Back to the ChroniclesThe record of the record
What changed, and when
v1.11.15October 6, 2026A failed act says so, once, with a way to try again
- When an act cannot be composed, the page now says so in one sentence, "Something interrupted the act.", with one Retry beside it. No server text, no number. The way out at the top left is never covered by the notice.
- While a new story composes after a Disruption, the header shows the series name alone instead of the act you left, then names the new act as soon as its title arrives. When an act finishes speaking, the header names that act.
- A voice request the voice service rejects outright is passed over at once instead of being retried for a minute and a half in silence.
- The story is now told which kind of turn it received. An Untold is composed as a deliberate turn toward a history the familiar accounts left out, and a Destination opens on the ground that was named, rather than either reading as a random rupture.
- A line of narration that says what a character does not do, arriving under that character's name, is now spoken by the narrator.
v1.11.14October 2, 2026Pause is the only transport
- The stop audio button is gone. While an act is being spoken there is one control at the top of the screen: pause. It turns amber and reads paused while held, and the voice picks up on the same second when released. Leaving the story is still the way out.
- There is no audio-off mode any more. The script appears as the voice reaches it, always. A listener who wants silence turns the device down.
- The framework the site is built on was updated to a release that closes a security flaw in how the site answers certain crafted links.
v1.11.13October 1, 2026A whisper is a whisper
- A quiet line is no longer whispered. One line in four that carried a delivery cue was being performed as a whisper, when the script had only asked for quiet, low, or soft, and a few had not asked for quiet at all. Now a line is whispered only when the script calls it a whisper. A quiet line keeps its voice and comes down.
- A hush now holds. When the travellers drop their voices because something could hear them, strangers within earshot or an animal close by, the answer comes back in the same hush, and it lasts until the narrator moves the scene or someone speaks up. When the quiet is one person's own, awe or a word that escapes, only that line drops and the answer is at full voice.
- The travellers no longer have the words to name what moves them. During a Disruption a character could say that the subject had changed, which is something only the audience knows. The phrase has been removed from everything the story engine reads.
- When the travellers meet people whose language cannot be rendered, a line that was only a description of what was said no longer leaves the story waiting in silence. The narrator says it.
- The five doors have been recorded again, so they carry all of the above.
v1.11.12October 1, 2026The recorder stops
- The end of an act now has a sound. As the last line finishes, a field recorder's stop is heard once: a click, a soft thump, the motor winding down, a breath of room, under a second and under the voice. The same sound marks the moment a break-in takes hold. It never plays when the page opens, never twice for one act, and never with audio off. A listener with the screen in a pocket can now tell an act that has ended from a connection that has dropped. The sound heard today is a stand-in made in code; a recorded one replaces it later.
- When a line of narration arrives wearing a character's name, it is now spoken by the narrator instead of by a stranger's voice. The script on screen is unchanged; only the voice that speaks the line is corrected, and only in live acts.
- A pricing page exists at /pricing, in plain words from one source of numbers: the two monthly plans, the top-up, when minutes renew, the refund rule word for word, and how to cancel. Its first line says the truth: nothing is on sale yet. The terms page links to it.
v1.11.11September 30, 2026A new generation of voice, and a floor that stops eating words
- Every voice in the cast now performs on a new generation of speech model. A line's delivery cue, quiet, pressed, steady, trembling, is given to the voice as a direction written before the line rather than as a set of dials, and the voice acts it. The narrator may now place one sound of the place inside a beat, where the ground earns it: never on a character's line, never a sound the travelers' devices make, and most beats carry none. The pace is a little slower than before; say if it needs a hand. Paying listeners hear the fuller model; everyone else hears its faster twin at the same quality of direction.
- A defect in the safety floor is fixed. Since late August, any line containing an ordinary word that happened to hold a short forbidden word inside it, the word about among them, was refused and never spoken, while its text stayed on screen. Short forbidden words are now matched as whole words, and spelled-out evasions are still caught.
