[taddyb]

Why

zeta = leakance_factor · area_z · K_D · (depth − d_gw) multiplies leakance_factor and K_D, so only their product reaches the physics: an exact scaling degeneracy. Measured over 346,321 CONUS reaches on a trained arm, rho(leakance_factor, K_D) = +0.9986, and the four learned outputs of that head hold 97 % of their variance in one direction. Physically a reach either sits above a losing aquifer or it does not, so leakance_factor should be a GATE selecting where leakance acts, not a continuous multiplier. A gate and a conductance are not degenerate with each other. On the same arm the model had repurposed d_gw as the switch (learned to +0.43 m against a median channel depth of 0.26 m so the losing-only clamp zeroed most reaches); with a real gate d_gw can go back to being a head difference.

This does not re-open the leakance verdict (research-status.md); it removes one degeneracy so a future arm can be judged on the gate rather than on the product.

What

A temperature-annealed sharpening sigmoid on the head's normalized leakance_factor output u ∈ (0, 1):

g = sigmoid( logit(clamp(u, 1e-6, 1 − 1e-6)) / tau ),   logit(x) = ln(x / (1 − x))

• src/training/gate.rs::leakance_gate — the transform. tau == 1.0 returns u untouched (bit-exact identity, not round-tripped through ln/exp). Generic over the backend, so training (Autodiff<I>), eval, probe and dump all call the same function. Exact autograd gradient through clamp → log → div → sigmoid; no straight-through estimator.
• params.leakance_gate: { temperature: {epoch: tau, ...} } (src/config.rs::LeakanceGate), absent by default. Epoch-keyed schedule resolved exactly like experiment.learning_rate (LeakanceGate::resolve: largest key <= epoch, first value before it). validate_leakance_gate rejects the block without use_leakance: true, an empty schedule, and any temperature that is not positive and finite.
• Applied in all three hand-written head readers, replicated rather than unified (WET): forward takes the epoch's temperature as a new gate_tau: Option<f32> argument threaded from the driver (asserted to be Some exactly when the block is configured); forward_eval_core and probe_forward apply the FINAL scheduled temperature; both dump_parameters copies dump the gated factor at the final temperature. The run log states the temperature in force (epoch E lr=… leakance_gate_tau=… in training; leakance gate: final temperature tau=… from eval, probe and dump).
• src/routing/, TimestepLeakanceOp, and every backward operator are untouched.

Epsilon

GATE_EPS = 1e-6. 1 − 1e-6 sits ~17 ulps below 1.0_f32 (ulp there is 2^-24 ≈ 6e-8), so 1 − u at the clamp keeps ~3 % relative accuracy; 1e-7 would be only 2 ulps away with ~20 % error in 1 − u. logit(1e-6) ≈ −13.8, so the clamp only alters values the head has already saturated past |pre-activation| > 13.8. Without the clamp a head sigmoid that has saturated to exactly 0.0 or 1.0 in f32 gives logit = ±inf and a NaN gradient. Everything stays f32.

Tests

• tests/leakance_gate.rs (new, 11 tests): tau = 1 bit-exact identity on the bare function and end to end through forward, forward_eval, probe_forward against the same config with the block absent; a sharp final temperature demonstrably changes all three readers and they still agree; gradient vs central finite differences at tau ∈ {1, 0.5, 0.2, 0.05, 0.01} (f64 formula at the repo tolerance, plus the f32 implementation at its ulp noise floor); monotone in u, 0.5 a fixed point at every temperature, tau → 0 limits; tau down to f32::MIN_POSITIVE with u exactly 0 and 1 gives finite values and finite gradients; forward rejects a mismatched gate_tau.
• tests/gamma_eval_parity.rs: two new rows (leakance + gate, everything + gate) at final tau 0.05 across all three readers.
• src/config.rs unit tests: block rejected without use_leakance, empty/non-positive temperatures rejected, schedule resolves at epoch boundaries and final_temperature.
• Docs: ddrs-dev/references/config.md §Leakance gate and validator rows; testing.md gate list.

Gates (Tier C; src/routing/ untouched)

cargo build                                   Finished
cargo test --lib                              377 passed; 0 failed; 3 ignored
cargo test --features fixtures --test gamma_eval_parity --test sp8_gradcheck
                                              4 passed / 20 passed
cargo test --test leakance_gradcheck --test leakance_off_parity --test zeta_accum --test leakance_reference_match
                                              16 / 3 / 11 / 5 passed
cargo test --test ddr_sandbox_match           1 passed
cargo test --features fixtures --test leakance_gate   11 passed
mkdir -p output && cargo run --release --example compare_ddr_sandbox
  overall max abs diff: 1.525879e-4 m³/s
  overall max rel diff: 2.412143e-6
  verdict: ABSOLUTE MATCH (max abs < 1e-3 m³/s)
python3 scripts/verify_doc_paths.py           0 unresolved in agent context, 75 in prose docs (pre-existing)

Note: the commit message lists the test counts of leakance_gate and gamma_eval_parity swapped (it is 11 and 4, respectively).

🤖 Generated with Claude Code

https://claude.ai/code/session_0143bx7gVz8ug2oc6DTy5jtz