The observation half of the improvement loop: what DiscoveryLab measures about itself between runs.
Five self-improvement signals (biology versions, regression bench, calibration, active-learning convergence, what changed since last time) used to live in separate artifacts. SelfImprovementSnapshot aggregates them into one typed Codable surface — exposed identically through a CLI (peptiter-self-improvement), through the improvement/snapshot MCP tool, and through an optional persistent ledger so the history is queryable over time. Pair it with the improvement/* proposal lifecycle (next section): the snapshot tells you what regressed, the loop decides what to do about it.
Every registered biology module reports its TensorLang semantic hash. v1 → v2 → v3_signed → v4_signed now sits beside biology_epi_v1 and biology_epi_v1_signed — a typed roster the snapshot reads from PeptiterMCPCatalog.defaultLeanModules and the matching .reasoner.json manifests.
eval_harness.py writes eval_badge.json on every PR with the delta-vs-previous strings. The snapshot lifts that out of the CI sticky comment so any in-process consumer — operator CLI or agent — can answer 'are we regressing?' without scraping GitHub.
Brier, ECE, AUROC, AUPRC for the platform's currently-committed calibration_report.json. Drift between snapshots flags subsystem-typed regressions: Brier ↑ or ECE ↑ hint at claim_threshold, AUROC ↓ hints at ranking_policy — both map directly onto ImprovementSubsystem cases so improvement/propose accepts them without translation.
Per-run Beta(α, β) convergence already lives inside V3LabLoopOrchestrator.Trace. The snapshot rolls it up across every cached lab_loop/*.json under --cache-dir so an operator can see 'across the last N runs, did max-variance shrink?'
Each event carries a subsystemHint that matches an ImprovementSubsystem case — claim_threshold, ranking_policy, mechanism_reasoning, safety_gate, eval_harness. The snapshot's Markdown renders the hint as a literal improvement/propose call so an operator (or agent) can copy-paste the next step. This is what closes the observation → intervention loop.
The same SelfImprovementSnapshot.read(...) drives all three. The CLI is the operator entry point and the only thing that writes to the ledger; the MCP tool is for agents that want the same shape without spawning a child process; the ledger is what makes the diff-based improvement events possible across invocations.
# Operator CLI — Markdown to stdout, append to a ledger
peptiter-self-improvement \
--repo . \
--cache-dir ~/.peptiter/cache \
--ledger-dir ~/.peptiter/ledger
# Agent — same snapshot via JSON-RPC
{"method": "tools/call",
"params": {"name": "improvement/snapshot",
"arguments": {}}}Live Markdown rollup the CLI prints for the current checkout — six registered biology modules including the epigenome signed twin, the eval bench at F1 0.818, one cached active-learning run, calibration status, and the improvement-event entry the snapshot derived from the prior ledger state.
# Platform self-improvement snapshot Schema: 0.1 ## Biology version coverage (6 modules) - biology_v1 → peptiter.biology_v1 · hash e065d18796a17c58… - biology_v2 → peptiter.biology_v2 · hash 07a0604088ee8b6f… · with attribution - biology_v3_signed → peptiter.biology_v3 · hash 20331e6e81c29318… · with attribution - biology_v4_signed → peptiter.biology_v4 · hash d11eefaddf4b4851… · with attribution - biology_epi_v1 → peptiter.biology_epi_v1 · hash 708129ac96b1cfe7… - biology_epi_v1_signed → peptiter.biology_epi_v1 · hash c445ed3e3e922096… · with attribution ## Regression bench - status: ok · headline model: v2 signed · 4 models evaluated - F1: 0.818 · safety precision: 1.000 · efficacy recall: 1.000 · tests: 25 ## Calibration - assay: receptorFitCheck · model ReceptorFitHeuristic v1.0 · selected curve isotonic - holdout: 36 observations (27 positive, 9 negative) - Brier: 0.0742 · ECE: 0.0472 · AUROC: 0.936 · AUPRC: 0.967 ## Active-learning history (1 cached run) - v3-il23-axis · 3/3 approved · max variance 0.0833 → 0.0833 (→ unchanged) ## Improvement events - Calibration regression on receptorFitCheck: Brier 0.0100 → 0.0742 (+0.0642). → improvement/propose subsystem=claim_threshold - Calibration regression on receptorFitCheck: AUROC 0.999 → 0.936 (-0.063). → improvement/propose subsystem=ranking_policy
The takeaway. The snapshot is the observation half of a coherent self-improvement story. When an entry in improvementEvents warrants a change — a regression bench delta went the wrong way, a lab loop stopped sharpening, a biology version rebuilt itself with a different hash — that's the signal to call improvement/propose and walk the proposal through evaluate → approve → promote (or rollback). The recursive discovery loop below is the intervention half.