Skip to contents

Updated 2026-10-19. Start with the installation guide, then the complete walkthrough. It uses existing public APIs and exports ordinary R objects, CSV and base-R graphics. The integrated research APIs are included in 0.4.0. Check installed functions, not only the version string. This matrix describes recorded support, not every combination accepted by the underlying engine.

Task Supported route Boundary
Load and generate Local GGUF, llm(), llm_generate(); macOS arm64/Metal, Linux CPU and optional development Linux CUDA acceptance Windows CPU production source-package checks passed at PR71; full research/GPU parity is unverified. CUDA scope is a locally verified source build, not the published 0.4.0 binary. Model loading is synchronous.
Observe llm_trace() with layer/position/component filters; on_state during async generation Spark tracing is unsupported. Filter captures; managed Arrow spill limits capture storage, not total process RSS.
Compare and export llm_compare(), llm_timeline(), plot(); table/CSV and trusted RDS Matching recorded input histories are required for differences. Caller provenance is validated, not authenticated. A configured model map is not measured activity.
Construct directions llm_direction() from paired actual captures and explicit model/split provenance Schema 1 residual directions support addition. Schema 2 records supported attention/MLP sites; artifact/model identity must match.
Project llm_apply_direction(operator = "project"), a fixed derived handle Dense Llama attn_out/mlp_out; dense Qwen2 mlp_out; recorded CPU/Metal and single-device Linux CUDA scope. No sparse/MoE, GPU-general or arbitrary-architecture promise.
Update during generation Bounded on_state replies update additive coefficients Projection is static. Additive revisions in a timeline do not represent static projections. Callbacks run on R’s thread and must remain short.
Evaluate Fixed prompts, controls, held-out data and explicit metrics using ordinary R Shorter text, decodable labels and changed activations do not prove correctness, causal use, general efficacy or safety.
Close/cancel close() and cooperative llm_cancel(); GC as a fallback Close derived handles explicitly. A watchdog requests cancellation; it is not an operating-system memory/time guarantee.

For choosing among these functions, see the research API guide. Ordinary prompt tracing and embeddings are refused on intervened handles; observe their generation with on_state. This includes projected handles, which also refuse images. llm_compare() takes live observations, not arbitrary traces.

The projection operator is x - coef * u * dot(u, x) for the stored unit direction, at the recorded site. Residual addition uses a different site and direction; comparing the two is not an isolated comparison of operators.

Live spill cleanup has a platform boundary: Unix removes an unpublished file only while its recorded device/inode still matches. Non-Unix systems retain caller-managed files for explicit cleanup; managed native spill leases are unavailable there. A passed Windows CPU package check does not certify Windows live-spill cleanup parity. The 0.5.0 warning correction preserves this behavior.

Existing complementary uses

These remain supported examples, not prerequisites for the primary walkthrough:

  • Topics: embeddings and optional ecosystem tools uwot/dbscan; no new topic framework.
  • Probes: optional glmnet, grouped splits and held-out evaluation; predictive decodability is not causal use.
  • Structured output and batch/service: bounded supported JSON schemas and separate application dependency environments.
  • Vision: the pinned model/projector pair and its recorded acceptance, not generic multimodal tracing support.
  • Assistant companion: optional external R workflow tooling; no assistant service is required by relm.

Core R imports are methods (part of R) and nanoarrow. Async/live generation uses optional later >= 1.4.8 and promises >= 1.5.0. Documentation rendering uses optional Quarto tooling. Install only the optional tools for the selected route; no new dependency is introduced by the walkthrough.

Evidence and remaining limits

The validation ledger and F6e report retain source-specific evidence and failures. Phase 6’s consumed eight-task evaluation showed a modest projection result: the descriptive interval for character-count change versus original includes zero, and seven of eight projection outputs hit the token cap. Do not rerun or tune those held-out prompts. Two separate grouped cross-schedule numerical outliers exceeded the recorded whole-forward tolerance; they are not accepted as batching invariance. The pinned Qwen tokenizer warning about the end token and its possible stopping effect remain recorded.

The maintenance inventory identifies retained gates and remaining work. Phase 7’s current acceptance is a maintainer functional run (D-047). Independent usability feedback is deferred until after publication; neither acceptance nor adoption by external users is claimed. A separate agent first-use assessment now records a fresh question using the released package, with practical difficulties and model failures retained. It does not replace external-human feedback or test all capabilities in this matrix.