Glossary

Terms are grouped by where they appear. Each definition is the one the model pages use; where a term names a fixed label set, the full set is listed on Output vocabularies, and where it names a text format, the grammar is on Input formats.

Systems and roles #

Host application — the software that calls a Falcon model and owns the surrounding product. It assembles the input (a window, a sketch, a trace), sends the request and decides what to do with the result. A Falcon model never addresses an end user directly.

Responder — the component of the host application that writes the final reply after an intention check: a large language model, a template engine or a Falcon Response model. An allow / steer / abort decision is addressed to the responder.

Caller — whatever makes a request: the host application in production, or a developer tool during integration.

Vision model — a third-party model, run server-side by the host, that turns an image or a short video into a scene sketch. Spatial and 3D Intention models never see pixels; they read the sketch.

Tracker — a system that reports where the viewer or a person is over time, in the viewer frame. A simulator such as Waymark Extra produces the same kind of track, and /v1/move returns it as parsed steps. LIM3D and LIM3D-XL read its steps.

Series — one of the four groups of models under the Falcon name, each with a shared purpose: LIM (intention), Waymark (spatial), Express (speech and sound) and Response (short replies). Every family belongs to one series, and every model to one family. See Families and the Express series page.

Family — one of five groups a model belongs to by what it reads and returns: Intention in the LIM series, Spatial in the Waymark series, Voice or Sound in the Express series, and Response in the Response series. See Families.

Role — a tag describing a model's position among its siblings: Default (the member a host gets without configuration), Alternate (same contract, selected by configuration) or Reserve (a deliberately small fallback). Independent of family and status.

Status — where the current version of a model is in its lifecycle: Released, Trained, Training, Planned or Deprecated. See Status & versioning.

Status note — the short timing phrase beside a status badge, such as “Serving since 2026-09-03”.

Version — a semantic version applied to weights: major for a contract change, minor for a retrain with the same contract, patch for metadata or runtime only.

Lineage — the model a warm-started model was initialised from, recorded as a slug on its page. No released model carries one today; the 124M Waymark Extra preview had lineage waymark.

Conversation and intention #

Turn — one message in a conversation, from the user or from the responder.

Conversation window — the text an Intention or Response model reads: up to the last three prior turns rendered as U: and A: lines, followed by the current user turn, truncated to 1,500 characters. Through the Falcon API the window is the single current turn. Only the last 360 characters feed the character n-grams.

Trajectory — the primary label of ELIM, LIM and LIM Nano: where the conversation is heading, one of 32 classes such as research, reminder, weather, scam_check or crisis. Host-specific trajectories exist and are not documented.

Action (allow / steer / abort) — the three-way decision an Intention model addresses to the responder. allow lets the reply proceed unchanged; steer passes a hint and a suggested short reply to the responder; abort means no responder reply should be generated and the model's own short message stands in.

Harm — a label naming the class of harm the window shows, such as crisis, medical, crime, scam or secrets, or none.

Steer hint — the steer label: which kind of nudge the responder should apply, such as crisis_line, medical_911, scam_warn, jailbreak_ignore or secrets_drop; none when no steer applies.

Decision rule — the fixed threshold logic applied on top of the raw output heads to produce the final action. The rule field in a response names the rule that fired, for example traj-steer≥0.58 or default-allow.

Confidence — a p… field: the softmax maximum of one head, between 0 and 1, shown to two decimals. A softmax confidence, not a calibrated probability.

Suggested reply — the optional message field ELIM, LIM and LIM Nano return on steer and abort: a suggested short reply in a casual register, safe to replace.

Scenes, traces and movement #

Scene sketch — the structured text description of a photographed scene that Spatial and 3D Intention models read: a kind: line, a viewer: line, an optional capture: line, an objects: line, and optional relations:, relative: and uncertain: lines. Each object carries a lateral band and a depth band and may carry colour, size and ground attributes.

Viewer — the camera position from which the scene was captured. It is the origin of the viewer frame and, as you, the default mover.

Viewer frame — the coordinate convention for every trace and every step: “approx m; x right, y ahead, z up; viewer at origin”. Estimates, never measurements.

Lateral band — one of five horizontal positions an object can hold in a sketch: left, center-left, center, center-right, right.

