Add the signal core: five shapes, features, one-way seed derivation

[guillaumelauzier]

What

First slice of the reposition toward generative art made from personal data. Backend only, not yet wired to routes — the pure core is worth proving before source-specific parsing multiplies against it.

Personal data streams are unbounded in variety but not in structure. Every one reduces to one of five shapes, so an adapter's entire job is producing one of those and everything downstream sees nothing else. Adding a source stays additive rather than multiplicative.

Shape	Examples
time_series	heart rate, glucose, steps, spend
interval_series	sleep stages, calendar events, trips
point_events	songs played, messages, transactions
spatial_trace	a run, a commute, a year of travel
categorical	genres, merchants, contacts

The chain: signal → features → seed = H(features ‖ salt) → artwork, feeding the mint seed slot that already exists (0015_mint_seed.sql).

Changes

migrations/0021_signal_core.sql — data_sources, signal_imports, derivations, consent_records.

There is deliberately no column anywhere that holds raw data, a sample series, or a feature vector. Adding one would silently turn this into a health-data store subject to GDPR Article 9. Only the seed, a hash commitment to the features, and coarse descriptive metadata persist. The schema is the enforcement.

src/signals/shapes.ts — the five shapes plus a strict validator. Strict on purpose: a malformed signal still produces a seed, and a seed gets minted permanently. It rejects millisecond timestamps (the classic adapter bug — silently yields a window ~50,000 years wide), non-finite values, reversed or zero-length intervals, and out-of-range coordinates, rather than quietly repairing them.

src/signals/features.ts — pure statistical descriptors, the irreversible narrowing step. Includes masd (mean absolute successive difference) and fragmentation specifically because they carry the character of a signal — a restless night versus a still one — which is what makes a piece recognisably someone's rather than decorative. Spatial features record bounding-box extent only, never coordinates, which would identify a home address.

src/signals/derive.ts — canonical encoding at fixed precision so the same night hashes identically across engines (IEEE-754 differs in the last bits), plus a salted one-way derivation.

The salt is required and fails closed. Without it the feature space is a dozen rounded numbers — small enough to brute-force a published seed back to plausible health data. Hashing alone is not sufficient here. featureDigest deliberately excludes the salt so a holder can verify a commitment without the platform's secret.

Testing

22 new unit tests over the pure core — the highest-risk code in the pipeline:

• every shape validates, and each rejection path is covered
• feature extraction is deterministic; masd separates a jagged series from a smooth one with identical mean
• canonical encoding is order-independent and rounds away float noise
• the seed leaks no feature value — asserted directly, since the seed is published on-chain and pinned to IPFS forever
• a near-identical night yields a completely different seed
• rotating the salt changes the seed but not the verifiable digest
• derivation refuses a missing or weak salt

70/70 tests pass (48 existing unaffected), npm run typecheck clean, and the migration applies locally (11 commands).

Not in this PR

Source adapters, import UI, generators, and manufacturing. Routes are not mounted yet — this is the core only.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01QuSa99HABSrWRarT9BQ5Tz

---

Generated by Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	generatedart	31cc773	Commit Preview URL Branch Preview URL	Sep 05 2026, 06:32 PM