v0.1.0-setup1 — Architecture complete, runtime unvalidated

First public snapshot. Honest status: Setup1 architecture complete, runtime unvalidated.

8/8 specialist architects shipped; all framework layers have templates, schemas, agents, workflows, and lint integration. Not yet end-to-end runtime-tested — no fork to an actual operational domain, no independent cross-validation. Hobby project — no commercial support, no roadmap promises.

The moat — predictive layer with invisible shadow hypotheses

Most AI assistants either don't predict counterparty behavior, or predict it in plain sight — which contaminates the prediction. The moment you read "the model expects Sarah to push back on Series B timing", you walk into the 1:1 framing the conversation around it. Self-fulfilling, self-preventing — either way the loop is broken.

Giovanni's predictive layer is three pieces, designed against this trap:

1. Branch-out (visible active simulation) — 3 likelihood tiers, no fake percentages, max t+2 horizon (further is fiction), hard-stop on shallow actors (no caveat-degraded predictions). /branch-out.
2. Shadow hypotheses (invisible — the moat) — predictions the principal never sees during the prediction window. Stored in memory/shadow/pending/. Not in digest, not in briefing, not in chat. They become visible only at /shadow-review, after the horizon date has passed and the outcome is structurally determined. Quarterly review runs an adversarial lookback: "strongest arguments this hypothesis was NOT fulfilled?" — default-skeptical on uncertainty, because generous verdicts corrupt calibration. >80 % accuracy triggers an immediate re-audit (usually means tier labels drifted).
3. Per-actor calibration (monthly) — /calibration-report aggregates hit rates per actor, per tier. Framework-level accuracy is meaningless; what matters is which specific stakeholders the model reads well. Score tunes the triage heuristic that gates branch-out runs.

Full binding rationale (anti-self-fulfilling rule, no-recommendation principle, canonical-moves discipline, calibration healthy-range bands) in docs/prediction.md.

What's shipped

Layer	Highlights
Predictive layer — the moat	Branch-out + shadow + calibration triangle. Anti-self-fulfilling shadow invisibility rule. Adversarial lookback at quarterly review. Per-actor monthly calibration with healthy-range bands (60–80 / 20–40 / 5–15 %).
4-layer memory	MAP → operational shortcut → topic shards → deep. Graduation criteria, hard limits (300-line / 2 % strikethrough), audit cadence (14 d light / 35 d full).
Living constitution	Anchored sections, commit-traceable, supersedes-pointer, auto-regenerating INDEX.
Per-stakeholder profiles	Sentiment trajectory as append-only time-series. 6-value relationship-type enum. Predicted-reactions section feeds the predictive layer.
Daily digest	12-step procedure. Parallel source-puller fan-out. Drift detection with 7-day ack expiry. Brief auto-gen for events ≤48 h. Shadow lookback integration.
Subagents	8 framework architects + 7 operational workers. Tool-scoped. Isolated context. Model-tagged.
Slash commands	/digest, /branch-out, /shadow-review, /calibration-report, /consistency-check, /market-radar, /review, /redline.
Adversarial review	SHIP / REWRITE / KILL verdict (no compounds). Strongest-counter-case requirement. Default-critical.
Governance + lint	11 pluggable lint rules. 8 hooks. INDEX/MAP auto-regen. Hard limits enforced. Classification rules.

Stats

• 16 agents (8 architects + 8 workers)
• 8 slash commands · 11 lint rules · 8 hooks · 8 generic scripts
• 13 memory templates · 14 Lattice-domain examples
• 1 living constitution template + 1 INDEX template + 1 governance config template
• 10 workflow / policy / design docs
• ~104 files · ~17 K lines

What Setup1 did NOT include

1. End-to-end runtime test against a real operational domain.
2. Independent cross-validation (same agent built schemas and filled examples — confirmation-bias risk remains).
3. Battle-tested fork-and-fill walkthrough (docs/setup-guide.md is architect-side intent, not validated).
4. CI/CD beyond pre-commit hook installer.

What comes next (Setup2)

Fork Giovanni into a clean repo, fill it with your domain content, run actual workflows. This is where the runtime gets validated. See docs/setup1-complete.md for the bootstrap summary and docs/setup-guide.md for the fork-and-fill walkthrough.

License

MIT — see LICENSE. Fork at your own risk. No SLA, no support, no roadmap.

Origin

Sanitized clean-room extraction from a private domain-specific implementation (expansion of an e-commerce platform into 6 EU markets); no proprietary content carried over. See docs/origin.md.