- A cold start no longer closes the door. When the first check against the usage store timed out, every line for the next few seconds was refused and the app was told to wait five minutes. The check now waits once more, patiently, before it gives up.
- The curtain now dissolves into the first act instead of cutting to it: the book's light carries over into the field log's rule, one motion across the handoff.
- One word per thing. An act is an act and a story is a story, in the account panel, the top bar and the facts panel alike, and a test now refuses the old synonyms so they cannot return. The top bar's act position reads in words.
- A listener who has paid can reach the provider's billing portal from their account to manage or cancel. The route exists and answers that billing is not enabled until the payment keys are placed by hand.
v1.11.10September 28, 2026The room dims, and the site says what it is
- Pressing a door is now the moment the room dims. The ensemble behind the landing darkens and softens over the whole four-second curtain instead of holding at full brightness until the first word, and the blurred backdrop behind the script continues that same fade rather than starting over. The book's page lifts, turns all the way over and settles, once, before the first word. The line under the book reads at arm's length.
- The site now says what kind of thing it is, in one sentence, in two places: under the masthead on the landing page, and as the first block of the About sheet, marked What this is. The sentence is the author's: an audio drama, created in the moment, that you can guide. Three plain consequences follow it, naming only the two ways to guide a story. The earlier line under the masthead is retired; its idea remains in the About sheet under How the world works.
v1.11.9September 27, 2026The door is open to everyone
- Sign-in is open to anyone with a Google account. The sign-in sheet now leads with one filled button, Continue with Google, above the fold, with the Privacy Notice line beneath it. The three promises about memory are stated in the future tense, because memory is not yet built, and the line that no story is ever gated stays word for word. The notify-by-email form remains as the secondary path, below.
- A checkout route exists and is off. A signed-in listener will be able to start a purchase of one of the three plans from the wall. Today the route answers that billing is not enabled and reads nothing, so the wall keeps its waitlist promise. It turns on only when the payment keys are placed by hand, and it checks the shape of each price against what the product sells before it will start a purchase.
v1.11.8September 26, 2026The curtain gets its room back
- The opening curtain, the spine of light that unfolds into a book while a chapter is prepared, renders at its designed size again. A landing-page composition change in late August had constrained the stage it plays in to the height of the entry row, which flattened the book to a strip and collapsed the spine at most widths. The stage now grows to the scene's height for the duration of the curtain and returns to the compact composition afterwards.
v1.11.7September 26, 2026A sent door shows its title
- A door link pasted into a message now unfurls with that story's title and description rather than the generic site card. The tags are emitted on the server for the named door only, so link crawlers see them without running the page; an unknown or absent identifier keeps the site's own tags.
- The payment webhook endpoint exists and is closed. It verifies the provider's signature over the raw request body within a five-minute timestamp window before reading anything, records each event identifier before acting so a repeated notice is a no-op, and answers 503 with no secret configured. It is exempt from the same-origin gate because provider servers send no origin, and that exemption is held to by a test. Nothing is switched on: no secret is configured and no provider account exists.
v1.11.6September 25, 2026A door can be sent, and the world has a byline
- Any of the five front-page stories can be sent as a link. The link opens the landing page with that story marked ready and holding focus, named in one sentence above the list, so a single press starts it; nothing plays until the recipient presses, because a stranger did not ask for audio. A share control under the opening act's title uses the device's share sheet where one exists and copies the link everywhere else, and states on screen which it did. Only the five curated openings can be shared; a typed destination never carries a share control. Unknown or absent identifiers fall back to the ordinary landing page.
- The About sheet now states what is written and what is live. The cast, their histories, the way each speaks, the relationships that carry between stories, the rules the world runs on and the disciplines every story keeps are written by Matthew Suber and published by Human Commons; each act is composed inside that world at the moment it is asked for, after the record has been checked and gathered. The disclosure line and the terms page carry the same distinction, and the page's structured data names the author.
