[taddyb]

Summary

Applies the loss-landscape learnability instrument to the per-dam reservoir parameters and records the result. Analysis, journal and docs only: no src/ changes.

• experiments/curvature_dams/: per-gauge training-loss landscapes (curvature at the optimum, 0.02 / 0.005 NSE bands, gradient at the engine init vs halfway, multimodality, train-to-test transfer and split-half stability) at 214 on-reach dam gauges and the 69 on-reach flood-control dams, on the laws_v6 offline harness. Also four fixed-table engine replays at pool sizes 0.05 / 1 / 5 / 20 d (binary ddrs-v6-77aa4a0, zero-step, test years) scored against the fair no-pool twin. report.txt has the full method, tables and caveats; five figures.
• research/journal/2026-10.md: entry and ledger rows for the four replays (plus one older unlogged ledger row in 2026-09.md).
• .claude/skills/ddrs-dev/references/research-status.md: what to do with each dam parameter.

Parameter	Verdict	Do
Pool size z	curved at its optimum, flat from the off init (0.05 d) to about 1 d	initialise open (5-20 d) or calibrate per gauged dam
Release target kc	stiff but multimodal (97 % of dams)	calibrate per gauged dam
Capture share phi	boundary optimum at 1; fixing it costs 0.0000	fix at 1
Rule-curve amplitude	curved, but 19 % of the training gain survives on test years	drop or shrink hard
Per-dam T0	flat below 0.5 d; stiff above 2 d, where half start on the shelf	learn without per-dam L2 from a 1-2 d init

Engine vs offline, the pool's own test gain over z = 0.05 d: 1 d +0.0042 / +0.0041, 5 d +0.0145 / +0.0142, 20 d +0.0272 / +0.0267 (per-gauge Spearman 0.92-0.98), so the shelf is in the engine's forward model. For the paper this is an offline pilot of E-17's T census axis and evidence for decision D-5; the engine census has no T axis yet.

The replay manifests stamp 3708ce4+dirty: that is the worktree at launch (another branch's uncommitted edits), not the binary that ran. The two .npz slice arrays (9 MB) are gitignored and regenerated by landscapes.py in about a minute.

Stacked on #52 (base dam-release-v2, where experiments/reservoir/ lives); retarget to master once #52 merges.

Test plan

• engine_summary.txt matches the report; each replay's config.yaml names its z table; all four manifests exist
• No src/, config or test changes
• CI

🤖 Generated with Claude Code

https://claude.ai/code/session_015TepsZAS2LkFC5EmvuZrz9