Roadmap

Recently shipped

Regime-aware research substrate, autonomous factory loop, and a control group for measuring it — landed as a single arc:

• EventAgent + premise tags + regime snapshots — every thesis now declares which policy/regime variables it depends on, drawn from a controlled vocabulary the EventAgent uses to tag Federal Register events. Intersection fires PREMISE_INVALIDATION alerts.
• Policy-neutral cohort — ThesisCohort.NEUTRAL, paired-hedge positions, cents cohort for per-cohort spread P&L. The control group that separates skill from regime beta.
• Hypothesis factory v1 — cents factory init | run | status | analyze. Walks a universe of symbols; opens paper theses where the orchestrator clears the entry threshold (paired by default); closes on target / stop / horizon / premise-invalidation. Direction-aware (bearish signals open SHORT), with per-premise-tag concentration cap and conviction-weighted preemption.
• Universe abstraction — cents universe create | list | refresh | set-default. Pluggable sources (static / watchlist / FMP index).
• LLM usage + pricing — llm_usage table, cents.pricing, cents usage summary reports input/output/cache tokens and estimated cost by agent / model / day / operation. Visible bill for autonomous operation.
• Live ops cost observed: one full factory run on a 14-symbol universe costs ~$0.02 in Anthropic spend. Daily refresh adds ~$0.05.

Next horizon

Most of the original “v1.0 productization” list shipped above. The remaining work falls into three buckets:

Quality of measurement (highest leverage)

• Cohort-stratified regime analytics — once enough closed theses exist (~30+ per cohort), stratify outcomes by regime_snapshot features. Answers “which kinds of theses cents generates well in which regimes?” — the actual deliverable the factory was built to produce.
• Beta-matched sizing for paired-cohort twins — today 1:1 dollar; regress 60-day returns to size the short leg properly. Currently TSLA-vs-XLY is the canonical mismatch case.
• Backtest mode for factory — cents factory run --as-of YYYY-MM-DD against historical event timestamps. EventRepository already supports historical queries; the factory just needs to honor as_of end-to-end.

Operational polish

• Cron / launchd scheduling docs — recommended cadence + sample plist + crontab. Turns “I run it manually” into “it runs at 4am and I check alerts in the morning.”
• Additional event sources beyond Federal Register — SEC EDGAR 8-Ks, SCOTUS opinion + oral argument calendar, USTR / Treasury / OFAC feeds, FOMC schedule. EventAgent pipeline is source-agnostic; each source is a new _fetch_* method.
• Chain preemption to fit larger pairs — currently the factory preempts at most one lower-conviction thesis to open a new candidate. Extension: chain multiple to free more notional.

Productization (still meaningful)

• Web API (Cloudflare Worker) — POST /research/{symbol} returning the same AgentResult JSON. Same shape, different transport.
• Slack / webhook alerts — CENTS_WEBHOOK_URL is plumbed for scan alerts; needs payload formatting + rate-limit niceties.
• Active threshold auto-tuning — learn entry_threshold and preemption_margin from prior outcomes. Risk: overfitting on small N; defer until factory has produced 100+ resolved theses.

Out of scope (deliberate)

• Real-time streaming data. Cents is a research and review tool, not a trading terminal.
• Live execution from the CLI. The broker integration is paper-only by design; live trading stays opt-in and manual. (The cents-tyy bead tracks an explicit broker-flag path for users who want it.)
• Hosted multi-user product. Local-first, single-user is a feature.

Suggestions and counter-proposals are welcome on GitHub issues.