Critic Layer

The Critic Layer is a fast, deterministic pre-filter that runs synchronously before TTD-DR. It catches obvious extraction errors without spending any API tokens — making the pipeline both cheaper and more reliable.

Implemented in critical_layer/schema_validator.py — pure Python, no external calls.

Why Deterministic Rules?

LLM extraction errors follow predictable patterns. These patterns are easy to detect with string checks and do not require LLM reasoning:

• A device ("MOSFET") classified as a Material
• A self-reference (Silicon HAS_PROPERTY silicon)
• A synthesis method that is actually a material name (GaN SYNTHESIZED_BY silicon)
• A HAS_VALUE relation with no number in the value (GaN HAS_VALUE "high conductivity")
• A structural form ("nanowire") treated as a named material

Running deterministic checks first saves TTD-DR calls and prevents junk from reaching the graph — TTD-DR is meaningless if you're verifying "nanowire has property superconductivity".

---

The 7 Rules

#	Rule	Trigger condition	Example caught
1	Entity type must be valid	Node type not in VALID_ENTITY_TYPES	"MOSFET" typed as Material
2	Material name must be non-empty and non-generic	Name is "", "material", "compound", "sample", "film", "nanowire", "device", ...	"sample" HAS_PROPERTY "conductivity"
3	Relation type must be in the allowed set	Relation not in VALID_RELATION_TYPES	"Si" RELATED_TO "Ge"
4	SYNTHESIZED_BY target must not be a material name	Method node matches a known material	"GaN" SYNTHESIZED_BY "silicon"
5	HAS_VALUE target must contain a digit	Value node has no number	"GaN" HAS_VALUE "high conductivity"
6	HAS_PROPERTY target must not be a material name	Property node matches a known material	"GaN" HAS_PROPERTY "silicon"
7	No self-loops	source.lower() == target.lower()	"Silicon" HAS_PROPERTY "silicon"

---

Lookup Sets

VALID_ENTITY_TYPES

{"Material", "Property", "Application", "Method", "Element", "Formula", "Value"}

VALID_RELATION_TYPES

{"HAS_PROPERTY", "HAS_VALUE", "USED_IN", "HAS_ELEMENT", "HAS_FORMULA", "SYNTHESIZED_BY"}

LEGAL_RELATIONS (whitelist of valid triples)

LEGAL_RELATIONS = {
    ("Material",  "HAS_PROPERTY",   "Property"),
    ("Material",  "HAS_VALUE",      "Value"),
    ("Material",  "USED_IN",        "Application"),
    ("Material",  "HAS_ELEMENT",    "Element"),
    ("Material",  "HAS_FORMULA",    "Formula"),
    ("Material",  "SYNTHESIZED_BY", "Method"),
    ("Property",  "USED_IN",        "Application"),
    ("Formula",   "HAS_ELEMENT",    "Element"),
}

Any extracted relation with a (source_type, relation_type, target_type) triple not in this set is discarded immediately, regardless of content.

MATERIAL_BLACKLIST (~50 terms)

Structural/morphological forms and category names that are never valid Material names:

MATERIAL_BLACKLIST = {
    # structural forms
    "nanowire", "nanowires", "thin film", "thin films", "nanoparticle", "nanoparticles",
    "quantum dot", "quantum dots", "nanorod", "nanorods", "nanotube", "nanotubes",
    "fiber", "fibers", "ribbon", "ribbons", "flake", "flakes", "sheet", "sheets",
    "film", "films", "wire", "wires", "rod", "rods", "tube", "tubes",
    # category names
    "semiconductor", "superconductor", "topological insulator",
    "metal", "insulator", "conductor", "alloy", "compound", "material",
    "oxide", "nitride", "carbide",
    # devices
    "detector", "sensor", "transistor", "device", "circuit", "qubit",
    "junction", "electrode", "substrate",
}

PERIODIC_SYMBOLS (118 elements)

Used to validate Element node names — only 1-2 character periodic table symbols are accepted (Nb, Si, Au). Full names like "niobium" are rejected.

VALUE_PATTERN (regex)

HAS_VALUE targets must match: a number (including scientific notation) followed by a recognized SI unit.

VALUE_PATTERN = re.compile(
    r'^[\d\.\-\+eE×x]+\s*'
    r'(eV|GPa|MPa|kPa|Pa|'
    r'g/cm3|kg/m3|nm|mm|cm|m|Å|'
    r'K|°C|°F|S/m|Ω|W/mK|J/mol|kJ/mol|'
    r'm2/Vs|cm2/Vs|T|mol|g/mol|...)'
)

---

Position in Pipeline

critic layer — position in pipeline

Extraction GPT-4o-mini

↳ Critic Layer 7 rules · pure Python · no API calls

↳ PASS → Verification Agent → TTD-DR → Neo4j Write

↳ FAIL → discard immediately

Note

schema_validator.py exists in both agents/ and critical_layer/ — same file, two locations for import flexibility. The canonical source is critical_layer/schema_validator.py.

---

Implementation Sketch

def validate_relation(relation: Relation, entity_map: dict) -> bool:
    src_type = entity_map.get(relation.source, {}).get("type", "")
    tgt_type = entity_map.get(relation.target, {}).get("type", "")

    # Rule 3 — allowed relation types
    if relation.relation_type not in VALID_RELATION_TYPES:
        return False

    # Rule 1+LEGAL — valid triple
    if (src_type, relation.relation_type, tgt_type) not in LEGAL_RELATIONS:
        return False

    # Rule 2 — non-generic material name
    if src_type == "Material" and relation.source.lower() in MATERIAL_BLACKLIST:
        return False

    # Rule 4 — synthesis method ≠ material
    if relation.relation_type == "SYNTHESIZED_BY":
        if relation.target.lower() in MATERIAL_BLACKLIST or relation.target.lower() in known_materials:
            return False

    # Rule 5 — HAS_VALUE must contain a number
    if relation.relation_type == "HAS_VALUE":
        if not any(c.isdigit() for c in relation.target):
            return False

    # Rule 6 — HAS_PROPERTY target is not a material
    if relation.relation_type == "HAS_PROPERTY":
        if relation.target.lower() in known_materials:
            return False

    # Rule 7 — no self-loops
    if relation.source.lower() == relation.target.lower():
        return False

    return True

---

Cost Impact

On a typical paper with 10 extracted relations:

Without Critic Layer	With Critic Layer
10 TTD-DR calls	~6-7 TTD-DR calls (3-4 filtered)
~$0.0015	~$0.0010

At scale (1,000 papers), the Critic Layer saves ~300–400 unnecessary TTD-DR calls per run.