- The billing decision layer is in place and inert. A pure function maps a normalised payment event to exactly one ledger action, with idempotency guaranteed by an event-id table, and tier changes never overwriting owner or invited accounts. It has no network access and no keys, and nothing is switched on: the schema exists as a migration file only, and no payment provider is configured.
v1.11.5September 24, 2026Terms of use, and voice direction by shape
- A plain-language Terms of Use page is published at /terms and linked from the footer beside the privacy page, the sitemap and the crawler index. It states what the product is and is not, that stories are generated and can contain mistakes, what the product refuses to present as history, acceptable use, that a parent or guardian should be present for a child, how to reach the publisher, and that the terms may change. It is not legal advice and does not claim legal review.
- The companion units' voice direction now describes the shape of each unit's speech, its vocabulary field, rhythm, what it notices and what it never says, and carries no example utterance. A test over the unit definitions fails on any quoted span of three or more words, closing the path by which a directive's example could be recited verbatim in a listener's ears. Cast, alias, sub-voice and voice-id tables are unchanged.
v1.11.4September 23, 2026The content level is stated
- The history panel now states the content level a chamber was written at, in plain words, as one line beneath the facts. The engine has reported the effective level on every act since late August; the client now reads it. The three levels are validated against the server's own definitions by a shared test, and an absent or malformed value shows nothing rather than a guess. There is no control: the shipped level remains the gentlest of the three, and a settings surface is deliberately not part of this release.
- A measurement script for the speaker-slot guard, which detects narration carrying a character's name, runs the guard over the pre-rendered front-page acts and reports every flagged line with its speaker label. It changes no behaviour and holds no keys; its purpose is to put a number on the guard's precision before any decision to connect it to the render path.
v1.11.3September 22, 2026Alternate histories are labelled
- A chamber that imagines a history which did not happen now says so on screen. A persistent label sits beneath the act title for every act of that chamber, and the history panel opens with a sentence stating that the real record follows. Both are exposed to assistive technology and never appear on a mainline chamber. Previously the only signal was a change of narrator, which is inaudible to a screen reader and easily missed by a listener.
- The redaction layer that keeps held material out of the model's context is now covered by a regression test that builds the production prompt and asserts no redaction pattern matches it. A failure identifies the pattern by position only. Two guard assertions prevent the test from passing on an empty prompt or an empty pattern list.
- Both changes were built and verified by the autonomous build lane, including a headless-browser check of the label in both states, and merged under the repository's protection rules.
v1.11.2September 20, 2026Verifier judgement hardening
- Entity extraction now covers the minimal request. A destination consisting of nothing but a personal name is treated as a claim that the person is part of the historical record and is verified as such. Previously a name that asserted nothing else could be read as asserting nothing, which left a class of unverifiable requests outside the gate's scope. Real individuals requested by name alone, and generic subjects containing no named entity, are unaffected.
- The substituted destination is validated in code. When an unverifiable name is removed from a request, the phrase that replaces it becomes the story's subject, so it is no longer accepted from the verifier unexamined: it must be a short phrase, not a sentence or an explanation, and it may not reintroduce the name that was declined. Where the verifier offers no usable phrase, the surviving destination is computed deterministically from the entities it did verify. A request with no verified entity is declined.
- Prompt hygiene for the verification layer. The verifier's instructions describe output shapes and contain no worked examples, removing a path by which an exemplar could contaminate both judgement and evaluation. A unit test holds the shipped template to that constraint, and no evaluation input may appear in the prompt it evaluates.
