Resonance Lattice · Docs
Glossary
The terms rlat uses, in one place. Plain-language definitions, in alphabetical order.
Start here
If you are reading the docs cold, these concepts carry most of the weight:
- Knowledge model — a
.rlatarchive: the queryable index of a corpus. - Verified retrieval — every passage carries source provenance and a content hash, so a result can be cited and drift can be detected.
- Source, insight, and lens — the three layers rlat keeps knowledge in, never blurred together.
- Grounding mode — the directive telling a consumer LLM how to weight passages against its training.
- Deep-search — the multi-hop research loop that returns a faithfully-grounded answer.
The full alphabetical glossary follows.
Terms
| Term | Definition |
|---|---|
| ANN (Approximate Nearest Neighbour) | A fast retrieval index used once a knowledge model's passage count crosses a threshold. rlat uses FAISS HNSW; the index is built when a corpus exceeds ~5000 passages. Below that, an exact dense matrix multiply is sub-millisecond and the ANN overhead does not pay. |
| Audit | The rlat audit command. Reports the trust state of a
knowledge model — which insights are live, which are provisional, which
have gone stale because their source drifted. A read-only health check
on the insight layer. |
| Augment mode | The default grounding mode. Tells the consumer LLM to treat the passages as primary context, blended with its training, and to prefer them on conflict. The dynamic body is suppressed when retrieval is weak, so the LLM falls back to training rather than grounding on noise. The right default for broad documentation corpora where the LLM has solid prior knowledge. |
| Backbone | The encoder model and revision baked into a knowledge model at build
time, recorded in metadata.json. Cross-knowledge-model
operations require matching backbones — different revisions produce
non-comparable embeddings. |
| Band | The embedding tensor inside a knowledge model — the
base band, 768d, always present. It is stored
compressed inside the .rlat archive. |
| Base band | The universal 768d gte-modernbert-base output. Every
knowledge model has one. Cross-knowledge-model operations always use it,
because it is the only band comparable across corpora (the
cross-model rule). |
| Bundled mode | A storage mode where source files are compressed inside the
.rlat archive itself. Self-contained — drop the file on a
USB stick, ship it to a colleague, query offline. |
| Capture surface | Any entry point that lands knowledge in the .rlat
file: the passive miner (opt-in, via rlat memory install-hooks
--mine), rlat capture-attribute,
rlat capture-env, and rlat memory add. |
| Citation | The provenance record on a retrieved passage:
(passage_idx, source_file, char_offset, char_length,
content_hash). Lets any result be traced back to the exact byte
span it came from. |
| Confidence lifecycle | The path an insight takes from when it is first created to when it is trusted or retired. A new insight enters provisional; real task outcomes then raise its confidence if it keeps paying off, or retire it out of retrieval if it is falsified. No up-front approval is part of the path. |
| Confidence ladder | The four rungs a memory lesson or an
insight can sit on:
low → medium → high → verified. Nothing is declared
confident; it earns each rung from accumulated evidence. Recalled items
always show their rung, so a low-confidence guess is never mistaken for
a verified one. |
| Connected-components clustering | A single-linkage clustering algorithm used by the dedupe operations. Two items cluster together if any chain of pairwise-similar links connects them, even when the two endpoints are not directly similar. |
| Constrain mode | An opt-in grounding mode that tells the consumer LLM to answer only from the retrieved passages and to refuse explicitly when the passages do not cover the question. There is no body suppression — the LLM does the refusal. The recommended posture for fact extraction, compliance, and audit work where a wrong-but-confident answer is worse than no answer. |
| Constraint (claim kind) | A standing hard rule of the world the knowledge model covers.
