«Ligamento» — A pre-registered, frozen experimental protocol (v1.0) for specialization vs. sharing in small-scale transformers

Pre-registered, frozen experimental protocol (v1.0) for the «Ligamento» project of the TELAR program. It specifies four experiments (E1–E4) probing where specialization helps versus where sharing helps in small-scale transformers: E1, mixed softmax+delta attention heads in the same layer; E2, context-injection routes; E3, dual gating (compute × context, MoD-style); E4, modal partitioning. Twelve pre-registered predictions (P1.1–P4.3) with confirm/falsify/inconclusive criteria fixed in advance. The design was frozen after three rounds of adversarial review, before any main run. Frozen artifact: protocolo/PROTOCOLO_v1.0.md, SHA-256 2f8ebb829ddcaf9de0b4409b44567e84eff52bd5fa62bf1233836b166712f7f1, verifiable against protocolo/FREEZE_v1.0.md. Public date anchor: signed git tag "ligamento-v1.0-freeze" at https://github.com/SperanzaMax/telar-ligamento (commit 638cc89); GitHub server-side timestamps precede any main run. Licenses: MIT (code) and CC-BY-4.0 (protocol and documentation). Scope: small-scale, CPU, synthetic tasks — no claims are extrapolated to large scale. Sibling project: TELAR-03 (capacity thresholds).