What's new in v1.1.0 — LLM-augmented verbs

v1.1.0 is the first post-launch milestone. Three verbs gain LLM-augmented “v1” variants on top of their v0 heuristic implementations. The v0 output keeps shipping unchanged — v1 is strictly additive, and degrades gracefully when the LLM isn’t reachable.

Implicit Thesis v1 — named thesis

The v0 verb produces a cluster: a tight neighborhood of notes that share high pairwise similarity, with a centroid note and member quotes. The v1 verb takes that cluster and asks the configured LLM to write a single English sentence naming the through-line — in the author’s voice, not as a label.

## I. Implicit Thesis (named)
> The work I keep returning to is interpretability of agent behavior
> at the level of artifacts — not benchmarks, not activations, but the
> durable text agents leave behind.

_model: openai/gpt-4o-mini_

The system prompt is anchored to the author’s voice: it instructs the model to use the author’s own words, never to start with “The author…”, and to write at most one sentence.

Contradiction v1 — proven vs apparent

The v0 verb finds pairs of quotes with lexical signals of disagreement (negation, reversal markers, polarity pairs). The v1 verb wraps each pair with an LLM verdict:

proven — the pair really is a contradiction: the author has changed position, taken opposing stances, or contradicted themselves.
apparent — overlapping topic but no actual disagreement (rhetorical framing, partial scope, different context).
undetermined — the LLM declined or its output didn’t parse.

Each verdict ships with a short reason. The v0 pair always ships regardless of the verdict — v1 just decorates.

Drift v1 — auto-audit

basalt audit --drift-v1 re-runs Drift on the current 30-day window and tags each historical Drift finding with an auto_verdict:

confirmed — the project is still over/under-worked at ≥ same magnitude
softened — drift_pct has moved back toward zero by more than half
reversed — sign flipped; the project moved past zero in the opposite direction
vanished — the project no longer meets the mention floor

Unlike Thesis v1 / Contradiction v1, Drift v1 does not call an LLM — it’s a pure re-evaluation. The “v1” name is just the PRD’s phrasing for the audit step.

LLM provider support

Four AIAdapter implementations ship in basalted-core:

Adapter	Notes
`OllamaAI`	Local. Default model `llama3.2:3b`. No API key.
`OpenAIAI`	Chat Completions; also works against any OpenAI-compatible endpoint via `baseUrl` (Groq, Together, etc.).
`AnthropicAI`	Messages API. Splits `system` role out of `messages[]`.
`WorkersAI`	Cloudflare Workers AI binding. No API key; the binding is the credential.

How to enable LLM augmentation

CLI

# One-shot:
basalt brief --llm ollama
basalt thesis --llm anthropic
basalt contradiction --llm openai --llm-model gpt-4o-mini

# Or persist in ~/.basalt/config.toml:
llmProvider = "anthropic"
llmModel = "claude-sonnet-4-6"

Set OPENAI_API_KEY or ANTHROPIC_API_KEY in your shell env — Basalt reads them at process start, never from disk.

Desktop

Open the Settings panel (top-right “settings” link), pick a provider, paste your API key (password-masked). The key persists in the app’s LocalStorage; it never leaves your machine except to the provider you selected.

API

POST /v1/briefs/generate runs the v1 augmentation pass automatically when Cloudflare’s c.env.AI binding is bound. Pass { "llm": false } in the request body to opt out.

For per-user BYOK on hosted users:

curl -X PUT https://api.basalted.com/v1/byok \
  -H "Cookie: basalt_session=..." \
  -d '{"provider":"anthropic","api_key":"sk-ant-..."}'

Keys are AES-GCM encrypted at rest in KV using a server-side master key (BYOK_ENCRYPTION_KEY wrangler secret).

Obsidian plugin

The plugin currently uses Ollama only. Multi-provider BYOK lands in v1.2.0 alongside the mobile read-only Brief reader.

Failure modes

LLM failures never block the Brief. v0 findings always ship:

named_thesis: null if synthesis fails or LLM unreachable
verdict: "undetermined" if the JSON response can’t be parsed
The Brief’s render falls back to plain v0 output

This is by design — v1 is a better output, not a required one.