- The live verification corpus is extended with name-only requests on both sides of the boundary and runs daily under the scheduled monitor. The new rules are mutation-checked.
v1.11.1September 20, 2026Verification resilience and gate telemetry
- The verification gate now degrades gracefully before it degrades open. Each verifier call runs inside a per-attempt timeout and an overall deadline budget, with bounded retry and backoff on transient upstream faults: rate limits, overload responses, dropped connections and timeouts. Vendor retry-after hints are honoured up to a ceiling. Permanent faults, such as rejected credentials or a malformed verdict under deterministic sampling, are not retried, so a listener never waits twice for the same answer. Failing open remains the policy of last resort for infrastructure faults. It is no longer the first response.
- Truncated verdicts are recovered rather than discarded. When the verifier's structured output is cut off at its token ceiling, the request is re-issued with a wider budget. This closes a path by which a long list of named entities could exhaust the verdict and pass unverified.
- Every fail-open is observable. A miss carries a typed reason (timeout, transport, status class, truncation or malformed output), and production fail-opens are counted per reason per day in a series kept separate from refusals. The governance posture is explicit: an unverified request is a countable event with a zero-tolerance alert threshold, not a silent fallback.
- Operational assurance is automated. A scheduled monitor runs the full gate suite daily: unit and type checks, generator preflight, the live verification corpus, production health, scheduled-job wiring and the previous day's gate telemetry. The live corpus distinguishes a wrong verdict from an unreachable verifier, because the two demand different responses. The monitor records a heartbeat on every run, so monitor failure is distinguishable from system health, and both the verdict and the alert are computed in code rather than delegated to a model.
- The transport policy is tested through an injected I/O seam with no network or clock dependency. Each behavioural test is mutation-checked: removing the behaviour it names fails the suite.
v1.11.0September 6, 2026Verification before generation
- A verification gate now runs before any story is composed from a typed destination. Every named entity in the request, whether a person, an event, a place, an organisation or a document, is checked against the historical record by an independent model call that returns a structured verdict. An entity that cannot be placed is never written. Where verifiable context survives the unverifiable part, the story opens on that context and the interface states the substitution. Where nothing survives, the request is declined with a plain explanation. The gate fails open on infrastructure faults and closed on judgement, and an extended integrity policy in the generation prompt provides defence in depth.
- Temporal policy is computed rather than inferred. Destinations at or before the present are held to the full integrity standard. Dates beyond a ten-year horizon open as explicitly speculative fiction, labelled as such in the interface, with the sourcing layer suppressed. The interval between is unavailable: neither history nor fiction can be told honestly there.
- Sourcing shown to the listener is filtered for provenance. Entries that do not name a locatable source are dropped rather than displayed, and the panel is labelled as the basis the engine stated rather than as verified citation.
- Playback is resilient to loss of connectivity. Prefetch depth is measured in seconds of audio and has been raised from 9 to 45. Per-line fetches retry against a 90-second deadline with exponential backoff capped at 3 seconds; transient upstream and quota-store failures are retried, deliberate refusals are not. When the buffer is exhausted, playback waits on the outstanding line instead of advancing past it, and a status notice appears after 1.5 seconds. Verified by fault injection against the live pipeline.
- Speech rate is register-conditioned. The base rate has been raised, and each delivery register carries a tempo offset within a bounded range, so variance in pace is a property of individual lines rather than a single global multiplier. The effective rate is reported per line for observability.
- Perception is rendered as process. The companion units draw from a rotating pool of instrument-read devices at a deterministic rate, so a finding is read, partial or corrected, before it is concluded rather than stated as instant knowledge.
- Stratified ground is classified in code. A detector for caste, segregation and colonial arrangements drives plain naming early in a chapter and a single observational aside per chapter at a ratified rate, with rotation to prevent house phrasing.
- Two latent hash-correlation defects are fixed. Low-bit parity between related FNV-1a seeds had left several chamber shapes statistically unreachable; a per-draw finalizer restores independence without altering seeds that frozen content depends on. Distribution tests now cover every seeded draw.
- Build integrity: pre-rendered content is fingerprinted against the engine that produced it and preflight fails on staleness. An integrity regression corpus runs in preflight, with an opt-in live corpus for the verification gate.