Constraints are served in full on every context serve — not retrieved
by similarity — so a hard rule is never missed because a query didn't
happen to resemble it. Captured via
rlat capture-attribute --kind constraint. |
| Content hash | A SHA-256 of a passage's source bytes, recorded at build time. At query time the live source bytes are hashed and compared — a mismatch is drift. |
| Cross-model rule | Any operation spanning two knowledge models always uses the base band — the band every knowledge model shares, so the numbers are always directly comparable. |
| CLS pooling | Taking the [CLS] token's hidden state as the sentence
embedding. The pooling convention rlat's encoder uses. |
| Deep-search / deep-research | The multi-hop research loop — plan, retrieve, refine, retrieve again,
synthesize — that returns a faithfully-grounded answer plus an evidence
union of citations. Two surfaces run the same loop: the
deep-research skill drives it inside a Claude Code session
with no API key; the rlat deep-search CLI verb exposes it
for non-Claude-Code agents, CI, and batch jobs and needs an Anthropic
API key. |
| Drift | Divergence between a passage's recorded
content hash and the live source bytes. Three statuses:
verified (hash matches), drifted (source
exists but the hash mismatches), missing (the source file
is gone). Computed at query time. |
| Drift cascade | When a source file changes, the insights
built on that source are automatically marked stale and drop out of
retrieval until re-verified. This keeps an earned summary from outliving
the ground truth it was drawn from. --include-stale
overrides the cascade when you want to see stale insights anyway. |
| Encoder | The gte-modernbert-base model that turns text into 768d
L2-normalised embeddings. A single recipe — no preset, no pooling
toggle, no projection knob — pinned to a specific Hugging Face
revision. |
| Expertise primer | A synthesised markdown summary of what the assistant has earned
about a project — the durable distillation of corpus, memory, and intent
into directly usable context. Produced by rlat expertise
and read at the start of a session so the assistant begins already
oriented. |
| Faithfulness gate | The check a deep-search answer must pass before it can become a provisional insight. An answer is faithful when every claim in it traces to a cited source passage. The gate replaces the retired up-front approve/reject step: an insight is kept because it is grounded, not because someone signed off on it. |
| Falsified finding (negation) | A record that something was tried and proved not to work, with its
evidence in the text. Served with the "Tried and falsified" framing so
an assistant stops re-recommending dead ends. Captured via
rlat capture-attribute --kind falsified. |
| Field | The router layer of rlat's three-layer architecture — the encoder plus dense cosine retrieval plus the ANN index. |
| Flat memory | The memory model rlat uses: one flat store of short, polarity-tagged lesson rows, each with a recurrence count and a confidence rung, scoped to a workspace. There are no tiers, no half-lives, and no tier-weighted recall — a lesson's standing comes entirely from its confidence and how it has performed. See the claim system page. |
| Grounding mode | The directive stamped at the top of LLM-facing markdown by
rlat skill-context and rlat search --format
context, telling the consumer LLM how to weight passages against
its training. Three modes: augment (default),
constrain, and
knowledge. Selected with --mode.
The directive header always ships, even when the dynamic body is
suppressed. |
| Insight layer / insight passage | The middle of rlat's three layers. An insight passage is a compact, cited summary earned from past research — every claim in it traces down to source. Insights carry a confidence and are served alongside source passages at query time, each clearly labelled. They enter through the faithfulness gate and live or retire by their confidence lifecycle. Insights are derived from the corpus; captured world claims — facts, constraints, falsified findings — come from you or your sessions and live alongside them, each labelled by kind. |
| Intent | A live record of what a user is trying to accomplish. rlat
intent manages them — add declares one,
list shows the active ones, path shows how an
intent decomposes into steps, activate switches the live
one, and accept / reject record whether an
intent was satisfied. Captured intent lets the harness condition
retrieval and attribute outcomes. |
| Knowledge mode | An opt-in grounding mode that tells the consumer LLM to treat the passages as a supplement to its training knowledge for the corpus's domain. It uses a lighter weak-retrieval gate than augment, because the directive already invites training-blending. Choose it for domains where training data may be stale but partial knowledge is still useful. |
| Knowledge model | The canonical name for a .rlat archive — the queryable
index built from a corpus. Always "knowledge model", never
"cartridge". |
Lens / .lens file |
A portable layer that carries one person's perspective on a
knowledge model — which sources they trust, which earned insights have
served them well, an optional written stance, and their accumulated
memory. A lens names patterns of trust rather than specific files, so the
same .lens can travel to a different knowledge model.
Several lenses can be merged into a team lens. A lens selects and
ranks, and is the review surface for what has been captured — it never
creates knowledge. See the
claim system page. |
| Local mode | The default storage mode. Source files stay on disk; the
.rlat carries embeddings, coordinates, and a recorded
source root. Drift detection works against
live filesystem state. |
| Memory confidence ladder | The confidence ladder as it applies
to memory: every lesson sits on low, medium,
high, or verified. Outcome corroboration,
corpus verification, and direct user confirmation move a lesson up;
confidence also orders recall, so the most trusted relevant lesson
surfaces first. |
Name-check / --strict-names |
A distinctive-token verification applied when grounded output is
emitted. It extracts proper nouns, acronyms, and alphanumeric IDs from
the question and presence-checks each against the retrieved passages. By
default it prepends a refusal directive when a token is missing from all
passages; under --strict-names it exits non-zero instead.
