One self, many local forms.
A personal, local-first communication compiler that drafts in your voice - the right voice, for the right people, at the right reach.
Pando is named for the Utah quaking aspen colony: what looks like a forest of forty thousand trees is a single organism sharing one root system. Same deal here. You are one person, but you do not have one voice - you have a repertoire, and which register you speak in depends on who's listening.
'I have the idea of building a corpus of my writing style that can be used to write in my voice, with a dial from shitposting on Discord to professional business communications, effectively a temperature knob for language. Tone, word selection, conciseness, they all shift depending on who you're talking to. This could be overlaid with a relationship map: close family weighted at this point on the dial, customer correspondence at this weight, banter friends at that one. All adjusted for reach, a 1:1 conversation vs. a public post.'
Pulling that thread led through forty years of sociolinguistics (audience design, register theory, conceptual pacts) and into a harder truth: a universal style dial is wrong. Manner (rhythm, warmth, terseness) interpolates smoothly between contexts. Membership (jargon, in-jokes, server emotes, private references) do not. The meme that named the failure mode: "oops, I used personality 2 with friend group 6."
So Pando became two systems in one: a voice model (learn my registers from my own writing) and a social-risk compiler (never let material scoped to one audience leak into another). The second half turned out to be the genuinely novel part, and the reason the architecture looks the way it does.
Pando compiles an explicit communicative intent + an audience exposure description + owner-defined policy + a register state into an inspectable draft with a local certificate showing what it believed, used, checked, and remained unsure about. Generative models do the wording; they are untrusted realizers with no authority over meaning, permission, or action.
Non-negotiables:
- It never sends anything. Drafts only; the human presses send.
- Statistics never create permission. No frequency, familiarity, or model confidence ever widens who may see what - only explicit owner grants do.
- History shapes how, never what. Personal corpus evidence may influence wording and rhythm; it may never invent a fact, commitment, or relationship claim.
- Wider audience ⇒ narrower vocabulary. Adding viewers can only shrink the eligible sensitive-material set, never grow it.
- A draft that's appropriate but sounds like it could be written by anyone is a partial failure. Voice fidelity is a floor, not a tiebreaker.
This is a personal research project, not intended to be commercialized into a product.
The system vocabulary is botanical, because the namesake earned it:
| Name | What it is |
|---|---|
| Pando | The whole system: one owner, many context-specific realizations |
| Heartwood | The stable core of manner: global and shared style factors (rhythm, warmth, directness), estimated with partial pooling |
| Sapwood | The living layer: audience-indexed expressive markers (slang, emotes, reaction habits) with per-context usage, freshness, and lifecycle |
| Cambium | The growth boundary: the owner-approved semantic plan: which propositions, references, and obligations a draft may actually communicate |
| Tropism | Leaning toward the interlocutor: the owner's characteristic accommodation to the current conversation |
| Abscission | Controlled shedding: decay, dormancy, and retirement of stale markers (a fossilized voice deploys last spring's slang; leaves must fall) |
| Phenology | The study of timing: temporal estimation, rolling-origin replay, lifecycle calibration |
| Etiolation | Growth in darkness: the low-confidence state of a dormant relationship or venue, reducing automatic use until fresh evidence arrives |
| Sport | A branch that differs from its parent: a relationship-specific register component that has earned its existence by measurably beating the defaults |
Under the metaphor sits a typed compiler: IntentIR (what you mean) + AudienceIR (who can see it) + PolicyIR (what's authorized) + MannerIR (how you sound there) → an untrusted realization engine → DraftIR plus a local certificate, gated by a small trusted reference monitor that evaluates grants, validates overrides, enforces identity compartments, and denies all communication actions. Full detail lives in the spec.
The plan runs product-cheapest-first, with the founding research questions insulated from product decisions:
-1A no-ingest register-control experiment (paired probes, my own writing)
-1B walking skeleton
-1C routine core loop - the minimum viable compiler, on real low-stakes messages
0 trust boundary, schemas, thresholds
1 source/corpus viability (can my history actually be exported well?)
2 audience/policy compiler + direct exposure test
3 manner measurement: do my registers actually separate? (H1)
4 generation: does my corpus beat register instruction alone? (H2) + hardware
5 optional: discovery, temporal modeling, calibrated retrieval
6 counterbalanced prospective shadow pilot → build / no-build
Three founding hypotheses anchor everything, and a product stop cannot cancel them:
- H1 - my voice is a measurable repertoire: distinct, audience-conditioned registers, not one style;
- H2 - a corpus of my own writing produces drafts I recognize as mine, per audience, better than register instruction alone;
- H3 - audience/reach conditioning prevents cross-context misfires without making output generic.
The one-sentence project test: Pando is worth building if explicit intent plus audience-aware register control makes my real communication measurably easier and more context-appropriate than unaided or generic drafting, in drafts I recognize as my own voice - without ever autonomously widening disclosure or taking a communication action.
The full project doc/spec can be read at project-pando_integrated-spec.md. Fable 5, GPT 5.6 Pro, and Grok 4.5 were used together for the project evolution and iteration.
Pre-implementation as of 24-Jul-2026. Specification at v5.1.1 after several review-and-revision cycles (the spec's decision log and changelogs carry the full lineage).