- The landing composition is stable across viewports: a viewport-anchored title lockup, labelled destinations with hover descriptions, a phone layout with the action beneath the field, a full-measure script column, and the act position in the header.
v1.10.0August 31, 2026The front page and the player
- A new landing composition: the ensemble artwork leads the page and the entry, the destinations and the explanation share one screen.
- What the product is, how it works and the mailing list now live behind the info control in the top bar.
- Media Session integration carries cover art and transport controls to lock screens and car displays for the duration of a chapter.
- Leaving a story tears down audio completely; a defect that allowed playback to continue after exit is fixed.
- A generator configuration regression that shortened chapters is fixed; chapters run their intended length.
- Synthesis is conditioned on adjacent lines, so delivery carries across a conversation instead of resetting per line.
v1.9.0August 29, 2026Weight where it belongs
- A global intensity ceiling that capped every chapter, including those standing on genocide, famine and slavery, has been removed. The factual content was never affected; the telling is now permitted to carry it.
- Chapters on mass suffering no longer converge on one emotional register. Reactions diverge by character, the machine characters remain precise about cost rather than moved by it, and no one argues another into their position.
- A pre-arrival caution is shown for destinations that stand on atrocity or on revisionist narratives: a plain statement of what the ground holds, informational and non-blocking.
- Arrivals are restored. Chapters had begun with the travellers already present; they now enter, and the place responds.
- Incidental characters are cast and written distinctly rather than assigned interchangeable lines and arbitrary voices, and a character keeps the same voice for the length of a story.
- Audio renders at a higher quality tier, and chapters are written by a newer generator model at lower temperature. Variety is meant to come from the story, not from sampling noise.
- Separation arcs resolve within a bounded number of chapters, and the telling cross-cuts to the missing party so the listener holds information the searcher does not.
- A listening allowance is enforced with honest messaging: a chapter in progress always completes, and the notice states when the allowance reopens.
v1.8.9August 28, 2026Destinations are screened
- A destination classifier evaluates typed requests before generation, with graded outcomes: allow, adapt, redirect, decline, or route to care. Refusals are delivered as application copy, never in character.
- Under-specified destinations are interpreted rather than refused, and the interpretation is shown so the listener can see the reading and change it.
- The intensity setting now affects rendering, not facts: the length of silences, proximity to reactions, and where scenes end.
- Field-log identifiers are derived deterministically from the ground, so the same destination files under the same reference for every listener.
- This page collapses each release to a single line that expands on demand.
v1.8.7August 28, 2026Stage directions stay silent
- A guard prevents stage directions from being emitted as spoken lines. Two pre-rendered chapters that carried the defect have been rebuilt.
- The daily composition cap and the daily voicing cap are brought into proportion, so neither is reached while the other has budget remaining.
v1.8.5August 28, 2026Accounts, one origin, and craft direction
- Sign-in with Google, and full account deletion that removes identity, listening history and every associated record, verified after the fact rather than assumed. No story is gated on signing in.
- The site and the application now share one origin. This is what makes persistent identity possible: a browser keeps a session per origin, and two origins meant two partial identities that could not be joined.
- Craft direction added to the engine: open late, end on the turn, let a room behave rather than be described, let reaction carry weight. Measured against the same chapter without it, narration uses roughly a fifth fewer words at half the length per line.
- Spatial rendering is ground-aware: the engine selects the one or two spatial techniques a place actually supports and uses only those.
- Eight in-world technologies now have explicit capabilities, limits and costs. A device without a limit resolves a story instead of complicating it.
- Six internal canon inconsistencies were reconciled by restoring missing frames of reference rather than overruling facts.
- A pre-rendered chapter with a protagonist discontinuity between its opening and its continuation has been rebuilt.
v1.8.0August 27, 2026The front page explains itself
- The explanation of what this is and how it works now sits beneath the destinations on the front page instead of on a separate route.
