Transmuter
A conversation archive contains your words, a model's suggestions, abandoned ideas, and decisions that may have changed. Transmuter would work through that material with its attachments, identify what each source actually says, and prepare grounded proposals for what the Lich could remember or use to develop its identity.
The name Transmuter and its two internal responsibilities are settled in this candidate study, reviewed 2026-09-11. Portfolio acceptance and executable contracts remain open; State of Work owns delivery.
From sources to proposals
The operator selects a source, purpose, permitted destinations, and processing bounds. Transmuter keeps a processing dossier while two internal stages work through the material:
| Stage | Reader question | Result |
|---|---|---|
| Extractor | What is here, who produced it, and how should it be named and classified? | Source-located units with attribution, proposed names, labels, and uncertainty. |
| Synthesizer | How does this relate to what is already known, and what could be adopted? | Evidence-linked knowledge and proposals for separately admitted use. |
For a first personal intake, a smaller model could identify and prioritize material from a bounded chat export or supplied phone history before a larger model synthesizes selected units. That split is a possible implementation of the two stages. Selection would need to account for deferred material and bring relevant corrections and contradictions into synthesis with the earlier claims.
Extractor and Synthesizer are internal parts of Transmuter, not nested Compositions, Extension Domains, fixed Agent identities, or independent coordinators. Naming and labeling belong to Extractor; no separate Labeler is introduced.
A chat archive and its images can yield a proposed preference, a project history, and a possible Persona revision. The operator's words remain distinct from the model's interpretations. A contradiction stays visible. Archive or Mirror decides whether to accept the exact proposal, and Transmuter records that outcome alongside what it prepared. In the designed memory path, Archive retains admitted records in the Phylactery; its Curator separately decides which revisions are eligible for recall.
Learn through Transmutation
Transmutation gives the two stages a method for learning from what they transform. Suppose an archive contains “keep answers short,” followed months later by “explain this design fully.” A model has summarized both as “the operator always prefers brevity.” The useful result must recover whose words support which claim, what their contexts permit, and how a later correction could reach the proposed lesson.
| Chapter | Question within this intake |
|---|---|
| Genesis | Could the statements express a general preference, a task-specific request, or a changed preference? Form alternatives from attributed units before settling their relation. |
| Trial | Compare each interpretation with the exact source passages and authorized surrounding context. Include the inconvenient statement; unprocessed material remains a gap. |
| Canon | Declare before extraction and synthesis what would support the proposed use: correct speaker, preserved scope, traceable evidence, and the receiving owner's requirements. More extracted claims or a smoother summary cannot substitute for these conditions. |
| Revelation | Separate the observed requests from the diagnosis that their contexts differ and the proposal for a scoped preference. The model's universal summary is another attributable claim to examine. |
| Distillation | Carry the supported distinction with its conditions, counterevidence, and proposed next use. A later correction should reopen the affected preference rather than erase unrelated history. No adoption may be the warranted result. |
These chapters can recur within either stage; they add no five-stage execution pipeline. Canon precedes the attempt. A later statement does not automatically defeat an earlier one: changed context, changed preference, and unresolved contradiction require different conclusions.
The Tree's relations sharpen the method. Chokhmah opens interpretations and Binah asks what supports their structure. Chesed gives a neglected source room to matter; Gevurah limits what may be inferred from it. Tiferet seeks a useful composition that preserves disagreement where it remains unresolved. Netzach sustains the search; Hod can acknowledge that the available material cannot support the desired conclusion. Their interaction matters more than assigning a separate Agent to each name.
The receiving relation reaches the handoff itself. “Prefers brevity” loses the condition that made the lesson usable; a scoped proposal carries that condition to Archive or Mirror. Likewise, an archive's description of the Lich may inform a Persona proposal without establishing what it should become. Distillation preserves a way back to correction within the dossier's privacy and retention bounds; the receiving owner still decides what may enter memory, identity, or training.
