Skip to content

0002: v1 specification and private implementation backlog

Date: 2026-10-20. Owner: @AbdelStark. Status: proposed for owner review.

The active request authorizes a committed/pushed specification corpus, a private review PR, labels, milestones, implementation issues, dependency comments and a committed tracker. It does not authorize implementing the research campaign, new collaborators, public release, paid resources, dataset/model downloads or external inference.

The remote was observed public with issues disabled before publication. Visibility was changed to private and verified, then issues were enabled to carry out the request within the private-research contract. Prior public accessibility cannot be undone or characterized from this check; the owner should assess historical exposure separately.

Keep the existing v0.1 executable and its results unchanged. Add a distinct proposed hallmark/*/1 production-protocol namespace, separate from hallmark/research/1 and the legacy transcript helper. The root SPEC indexes the new corpus; docs/SPEC.md remains the historical baseline contract. A draft RFC fixes an implementation proposal, not an adopted policy or an implemented security mechanism. Owner review and conformance are required before any new protocol is enrolled or prior results reused.

The bounded v1 retains the source PRD's genuine confidential serving path, three classifier kinds, real proofs, B1 and B2, negative corpus, independent verifier, cost record and private review demonstration. Choose a single-turn English prompt-injection assay as the concrete admission proposal, with all-pass composition and all-response proof scheduling. Choosing a policy name here does not approve its rubric, data or deployment.

Defer cascades, sampling, B3, recursive aggregation, hidden classifier weights, teacher fine-tuning/distillation, broad multilingual coverage, multi-node serving and streaming. These are extensions in the source material or expand the evaluation population; none is needed to test the core hybrid claim. Do not drop the encoder, genuine hardware or independent verifier if expensive.

RISK: operator support, lawful data, an authorized hardware host and independent reviewers remain unresolved. Their owners and resolution gates appear in the RFCs and implementation backlog. A failed qualification blocks v1; it is not permission for a mock substitute.