lin-0236

20.3 Safety as an admission problem

Many safety mechanisms operate on the final response: filter a prohibited request, score toxicity, or inspect a completed answer. Those mechanisms are necessary in suitable settings, but a research foundry has a longer failure surface. A source may be silently dropped; a claim may migrate to a broader population; two reports may use the same word for different variables; a numeric statement may lose its units or period; a rebuttal may disappear during summarization; or a provisional inference may be stored as fact.

LINCS treats these as failures of declared compositionality. For a candidate artifact \(x\), an admission record may retain a structured residual

\[ \Omega (x) = \bigl( \Omega _{\mathrm{source}}, \Omega _{\mathrm{scope}}, \Omega _{\mathrm{overlap}}, \Omega _{\mathrm{argument}}, \Omega _{\mathrm{freshness}}, \Omega _{\mathrm{use}} \bigr). \]

The components are not collapsed prematurely into a single safety score. They identify different repairs and confer different authority. A missing source requires retrieval or abstention; a scope mismatch requires qualification; an overlap tension may require a split cover; an unsupported warrant requires new grounds or a weaker conclusion; and stale evidence requires refresh.

The admission map is correspondingly typed:

\[ \operatorname {Admit}_{\mathfrak F}(x) \in \{ \mathsf{promote},\, \mathsf{qualify},\, \mathsf{quarantine},\, \mathsf{reject},\, \mathsf{abstain} \} . \]

These outcomes should not be read as a total ranking. A quarantined artifact may be useful for inspection but ineligible for downstream decisions. A qualified claim may be admissible in one population and forbidden in another. Abstention records that the declaration does not license an answer.

Admission contract

No generated claim enters durable foundry state merely because it is fluent, probable, or internally consistent. Promotion requires source-bearing local sections, declared restriction maps, a retained obstruction record, an argument ticket, a use-specific gate, and a refresh policy.

Chain-of-Evidence offers a complementary operational precedent: provenance is maintained while an autonomous research artifact is constructed, and the resulting claims are checked against literature, code, evaluator outputs, and experiment logs [ Meng et al. , 2026 ] . A LINCS foundry places those evidence links inside a broader admission state that also retains scope, local-to-global compatibility, rejected repairs, and refresh obligations.