- Mailing-list capture is a field and a button rather than a mail client handoff, which silently lost anyone without one configured.
- Horizontal overflow on phones is fixed: a decorative element wider than the viewport was allowing the page to scroll sideways.
v1.7.4August 27, 2026The fourth wall holds
- The narration no longer refers to the listener. The term was present in the engine's own instructions where it could be reproduced; it has been removed, the rule now rides every chapter rather than only the first, and an output filter provides a final check.
- A named character introduced without a voice now receives a consistent one instead of falling through to the narrator.
- A rarely appearing cast member keeps her own voice in every line she speaks.
v1.7.3August 27, 2026Lead rotation and dialogue asymmetry
- Chapter leadership is distributed at roughly two to one rather than four to one, so a listener sampling several destinations hears both leads drive.
- Ensemble visits are tuned to about two chapters in five. Frequency changed; who arrives is still determined by the ground.
- Dialogue asymmetry is a function of subject, not rank: whichever lead is on their own territory speaks at length and the other briefly, and the pattern inverts when the subject does.
v1.7.2August 26, 2026Pre-rendered chapters keep the opening sequence
- Pre-rendered chapters now play the same short opening sequence as live ones, which is also the moment their audio is loaded, so the first line never stutters.
- The opening label is accurate to the mode: a pre-rendered chapter is described as opening, not composing.
v1.7.1August 26, 2026Embodiment corrected
- Ren's companion is described consistently with canon as an embodied machine with mass, position and audible movement, rather than drifting toward the projected presence of Sage's companion.
v1.7.0August 26, 2026Instant destinations
- The five front-page destinations are pre-rendered, so their first chapters begin immediately. About twenty-eight minutes of finished audio drama sits behind them. Every chapter after the first is generated live.
- Chosen journeys no longer open with a description of travel. A chapter opens on the place, through its sensory particulars, before anyone states that the travellers have arrived.
v1.6.1August 26, 2026A way out, and a Continue that survives backgrounding
- The whole product lives at one address, so browser navigation had nothing to return to and an installed app had no browser at all. A persistent leave control now sits in every story.
- Continue no longer depends on the tail of a streamed chapter arriving. Locking the screen or switching apps mid-chapter could drop that tail and leave the control inert.
v1.6.0August 26, 2026Five ways in
- Five curated destinations on the front page: a temple older than agriculture, the deep Cretaceous, a stone city whose builders were denied for a century, the first programmer, and twelve seconds of powered flight. Free-text destinations remain.
- Installable as a home-screen app, with Media Session support for lock screens and car displays.
- First version of connectivity resilience: audio is prefetched ahead of the playhead, and a failed line is retried rather than dropped.
- Rarely, a chapter ends on a moment the travellers themselves do not witness.
- Ruptured journeys are attributed to no one; the travellers can suspect a cause but are never told one.
- An unplanned arrival has no briefed objective; orientation and survival come first and the chapter's purpose is discovered on the ground.
- A stakes rule: no chapter builds a constraint the next chapter dissolves by leaving. What is at stake is what leaving would cost.
- Reunions after separation seed the next chapter rather than closing the current one, and visiting travellers are placed where the ground calls for them and may carry over between chapters.
v1.5.0August 26, 2026Openings and endings vary
- Some chapters open on the people of the place mid-task, with the travellers arriving as an intrusion on a world already in motion.
- Some chapters close on an unexplained detail after the story has landed.
- Long-running mysteries keep a cadence: roughly one chapter in four carries none, deliberately.
- Two previously unreachable chapter types are now possible: a clue chapter, where something in the record is too exact, and a quiet domestic one carried by what is left unsaid.
- Tone follows the ground. Deep-time chapters read as wonder rather than gravity.