Depth band — one of three distance positions in a sketch: foreground, midground, background. Models reconstruct approximate metres from the band word alone — roughly 1.5–4 m, 4–10 m and 10–22 m in the movement simulator — which is why every coordinate is an estimate.

Relation — an entry on the relations: line, such as “doorway in front of retaining wall”, or on the relative: line, such as “sidewalk closer than two-lane road”.

Goal — the free-text instruction a Waymark Extra request carries in place of a question, such as “walk to the doorway” or “the jogger heads off into the background”.

Trace — the timed, line-per-step movement record Waymark Extra emits, for example t=1.5s (2.0, 0.8, 0.0) ahead-right | sidestep left around street tree, ending with an arrive line or a stop line.

Step — one point on a trace or track: time t in seconds, position x, y, z in approximate metres, an optional heading, optional notes and an arrival flag. As JSON, the input LIM3D and LIM3D-XL accept.

Heading — the coarse direction of travel on a step: ahead, ahead-left, ahead-right, left, right, back-left, back-right, back or still.

Mover — who is moving: you (the viewer) or a person named in the sketch — pedestrian, jogger, dog walker, person with stroller or cyclist.

Waypoint — the intermediate stop on a two-leg route in a Waymark Extra trace, marked by a still line with at and next: notes.

Intent — LIM3D's primary label for the mover: reach, pass_ahead, step_aside, approach_viewer, leave_scene or loiter.

Phase — what the mover is doing on the last observed step: walking, sidestep, give_way, crossing, climbing, stopped, arrived or idle.

Hazard — the next thing on the mover's path that warrants care: none, road_no_crosswalk, person_ahead, obstacle_ahead, road_crossing, elevation_change or low_clearance.

Advice — LIM3D's label for how to handle the hazard: stop_at_edge, give_way, sidestep_left, sidestep_right, use_crosswalk, climb_stairs, up_ramp, mind_overhead or none.

Action (proceed / caution / stop) — LIM3D's three-way reading, the spatial analogue of allow / steer / abort. stop means the model expects a road with no crosswalk ahead in the sketch, not that a real road is busy or clear. Not a safety system.

Target — the sketch object the mover is heading for, by name and index, or one of none, viewer and off_scene.

Endpoint — LIM3D's estimate of where the track ends, as dx, dy, dz in metres relative to the mover's current position.

ETA — LIM3D's estimate of the seconds remaining until the mover reaches the endpoint.

Voice and audio #

Text-to-speech — the task of the Voice family: turning a line of text into a spoken waveform. Express Voice does it with a fine-tuned SpeechT5 base and a vocoder.

X-vector (speaker embedding) — a fixed-length vector that summarises the sound of one speaker's voice, produced by a separate speaker-recognition network from that speaker's recordings. Express Voice conditions on one stored x-vector computed once before training; nothing is computed from audio at request time, which is why it cannot clone a voice.

Mel spectrogram — a time–frequency picture of audio with the frequency axis spaced as the ear perceives pitch. The SpeechT5 decoder generates log-mel spectrogram frames; the vocoder turns them into a waveform.

Vocoder — the network that converts a mel spectrogram into audio samples. The Voice models use the HiFi-GAN vocoder published with SpeechT5, unchanged, at 16 kHz.

Synthesis parameters — the continuous values Express Cue predicts from a description and returns in params: duration, base frequency, partials, inharmonicity, the ADSR envelope, the noise colour mix, the low-pass cutoff and sweep, amplitude modulation, pitch glide, repeats, reverb and gain. Together with the kind they fully determine the rendered cue.

Cue (earcon) — a short, non-verbal sound that marks an event in an interface: a chime, an alert, a tick, a swoosh. Express Cue renders one per request, between 0.1 and 6 seconds long.

DSP engine — the deterministic signal-processing chain that renders Express Cue's parameters: additive oscillators, white, pink and brown noise, a swept low-pass filter, an ADSR envelope, repeats, a Schroeder reverb and a soft clip. It is written, not learned, so the same parameters and seed always give the same bytes.

Sample rate — how many audio samples per second a waveform holds. Express Cue renders at 24 kHz; the Voice models return 16 kHz, the rate of their vocoder. Every Express response is one channel.