Catches the name-aliasing failure that score-based gating cannot. |
| Outcome ledger | The record of how tasks actually turned out — did a task succeed, and how much had to be re-done. Outcomes are what make confidence honest: a memory lesson or an insight is trusted because the sessions that used it went well, not because it was written down. |
| Passage | A chunked text fragment from a source file — the unit retrieval works over. Built by rlat's chunker, which falls back from paragraph to sentence to a hard split, with defaults of roughly 200–3200 characters per passage. |
| Passage_idx | A zero-based row index identifying a passage. It maps to a row in every band tensor and a line in the registry — the bridge between the field layer and the store layer. |
| Probe | The rlat probe command — a diagnostic that inspects how
a knowledge model behaves on a query or set of queries, used to
investigate retrieval quality without changing anything. |
| Provisional insight | An insight that has just passed the faithfulness gate but has not yet earned confidence from outcomes. Because it is grounded, it is safe to serve right away — it is served carrying a visible provisional confidence, so the reader knows it is new. From there it hardens if it keeps paying off or retires if it is falsified. |
| Re-verification | The rlat reverify command. When source drifts and the
drift cascade marks insights stale,
re-verification walks those stale insights so each one is checked again
against current source and either kept or retired. |
| Registry | The list of records inside a knowledge model mapping each passage_idx to its source coordinates. |
| rlat | The command-line entry point. resonance_lattice is the
Python module; rlat is the command. |
| RQL (Resonance Query Language) | The typed Python API surface — a curated set of operations grouped Foundation / Comparison / Composition / Evidence / Experimental. Some CLI subcommands wrap RQL operations; the rest are library-callable. |
| Remote mode | A storage mode where source is resolved over HTTP(S) against a
SHA-pinned manifest baked into the .rlat at build time. The
first query downloads cited source to a per-knowledge-model cache, with
SHA verification on every read. |
| Skill-context | The rlat skill-context command, designed as the body of
an Anthropic skill !command block. Its stdout is markdown —
query-relevant grounded passages with citation anchors, per-passage drift
status, calibration data, and a grounding
mode directive — that replaces the placeholder before the model sees
the skill. |
| Source coordinates | The (source_file, char_offset, char_length) triple
identifying a passage's exact byte span in its source file, plus
content_hash for drift detection. |
| Source layer | The bottom of rlat's three layers — the raw passages from your
files, hash-verified and drift-checked. Source is ground truth: it never
changes meaning, and everything in the
insight and lens layers
traces back down to it. --source-only restricts any search
to this layer alone — the honest baseline. |
| Source root | The filesystem directory that a knowledge model's
source_file paths are relative to. Local mode needs it; it
is recorded at build time, and the CLI --source-root flag
overrides it. |
| Stable passage_id | A hash of a passage's source-file slice that survives
refresh and sync deltas. A passage that changes
content but keeps its file slice keeps the same id; one that moves within
a file gets a new id. Stable ids keep citation chains intact when
passages are added or removed. |
| Store | The authoritative content layer of rlat's three-layer architecture. It resolves source coordinates to text and computes drift status. Three storage modes: bundled, local, remote. |
| Trace | The rlat trace command. Follows a single
insight all the way down to the source
passages it rests on, so you can see exactly what an earned summary was
built from. |
| Verdict lifecycle | The path an insight follows as evidence accumulates: it enters provisional, hardens up the confidence ladder as outcomes corroborate it, and retires out of retrieval if it is falsified or if the drift cascade marks its source stale and re-verification cannot save it. |
| Verified retrieval | The contract that every passage in every .rlat carries
source provenance and a content hash, so a query result can be cited
back to source and drift can be detected. rlat's single biggest
architectural advantage over opaque-embedding-store retrievers. |
| Watch | The rlat watch command — a live, silent,
self-discovering refresh loop running the same incremental delta-apply as
rlat refresh. Filesystem events debounce-trigger a
per-archive refresh. --once is the synchronous one-shot for
CI and pre-commit. Local mode only. |
| Workspace | The scope a memory store belongs to — usually a project, resolved from the git repository. Lessons learned in one workspace stay attached to it. |
| World fact / attribute (claim kind) | A stable fact about the knowledge model's world — a capacity, a version, a region, a policy. Captured from a single mention; at serve time the newest value per subject wins, so a corrected fact replaces the one it supersedes. |