- Visitors who unsettle a timeline can bring its consequences with them.
v1.4.0August 26, 2026Chapter shape is selected before generation
- Every chapter had followed one structure: arrive, orient, witness, discuss, close. The engine now selects a chapter shape before composing, so variety comes from the telling as well as the history.
- Some shapes are not about the history at all: a traveller unwell from the crossing, or a chapter that turns on the companion AIs. The requested subject remains the spine.
- Every chapter carries a dramatic question the record cannot settle, and the ending answers it or explicitly declines to.
- The tone-to-ground rule is enforced rather than encouraged: a chapter standing on atrocity cannot be played for comedy.
- Visiting travellers arrive with their own business and their existing relationships, conveyed through behaviour rather than exposition.
- A visitor whose presence changes nothing is treated as a defect.
v1.3.0August 25, 2026The rare chapter, and untranslatable ground
- The rarest chapter type now opens on the listener's own travellers and makes the route there part of the story, and its frequency is gated so it is an occasional event rather than a common one.
- Cast voices speak only what a person would say; descriptions of speech and stage directions are no longer voiced.
- Where no translation is possible, the story renders the experience of not understanding from the travellers' side: what their companions can and cannot decode, and what the body language says.
- Visiting travellers are introduced through what they notice and pursue rather than through stated credentials.
- If the record store is slow to wake, the first request of the day waits briefly instead of failing.
v1.2.0August 25, 2026Public readiness
- Abuse protection is in place, which is the prerequisite for a public link.
- Interruptions are explained: a plain notice states what happened and when to return, instead of text continuing without a voice.
- A daily generation ceiling applies during beta. It is reached only after a long session and resets each day.
v1.1.4August 19, 2026Travelling together again
- Both leads are present in every chapter, and leadership alternates between journeys.
- The two companion AIs no longer speak over each other; the embodied one speaks when acting.
- A visiting traveller's own AI is distinguishable by voice and by what it can physically reach.
- Internal chronology for a recurring chapter is consistent, and ensemble appearance rates match their intended frequencies.
- Narration varies sentence length with the scene, and tonal cues are audible in delivery.
- Exposition moves to the narrator and the AIs; the travellers talk to each other.
- Time to first line is shorter.
v1.1.3August 18, 2026Arc awareness
- Chapters know their position in the arc: middles hold tension, the turn lands where intended, and endings are deliberate. The closing sign-off is reserved for a genuinely finished story.
- Per-chapter state carries who is where, who knows what, and what has been said, reducing jumps and repeated exposition.
- The field log opens every journey, not only the first.
- Redirects hold their new ground, and suggested destinations derive from the current chapter rather than a fixed list.
- Long chapters carry their thread across parts and remain fixed in time.
v1.1.2August 18, 2026A populated world
- Locals and incidental characters speak in their own voices, and a character met twice keeps the same voice.
- Companion AIs speak for themselves wherever they appear.
- Firmer historical integrity limits: what people of an era could have known or quoted, numbers stated with more precision than the record supports, and words attributed to real people.
- Chapters with research behind them lean on it; chapters without it are conservative rather than confidently wrong.
- Field-log numbers increment across journeys, characters keep their own idiom, and grave chapters hold their weight more reliably.
v1.1.1July 10, 2026Grave history carries its weight
- The cast's lighter verbal habits step back in grave chapters and return elsewhere.
- Locals in English-speaking settings speak directly, without a translation relay.
- This changelog is introduced.
v1.1June 2026Six new voices
- Every companion AI has a distinct voice; six join the cast.
- Non-English settings render locals in their own languages, translated within the story.
- Deeper character canon covering danger, the hardest eras, and the cast's relationships.
- Tighter narration discipline.
v1.0.4June 2026Stability and polish
- Smoother playback, fewer interruptions, and story-flow fixes.
v1.0June 2026The doors open
- Chronicles of Sage Orange launches: interactive, fully voiced journeys into real history with Sage, Ren and their companions. Change the subject mid-story and the world responds.
© 2026 Human Commons