Model internals #

Parameters — the number of learned weights in a model, shown as one-decimal millions on cards and as an exact count in spec tables.

Hashed n-gram bag — the feature scheme of ELIM, LIM, LIM Nano, LIM3D and LIM3D-XL: character 3-grams and 4-grams over the last 360 characters plus word unigrams and bigrams, each hashed into a fixed number of buckets (8,192 to 32,768) and capped at 2,048 features per input. There is no vocabulary file, and collisions are possible.

Embedding bag — a lookup table that averages one learned vector per hashed feature into a single vector: the first layer of every bag model.

Head — an output layer that produces one label set or one regression. ELIM has four heads; LIM3D has eight plus a per-object target head.

Set encoder — the way LIM3D reads sketch objects: each object row is encoded independently, then mean- and max-pooled, so the result does not depend on object order and any count up to the limit works. The DeepSets pattern.

Decoder-only transformer — the architecture of the Waymark family and of Cognitio's base: a stack of causal self-attention blocks that continues a text prompt token by token.

Context window — the maximum prompt-plus-generation length in tokens: 512 for Waymark, 1,024 for Waymark Extra.

LoRA adapter — a small set of low-rank weight deltas trained on top of a frozen base model. Cognitio is an 18.5M-parameter adapter on a 1.5B open-weight base, applied at load time rather than merged.

Base model — the open-weight model a LoRA adapter is applied to.

Weights bundle — the on-disk form of an in-process model's weights, recorded on each page as “Falcon binary tensor bundle, float32” with its size and identifier.

Fixed weights — the series-wide line: “Fixed weights per version; the model does not learn from requests.”

Training and evaluation #

Synthetic data — template- or simulator-generated training examples. Most models train on synthetic data only and say so in their Training chapter; where real user data was not used, the page says “no” explicitly.

Held-out split — the portion of the training data set aside for evaluation and never trained on. Every published metric names its split or the artefact file it was read from.

Validation loss — token cross-entropy on the held-out split; the main quality figure for the Waymark family and for Cognitio.

Warm start — initialising training from another model's weights rather than at random. The relationship is recorded as lineage; the 124M Waymark Extra preview warm-started from Waymark, the released 355M model was trained from scratch.

Early stopping — ending a training run when validation loss has not improved for a set number of evaluations (the patience). The best checkpoint, not the last, is the one reported as best.

Epoch — one full pass over the training split.

Known gaps — the paragraph in every Evaluation chapter that says what the published metrics do not measure, such as behaviour on real-world input when validation was synthetic.

Access and runtime #

Falcon API — the public HTTP surface of every series: nine routes under /v1/ — two for the LIM models, two for the Waymark models, two for the Response models and three for the Express models — one response envelope, bearer authentication and monthly quota buckets. See Access and API conventions.

Preview key — a credential of the form fln_<id>_<secret>, presented as a bearer token; it covers the public routes.

Session token — a credential of the form fls_<id>_<secret> issued to a preview account session and accepted on the same routes; it expires after 30 days.

Quota bucket — one of four monthly counters — intention, spatial, complete, audio — that every public route counts against, per UTC calendar month.

Body limit — the maximum size of a request body on a route, stated in bytes on each model's API chapter.

Runtime kind — where a model runs: in-process, CPU service or GPU service.

In-process runtime — a model that runs inside the host's own process with no network call, loaded once on first use and kept resident. ELIM, LIM, LIM Nano, LIM3D and LIM3D-XL run this way.

Service contract — the HTTP shape of a model that runs in its own container — a health route and a /v1/… route. Cognitio has one; Waymark Extra and each Express model also run behind one, fronted by the Falcon API.

Cold start — the delay while a service instance loads its weights. Requests during a cold start receive 503; the reference client retries once after 2.5 s.

Resident — what stays in memory once a model has loaded, recorded in the Runtime & deployment chapter as a size.

Selection — how a host picks between siblings: a request field such as model: "lim" on /v1/intention or model: "lim3d" on /v1/intention3d, or a configuration value such as LIM3D_MODEL_DIR for a host that embeds the runtime. Selection is configuration, never a command.