Sources and machinery
The candidate scope includes mobile shared items, conversation exports such as ChatGPT or Claude archives, social exports and links such as Facebook material, documents, images, recordings, and supplied system-inventory snapshots. These name intended inputs, not delivered adapters or current platform access. A device inventory remains a dated observation; live discovery and host control are separate work.
Reliquary would retain original bytes under its custody contract. Scout can supply admitted acquisition, while Prism and Echo supply media processing. Synthesizer's receiving-owner map distinguishes Memory, Persona, training, and code candidates. Each job selects the branches its purpose needs.
Using several Extension mechanisms does not make Transmuter a Suite. A promise that actively coordinates separate Composition-owned Invocations requires a Suite. This study supplies no general installation, self-extension, or autonomous evolution coordinator.
Computer-session capture without a service API
A call or working session could enter this candidate intake through a local recording, without integrating the meeting service's API. The operator would select microphone input and playback output as separate synchronized tracks, with optional video of one window or monitor. On Linux, PipeWire supports playback capture, while the ScreenCast portal offers desktop-approved screen sources. This records the selected local experience; it does not supply participant identities, separate remote-speaker streams, chat history, or undisplayed content. Whole-output audio may also contain unrelated applications and notifications.
The proposed recorder would run as an explicitly authorized helper in the logged-in desktop session. Routine capture generally needs the user's session permissions rather than root; PipeWire access control still applies to each client. A containerized or remote Vessel would need a narrow typed device integration under the host-effect boundary, without acquiring a general shell or ambient desktop access. Audio and visual permissions remain separately scoped.
An Altar control or CLI Intent could request the same bounded capture, optionally associated with a Project or call. The proposed controls would select sources, a private destination outside the checkout, retention, duration and storage limits, and expose active state and explicit stop. Finalization would return exact artifact references and locations, with gaps or partial results preserved; an unconfirmed stop remains unknown. Saved presets or later event triggers would reuse this contract under current authority rather than grant permanent recording permission. These are candidate interactions, not an implemented recorder, command grammar, or automation.
Recording and interpretation remain separate. Original tracks would retain their common timeline and acquisition facts; admitted speech transcription and visual extraction would produce source-located derivatives for Extractor and Synthesizer. Summaries, decisions and proposed tasks would link back to exact time intervals. Echo retains its speech boundary, and recording alone authorizes neither analysis, remote transmission nor adoption into memory.
The processing dossier
The proposed durable dossier lets the operator see what was admitted, what each stage produced, and what remains unresolved. It owns the processing purpose and scoped authorization, source manifest and fixed revisions, privacy/retention/resource bounds, and stage implementation and configuration revisions. As work proceeds, it keeps progress, extracted units, corrections, knowledge candidates, and their dependencies together.
For each proposed handoff, retain the exact payload, attempt identity, receiving receipt, and separate outcome. The dossier references original artifacts so a reader can return to the evidence; it does not become their byte-custody owner.
Completion means the admitted scope was processed and every proposed handoff has an accounted disposition. It does not require every proposal to be accepted. Partial work identifies unfinished sources or proposals; refusal names the blocker; cancellation stops new work. An unknown receiving effect remains unresolved until reconciled and cannot be reported as successful completion. Committed receiving-owner effects survive local cancellation.
The stage pages specify how their attempts can be retried or corrected. Deletion must follow the policies of artifacts, Archive, and Mirror, retain only permitted lineage, and mark dependent material invalid where required. That process can account for stored records and proposals; it cannot promise to remove influence from already trained weights.
What would establish the application
A bounded historical chat import and ongoing mobile captures are two independently useful cases for the same dossier and judgment. Each mobile arrival remains bounded; connecting a source grants no unlimited background collection. The candidate must establish exact schemas, Patterns, admission and custody, implementation, and receiving contracts before execution.
The first network-disabled synthetic fixture should include duplicate imports, changed source content, a model claim misattributed to the operator, conflicting preferences, an uncertain image label, a denied destination, cancellation and resumption, and an unknown downstream outcome. It must demonstrate provenance-preserving reuse, correction, honest partial completion, and refusal to repeat an unresolved effect. This is proposed acceptance work, not an implemented test or source scaffold.