Gateway Forge

The bundled voice is a Piper/VITS model fine-tuned on real speech, and its training clips were trimmed tight — so the model learned to stop mid-breath. A short line like "Build the pattern." ended at rms 0.0061 with the mic still open, against roughly 0.0007 for a clean stop. The fix isn't a fade — fading a real recording eats the same consonant it's meant to protect — it's giving the model two extra frames of silence to decay into before the cut.

Three frames is worse than two — past a point the model has room to voice into the extra silence rather than settle into it. It's a measured optimum, not a rounder-sounding guess, over seven real lines and five draws each.

FFT analysis across all fifty original tapes, keeping only genuine offset pairs (a real binaural signal, not two centred peaks) and weighting by how long each one holds. The whole programme collapses to three signals:

The app never plays a looped recording of these — it regenerates the pair live from the measurement, so the differential can move continuously through a session's transitions instead of cutting at a seam. Transitions sweep rather than glide: the differential widens before it narrows, matching how a transition was actually described rather than a straight linear fade.

Native Swift and SwiftUI throughout, no web engine and no bundled interpreter — the build fails outright on any Python file anywhere in the tree, on purpose. Speech runs through the same Piper/VITS-over-ONNX engine Voice Forge was built to expose, here bundled as a single fixed voice fine-tuned on the author's own speech.

A checkout of this public copy has its corpus-dependent suites — the ones that read the original tape transcripts or the Institute's manuals — stand down by name instead of either failing on files that were never meant to ship here, or quietly reporting health they never actually looked for. note [source tapes] no tape transcripts in this tree — corpus suites stand down (this is the expected state of a build made for distribution) is what that looks like in practice.

A ledger the app keeps, not the Institute's recordings

The three retained resonant-tuning cues and the wake-up signal that used to come from sampled Monroe tape recordings are now synthesised: the bed generates its own tuning and its own return signal, live, at the measured frequencies above. Nothing sampled from the original tapes ships in this build.

v5.4.0 · 21 October 2026

Rendering no longer quits the app on macOS. In v5.3.1, pressing Render closed Gateway Forge on every Mac except the one that built it. The release was compiled on a build server whose Swift looks for the voice in a different place inside the app than the one the voice is installed in, and gives up by stopping the program. The readiness check looked in the right place and said the voice was ready, so nothing warned before the crash. The voice is now found by the same lookup the readiness check uses, and the build refuses any code that reaches for it the other way.

The journal has its own place. Journal sits in the side rail between Focus and Studio: every entry across every level as one dated log, carrying the session report's fields — start and end times measured by the player, session title, how the body felt, and your own account. One entry or all of them export as PDF, the same report the session report form on this site writes. Notes that used to live inside a session's folder move into the journal on first launch, and clearing out old sessions no longer leaves empty rows behind.

The whole default path builds and exports from one place. Studio ▸ Production can queue every lesson on the default path that has no session yet, then export every assembled lesson to one folder with the bed mixed in, numbered in path order so a phone lists them as the tapes teach them.

Commands get room. In the ten-point relaxation each command is its own line with a pause after it, and the three Affirmations are said in phrases with time to repeat each one. The wording is unchanged. Narration made before this update renders again.

The speech engine is shared. espeak-ng, the phonemizer and the voice's inference now come from voice-core, the package Voice Forge and Interpreter also use. Across all 310 library lines it gives the same phonemes as before.

v5.3.1 · 3 October 2026

The default path answers clicks where its lessons are drawn. On macOS, every row in the side rail took clicks two lines above where it appeared, so choosing a lesson opened the one below it. The rail drew its list a toolbar's height lower than it hit-tested it — a quirk of how newer macOS runs apps built for macOS 14 — and the list is now built so that both agree. Nothing else changes; the Windows and Linux builds are the same as v5.3.0's.

v5.3.0 · 25 September 2026

Windows and Linux assemble and play a session now. Previously the cross-platform build could read the library and sound the bed but not make a tape. The assembler and the session transport are both ported: it renders what is missing, lays the takes down with the silences the script asks for, records where each piece actually landed, and plays the result with the bed generated live on top — one audio clock, the pan envelope, the arrival hold and the return. Both builds now produce the same tape from the same takes byte for byte, checked against a fixture neither side can fake.

The panning that shipped in v5.2.0 was wrong. @pan applied to the whole session rather than the segment that declared it, so Headphone Orientation left every word after it in one ear — the opposite of the check it was added for. It is scoped to its own segment now, and exports carry it, which they did not before.

The resonant tuning never sounded. Nothing placed it: the step that would have put the hum on the tape did not exist, and a check actively forbade writing one. Both are fixed, and the hum is where the script says it is.

The Projected Energy Bubble ran backwards. Its creation described a current falling down the inside of the body and rising outside, while every line after it — the strands, the recreation, the dissolve — described the opposite. It is one circuit again: out of the crown, down around the body, in at the soles, up through you.

The ten-point relaxation counts to ten in both densities. At full detail it opened at four, leaving the first three points unspoken, while the bare count spoke all ten. Someone who has stepped down to anchors alone was being taught a different shape of the same induction.

v5.2.0 · 9 September 2026

Export a session as a single stereo file. An assembled session.wav is the narration alone — the bed is generated live underneath it while you listen — so copying that file to a phone would hand over a voice talking into silence. Export mixes it down instead: the same bed engine, over the plan the session was assembled with, at your own listening levels. Open a session and choose Export as WAV, or run gfrender --export-session.

The voice is panned, at last. @pan has been in the script language, and in every session the generator writes, while nothing between the parser and the speakers read it. Headphone Orientation has been asking listeners to confirm they hear the voice in their right ear — an instruction that was unfollowable, and that told anyone wearing their headphones correctly to turn them around. It reaches the audio now, and a check refuses any session that uses that segment without panning. Sessions already on disk are untouched and still play centred; only what is assembled from here on picks it up.

The iOS companion is suspended. Set aside rather than removed — the client, the LAN transport and their checks all remain, and the phone side will be picked up again. Nothing listens on the network and no device can pair meanwhile. Export as WAV covers the reason it existed.

v5.1.1 · 6 September 2026

The first release with Windows and Linux builds in it. Both are packaged by CI from the tag, checked before they are built, and published with their checksums — previously only a macOS build existed.

Resonant Tuning is now an original humming exercise. The old segment followed the Institute's own breathing instructions closely; it has been replaced with three level-scoped variants written from first principles — vagal tone, bone conduction, and the fact that humming quieter drives proportionally more vibration into the skull than into the air. Low and in the chest through Focus 10–15; rising through the throat and jaw at 18; deliberately quiet from 21 upward, where it is felt more than heard. Nothing in it is paraphrased from the Institute's material.

The Resonant Energy Balloon becomes the Projected Energy Bubble. The three-strand construction was already the author's own; what is new is the load-bearing logic said out loud — strands at different depths pulling against each other like woven fibre, triangulating into a shell — and a crown-to-feet circuit tied back to the breath rhythm the ocean segment already establishes.

Verbosity now follows practice, per level. Full detail by default; ten completed sessions at a level steps it down to guided, twenty to anchors alone. A month away from a level resets it — earned ease is for someone actively practising there, not a permanent credential. Any level can be held at a chosen verbosity from its own page, and a choice made that way does not decay.

An app icon, and one-line installers for macOS, Linux and Windows that verify checksums before they unpack anything.

v5.0.0

First public release. macOS only — the native SwiftUI app, the authored library, the live-generated bed, and the ledger that keeps a listener's own practice on their own machine.