Make the observer LR gate two-channel (entropy + drift)

[InauguralPhysicist]

The defect

observer_matrix_scale — the production learning-rate gate in the transformer trainer — throttled a weight matrix to 0.5x whenever its entropy channel was quiet. That entropy is computed over a single SGD step's worth of weight change, which sits far below thresholds calibrated for scalar convergence loops, so the "quiet" test was effectively always true.

Instrumented over 2,597 steps x 42 matrices:

verdict	matrix-steps	share
0.5x "stable"	108,514	100.0%
0.3x oscillating	14	0.0%
1.0x full rate	0	never fired

An adaptive gate that never adapts is a constant 0.5x drag — and it was applying that drag while the loss was plainly still falling (6.09 -> 0.54).

Why the observer couldn't see it

The observer buffer here carries only entropy-derived state — entropy, dH, last_entropy, obs_age, prev_dH. There is no value channel at all, so slow-but-steady motion is invisible to it. This is the observer's own failure mode: a meter blind to slow motion certifying a moving system as converged.

The fix

Add the missing channel as relative drift ||grad|| / ||w||, and release the "stable" verdict when the weights are still moving. The 0.3x oscillating and 0.1x diverging verdicts are left untouched — those are genuine instability signals the entropy channel reads correctly.

The two channels disagreed on 100% of matrix-steps, and ground truth says drift was right every time. The gate now releases 99.6% and throttles only real oscillation (0.4%).

Three-way A/B

Fixed checkpoint, 400 windows, identical seeds:

step	entropy-only (status quo)	gate OFF	two-channel
100	1.554582	1.205151	1.237571
200	1.405466	1.012344	0.975602
300	0.728710	0.542757	0.553253

Two-channel recovers ~all of the gate-off gain (~25% better loss than the status quo) while retaining the ability to throttle a matrix that genuinely settles — which "off" permanently gives up.

Honest limit: this A/B cannot separate off from two-channel. During active learning "release everything" and "off" are the same policy, so the distinction is structural and would only show in a long run where something actually settles. What it does establish decisively is that the status quo is the worst of the three.

Controls retained

All three arms stay selectable for future A/B: EIGS_OBS_ENTROPY_ONLY=1 (old single-channel), EIGS_OBS_SCALE_OFF=1 (no gating), EIGS_OBS_DRIFT_EPS (release threshold, default 1e-3).

Testing

Suite 3164/3164 on the http+model build; [47c] gradcheck still pins the batched path to the per-position oracle.

🤖 Generated with Claude Code