Imprint
Canonical JSON
Machine-readable profile contracts with build metadata, support summaries, confidence, and compatibility warnings.
Evidence. Expression. Identity.
Public-first identity and expression profile compiler
Imprint analyzes human-authored artifacts — writing, messages, transcripts, notes, and documents — then compiles structured expression profiles that downstream systems can use without reading the raw private corpus.
Open-source · Local-first · MIT License
What it is
Evidence in, portable profiles out
Imprint is an independent expression compiler. It turns human-created artifacts into versioned, evidence-backed expression profiles that downstream systems can consume without direct corpus access.
- Local ingestion and private connector framework
- Deterministic classification and signal extraction
- Versioned profile compiler
- Public-safe JSON, Markdown, and first-run exports
- Mosvera, Broadside, agent, and CLI consumer contracts
- No-raw-text privacy and compatibility warnings
Evidence first — every claim answers why
Confidence ≠ truth — support strength, not certainty
Store less — durable signals over maximal capture
Local-first — private corpora under your control
Public-safe — synthetic examples only in git
Exports
Public-safe export contracts, ready for downstream use
Compiled profiles export as deterministic JSON, human-readable Markdown, first-run "What Imprint Learned" summaries, Mosvera expression overlays, and consumer-safe projections — all with compatibility warnings and no raw text, provider prompts, or private locators.
Markdown summary
Concise observed-pattern reports that explain what the profile is based on without overstating certainty.
First-run output
"What Imprint Learned" turns compiled profile data into a useful first experience, not a personality assessment.
Mosvera overlay
A bridge contract for expression summaries and avoid-patterns, while Mosvera keeps aesthetic intent separate.
What it is not
Clear boundaries, honest claims
Imprint observes expression patterns supported by evidence. It does not diagnose people, infer mental health states, or simulate a person.
Imprint is
- An expression profile compiler
- Evidence-backed and versioned
- Provider-neutral (BYOM / BYOP)
- Designed for portable export contracts
- Local-first by default
Imprint is not
- A memory system or data lake
- A publishing platform or model router
- A voice-cloning or digital twin service
- A personality test or diagnostic system
- A writing automation product
Profile theory: Imprint may describe observable expression patterns — "frequently uses operational evidence before generalization" — but never generates diagnoses, personality claims, or intent inferences.
Why it matters
Portable profiles, not pasted prompts
People increasingly use AI to draft, publish, and represent them — but the reusable contract for a person's expression is weak. Raw few-shot prompts leak private samples, age poorly, and cannot explain confidence. Memory systems store facts but do not separate expression, provenance, and contamination.
Imprint exists to produce a durable, auditable profile contract that downstream tools can apply without owning your raw corpus.
Writing assistants
Optimize for immediate draft generation. Imprint produces the durable profile contract those tools can consume.
Memory systems
Store facts and preferences for agents. Imprint compiles expression signals — it does not compete as a memory store.
Digital twins
Simulate conversation and delegation. Imprint stays bounded to evidence-backed expression, not persona simulation.
Brand voice tools
Enforce approved language in enterprise workflows. Imprint exports portable profiles; downstream adapters own generation.
Current status
Built, reviewed, and running
Imprint is no longer just a project thesis. The repository now contains the full evidence pipeline: ingestion, classification, signal extraction, profile compilation, public-safe exports, consumer contracts, and a private connector framework.
Implemented pipeline
Adapters, connectors, classifiers, signal extractors, compiler, exporters, and consumer projections are implemented and tested.
Trust boundaries
Opaque source IDs, no-raw-text public exports, evidence-backed claims, version metadata, and compatibility warnings are built in.
Architecture reviews
Each major layer has been reviewed adversarially for privacy, scope creep, versioning, and boundary failures.
Current focus
Public web polish, future private-source adapters, profile comparison, and downstream integrations built on existing contracts.
Integration surfaces
Contracts first, integrations second
Imprint exports canonical profile contracts. Other systems consume projections of those contracts without owning profile inference, raw corpus access, prompt assembly, or runtime behavior.
Mosvera
Consumes expression overlays. Imprint compiles expression; Mosvera compiles aesthetic intent.
Broadside
Consumes publishing constraints and expression summaries, not prompts, workflows, schedules, or generation settings.
Agents & apps
Consume safe profile projections, evidence summaries, limitations, and compatibility warnings.
Private sources
Generic connectors support future private adapters while keeping real credentials, corpora, and source paths out of the public repo.
Privacy
Local-first, public-safe by design
This repository is designed to be public from the beginning. Real corpora, secrets, local configs, and private exports stay out of git.
No credentials required
Default local mode runs without cloud providers, private connectors, or network calls. Synthetic examples power public demos.
BYOM / BYOP
Bring your own model and provider. Remote inference is optional, visible when used, and never required for basic operation.
Export modes
Public-safe exports contain no raw private text. Richer metadata and excerpts exist only in ignored local paths you control.
Fail closed
Public-safe mode rejects raw examples. Provider secrets, private URLs, and account identifiers never belong in canonical schemas.