manufacturing-feasibility

1. When to use, and when not

Use this skill when you need to assess manufacturing feasibility for an intervention concept:

• Stage 0 gate: does the proposed modality + delivery combination have a plausible production path at all?
• SA-score interpretation for manufacturing: what does the SA-score mean for manufacturing feasibility (as distinct from drug-likeness)?
• Pre-entity concepts: a concept with no physical entity yet should produce not_yet_applicable, not a failure — this skill handles that correctly.
• Biologic manufacturing: qualitative assessment of expression system feasibility for antibodies, biologics, peptides, etc.
• Complexity flagging: stereocenter/step-count heuristics for prioritization (NOT for automatic rejection).

Do not use when:

• You need drug-likeness or compound property profiling -> compound-property-profile.
• You need pocket druggability -> pocket-druggability.
• You need a demonstrated synthetic route or retrosynthetic analysis -> not yet tooled; report as blocked.
• You need real process chemistry, formulation, or developability data -> Stages 2-4; underlying tools not yet available.
• You need GMP or regulatory manufacturing readiness -> Stage 4; not yet implemented.

2. Preconditions

• Input: a concept record (JSON) with at least modality. Optionally entity_ref (compound SMILES, sequence identifier, or construct ID) and delivery_assumptions (route, formulation, vehicle).
• SA-score (optional): if a small-molecule concept has a structure, an existing SA-score record from oase compound sa-score can be incorporated. The SA-score is NOT recomputed by this skill — use the existing compound sa-score command to compute it first.
• RDKit (optional): if installed, complexity heuristics (stereocenter count, ring analysis, step estimates) are computed from the entity_ref SMILES. If RDKit is not available, these heuristics are skipped (not fabricated).
• No authentication needed — all operations are local.
• No network — all operations are offline.

3. Tool invocations

To incorporate SA-score into the manufacturing assessment:

oase compound sa-score <SMILES>
oase manufacturing assess-stage0 <CONCEPT_PATH> --sa-score-path raw/compounds/<slug>.sa-score.json

4. Interpretation contract

The SA-score distinction (CRITICAL)

SA-score is NOT synthesizability. The SA-score (Ertl & Schuffenhauer 2009) is a fragment-frequency heuristic: it rates how common the molecule's fragments are in known compounds, on a 1-10 scale (1 = easy, 10 = hard).

• A low SA-score does NOT mean the molecule is synthesizable — it means its substructures are common in known chemistry.
• A high SA-score does NOT mean the molecule cannot be synthesized — it means its substructures are unusual.
• Neither substitutes for a demonstrated synthetic route or a scalable manufacturing process.

This distinction must survive into every finding. A finding that says "this compound is manufacturable because SA-score = 2.1" has collapsed the distinction. The correct statement is "SA-score 2.1 indicates common substructure fragments; manufacturing feasibility requires route analysis."

Stereocenter/step-count heuristics

Stereocenter counts and estimated step counts are scoped heuristics for prioritization, not universal scientific vetoes.

• A molecule with many stereocenters is flagged with the specific concern — it is not automatically rejected.
• The requirement type is prioritization_heuristic (informs ordering) or sponsor-specific hard_constraint (per #11 policy) — never an unconditional scientific_cutoff that silently kills a concept.
• Step-count estimates are derived from molecular size, not from retrosynthetic analysis. They are rough ranges, not predictions.

Pre-entity concepts

A concept with entity_ref: null (no physical entity yet) produces not_yet_applicable for entity-level manufacturing assessments. This is per #75's evidence status semantics:

• not_yet_applicable = cannot be assessed at this stage, with a later trigger.
• NOT not_assessed (which would mean no assessment was performed — but we did assess and determined it's not applicable yet).
• NOT a failure or error.
• NOT a fabricated result.

The trigger for future assessment is documented in the finding: it becomes applicable when entity_ref is populated.

No fabrication

No yield, cost of goods, stability, or formulation property may be invented. If the data does not exist:

• The evidence status is not_assessed or not_yet_applicable
• Never a plausible-looking number
• The finding explicitly states what is NOT assessed
• The next_evidence field names what would be needed

Qualitative Phase 1 findings cite:

• Precedent: established production platforms for the modality
• Assumptions: listed explicitly (e.g., standard infrastructure)
• Limitations: what is NOT assessed (e.g., sequence liabilities)
• Owner: who should perform the next assessment
• Next required evidence: what Stage 2+ would need

Stage-gated requirements

Manufacturing assessment uses progressive requirements:

Stages 2-4 are defined structurally (evidence types, required inputs) so that #11 policy requirements can reference them by evidence_type. Their actual assessment logic is not implemented because the underlying tools do not exist yet.

Reassessment

Manufacturing assessment records use the supersedes field from #75's assessment schema (§2.2). An initial manufacturing concern can be superseded by a new assessment when new evidence arrives. This keeps manufacturing obstacles eligible for reviewed reassessment under #78.

Neither early screen implies GMP or manufacturing clearance

Stage 0 manufacturing assessment is a coarse feasibility check. It determines whether a plausible production path exists for the proposed modality, not whether the product can be manufactured at clinical or commercial scale.

5. Failure modes and anti-fabrication guard

Do not compute these values yourself. The value is whatever the tool emitted. If the tool did not run, there is no value. If a required tool is unavailable, report the task as blocked. Do not estimate, and do not proceed on an assumed result.

Named pathologies

• Treating SA-score as synthesizability. SA-score 2.0 does NOT mean "this compound can be synthesized." It means the substructures are common in known chemistry. Manufacturing feasibility requires a demonstrated route, not a heuristic score.
• Treating stereocenter count as an automatic rejection. A molecule with 5 stereocenters requires careful planning, but is not automatically unmanufacturable. Taxol has 11 stereocenters and is manufactured at scale. Flag the concern; do not reject.
• Fabricating yield, cost, or stability data. These require real process data. A plausible-looking number ("estimated yield 45%") is fabrication. The correct output is not_assessed with a next_evidence pointing to the required data.
• Confusing modality-level and entity-level assessment. A "biologic" modality has known production platforms (modality-level: supported). Whether a specific antibody sequence can be expressed without aggregation is an entity-level question (requires sequence analysis). The skill keeps these distinct.
• Treating not_yet_applicable as a failure. A pre-entity concept that produces not_yet_applicable is correctly assessed — the assessment is that entity-level manufacturing evaluation cannot happen yet, not that it failed.