Self Improvement

Self-Improvement

MOSAIC's self-improvement stack has four parts: Autoresearch (prompt evolution), PRISM (multi-regime training), JANUS (cross-cohort meta-weighting), and MiroFish (reflexive simulation).

Autoresearch — prompt self-evolution

mosaic/autoresearch/ — selects an agent, has an LLM rewrite its prompt, commits the mutation on a git feature branch, runs a two-stage backtest to compute ΔSharpe, then keeps (merge to main) or reverts (delete branch) by threshold.

• git_ops.py (GitOps) — thin, fail-loud wrapper over git. Mutations are committed inside a throwaway git worktree so the operator's working tree is never touched. Keep = merge_to_main, revert = delete_branch.
• constraints.py — check_cooldown (24h per agent), check_monthly_cap (≤100/cohort/month), check_keep_lockout (3 days after a keep).
• evaluator.py — computes ΔSharpe over the evaluation horizon (default 5 trading days).
• decider.py — keep iff delta_sharpe ≥ keep_threshold_delta_sharpe (default 0.1).

Branch naming: cohort/{name}/auto/{agent}/{YYYY-MM-DD}.

Optional: mirror kept mutations to a self-hosted git server

autoresearch.git config (default OFF): when push: true, the keep-path runs git push <remote> main after a successful merge. A push failure is logged and swallowed (the keep decision still stands locally); a failed merge skips push. Credentials are the operator's responsibility (SSH key / credential helper). Config: { "push": false, "remote": "origin" }.

Defaults (mosaic/default_config.py → autoresearch): agent_mutation_cooldown_hours: 24, keep_revert_lockout_days: 3, keep_threshold_delta_sharpe: 0.1, monthly_modification_cap_per_cohort: 100, evaluation_horizon_trading_days: 5.

PRISM — multi-regime training

mosaic/prism/ trains prompt evolution across 7 market-regime cohorts, sequentially per cohort with bounded intra-layer concurrency. Cohorts (prism/cohorts.py):

Cohort	Window	Regime
bull_2007	2006-01-04 → 2007-10-16	牛市顶 6124
crisis_2008	2007-10-17 → 2008-10-28	暴跌 70%, 1664 见底
bull_2016	2016-01-29 → 2017-12-29	慢牛 + 白酒
crisis_covid	2018-10-19 → 2020-03-23	贸易战 + 疫情合并
recovery_2020	2020-03-24 → 2020-12-31	疫后宽松反弹
euphoria_2021	2020-07-01 → 2021-02-18	茅指数高峰 (启动 cohort)
rate_tightening	2022-04-01 → 2023-12-31	中特估 + 量化退潮 + Fed 加息

CLI: prism list|train|status|compare.

JANUS — cross-cohort meta-weighting

mosaic/janus/ computes softmax meta-weights across cohorts (so the live system blends regime-specialized prompts by current-regime fit). CLI: janus run|weights|regime|history; RPCs janus.run_daily|get_weights|regime|get_history.

MiroFish — reflexive simulation

mosaic/mirofish/ synthesizes forward scenarios from a behavioral agent swarm and grades recommendations against them.

• engine (config.mirofish.engine): montecarlo (default — i.i.d. correlated paths + optional reflexivity kernel) or swarm (agent-to-agent interaction). Swarm is opt-in.
• scorer (config.mirofish.scorer): terminal (default — direction × cumulative return) or path_aware (drawdown-penalized equity curve; the --path-aware shorthand on mirofish train).
• inject_context (config.mirofish.inject_context, default OFF): append the latest scenario context to the CIO prompt (see Agents).
• An OASIS adapter can drive a real external MiroFish engine over HTTP (MOSAIC_MIROFISH_URL).

CLI: mirofish generate|train|history; RPCs under mirofish.*.

Memory/persona (7M.2/7M.3) is deferred/NO-GO per gain validation; documented in mosaic-tsplan.md §11.8.1.