[sudoTomas]

What changed

Two new Sources, both shadow signals: computed and journaled on every decide of a Bot that subscribes, scored against later bars, and kept out of the decide prompt unless a per-Bot toggle is on (off by default).

• kronos (src/trader/data/kronos.py, KronosSource): sends the symbol's last 400 1h bars to a Kronos-small sidecar over HTTP. The sidecar runs each path as its own batch row with sample_count=1, so the N sampled paths stay separate instead of being averaged. The trader summarises them as share of paths up, plus median, p10 and p90 return TRADER_SIGNAL_HORIZON_BARS (24) bars ahead. The model code is pinned by commit and the weights by Hugging Face revision; the trader journals that revision with each forecast.
• crypto_derivatives (src/trader/data/derivatives.py, CryptoDerivativesSource): BTC and ETH USDT-perp funding rate, open interest and its 24h change, and mark-vs-index basis from Binance, Bybit and OKX through ccxt public endpoints (no keys; ccxt added as a dependency). Fetched on each tick and never stored as market data. A venue that fails drops out of the reading. Replay reads nothing from it, because today's venue numbers are not point-in-time.
• Journal: new column decisions.signals (migration 0019) holds {source: {version, as_of, in_prompt, reading}} for every reading, in the prompt or not. Replay steps record it too.
• Toggle: set_signal_prompt(name, source, in_prompt) on services, API (POST /bots/{name}/signal-prompt), MCP and CLI (trader bot signal NAME kronos --in-prompt). Each flip is a new, audited Bot version. Bot Flip briefs into the decide lane (was EAI-192) #1 is refused: it follows Settings and subscribes journal-only with TRADER_SHADOW_SIGNALS_ENABLED.
• Scoring: signal_scores (GET /signals/scores, MCP get_signal_scores, trader signal-scores) scores the journal against bars fetched for that call only:
  • Kronos band coverage (target 80%) and directional hit rate, per symbol and overall.
  • Rank IC of funding, OI change and basis against the forward return.
• Sidecar: infra/kronos/{Dockerfile,server.py} plus a compose service under the kronos profile. It has mem_limit: 2g, a GPU reservation and a port bound to localhost only. Runbook: docs/runbooks/shadow-signals.md.
• Docs: settings in ops/config.py (regenerated docs/configuration.md and .env.example), the "Shadow signal" glossary entry, and an architecture note.

Acceptance criteria

• Sidecar memory stays under 2 GB: the compose service has a 2 GB hard cap, and GET /health reports peak_rss_mb. The runbook also says to check docker stats and nvidia-smi on the unified-memory box. This needs the operator to run it; CI has not run it.
• Inference time per symbol logged: the sidecar prints one JSON kronos_forecast line per call with inference_ms. The trader logs kronos_forecast and journals reading.inference_ms.
• A report after two weeks of shadow data: the scoring tool and the criteria are in the runbook. The report itself can only be written after two weeks of data have been collected.
• Per-Bot toggle, off by default: tested (test_the_toggle_*).
• Trader works with the sidecar down: tested (test_the_sidecar_down_drops_only_its_reading).

How to switch it on

1. Build and run the sidecar on the GPU box: KRONOS_BASE_IMAGE=nvcr.io/nvidia/pytorch:<tag>-py3 docker compose -f infra/docker-compose.yml --profile kronos up -d --build kronos.
2. Set TRADER_KRONOS_URL (http://kronos:8095 or http://localhost:8095).
3. Set TRADER_SHADOW_SIGNALS_ENABLED=true for Bot Flip briefs into the decide lane (was EAI-192) #1, or subscribe a store-defined Bot to kronos / crypto_derivatives.
4. After two weeks, run trader signal-scores --days 14. Only then consider trader bot signal <bot> kronos --in-prompt, through the decide-change gate.

With nothing set, this deploys with no change to any prompt or decision.

Tests

tests/test_shadow_signals.py: 22 tests covering the clients (HTTP and ccxt mocked), the Sources (point-in-time Kronos bars), context and journal, the Bot spec and toggle, scoring, and the toggle and scores through every surface. Full suite: 953 tests, ruff clean, config reference --check clean.

Closes #144

https://claude.ai/code/session_01UDNLNu854ULHbHSL59UJKu