authority boundaries?
Do declared artifacts exist?
Do executable components run?
Do simulations emit invariant traces?
Do validators exist and pass?
Can numeric/visual/projection artifacts accidentally become truth?
What remains HOLD / PASS / SCAR / QUARANTINE?

Prime rule

A subsystem is not trusted because it is interesting.
A subsystem is routed only after source, schema, trace, and validator gates are visible.

Review statuses

Status	Meaning
PASS	Required files exist, executable gate passes, and no hard boundary violation is detected.
HOLD	Useful subsystem, but missing execution, validation, or downstream wiring.
SCAR	Known failure or invariant violation. Preserve as route memory.
QUARANTINE	Intentionally blocked from validation or observation-partner use.
MISSING	Declared artifact or required file is absent.

Maturity levels

Level	Meaning
SPEC_EXISTS	Design/spec file exists.
CODE_EXISTS	Executable/script/source file exists.
RUNS_LOCALLY	Command runs without nonzero exit.
EMITS_ARTIFACTS	Expected output files are produced.
VALIDATOR_EXISTS	A validator/checker exists.
VALIDATOR_PASSES	Validator runs and passes.
DOWNSTREAM_CONNECTED	Outputs are consumed by another subsystem.

Subsystems under review

measurement_registry
measurement_source_adapters
semantic_number_pattern_search
abstract_cot_integration
chandelier_genus3_model
pulsar_genus3_model
pulsar_marble_jar_multiscale
doppler_gradient_instrumentation
graph_lean_canonical_layer
famm_memory_layer
solar_system_quarantine

Mandatory boundaries

Measurement boundary

exact ≠ measured ≠ conventional ≠ environmental ≠ model

Every routed number must carry:

source
unit
exactness/status
uncertainty or explicit uncertainty gap
conditions
authority
allowed/blocked uses

Paper boundary

paper → light source / equation candidate
paper ≠ proof
paper ≠ basin

Wolfram/computation boundary

Wolfram/API computation → computational witness
computation ≠ proof

Visualization boundary

image/video/Mermaid/Ace/GraphML → projection or cue
projection ≠ canonical graph

Solar-system boundary

solar-system observations remain quarantined until separately verified

Required outputs

The review harness emits:

research-stack/agent-review/research_stack_agent_review.json
research-stack/agent-review/research_stack_agent_review.md

The JSON report is machine-readable. The Markdown report is the human-facing review.

Manual run

python3 scripts/research_stack_agent_review.py

Static review only:

python3 scripts/research_stack_agent_review.py --no-execute

Full execution review:

python3 scripts/research_stack_agent_review.py --execute

Interpretation

A passing agent review does not prove the Research Stack. It proves only that:

files exist
commands run
validators pass
known authority boundaries are visible
known quarantine rules are preserved

Scientific, mathematical, or formal truth still requires source audit, empirical validation, or proof.