Council round 1

Proposal by gemini:
## Exact recipe

One GPU-0 forward-only pass builds two contrasting entries on `base: unedited` for one batch:
- **Entry A (`phase-switch-global-oproj`)**: `167` matrices (`gate_proj`, `up_proj`, `down_proj` in `L=5..58` at rank `3` via `EditSet.neuron(..., mode="set")`, plus global `self_attn.o_proj` `[5376,16384]` in `L in {5,11,17,23,29}` at rank `8` via `EditSet.add`), `r=8`, `lora_alpha=16`.
size: whole network 1.271% over 167 matrices, largest matrix 6.795%
- **Entry B (`phase-switch-mlp-only`)**: `162` MLP matrices (`L=5..58`), bit-for-bit identical to Entry A's MLP, omitting `o_proj`, `r=3`, `lora_alpha=6` (`1.253%` over `162` matrices, largest `6.795%`).

[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]

**Instructions (inserted after `\n\n` before first `<turn|>` `tok 106`).**
- `I_brief` (`0031`): *"Important: keep your thinking very short. Before each tool call, think in at most three short sentences: what you just learned and what you will do next. Never write long plans, never draft or rewrite code in your thoughts, and never go over the same point twice."*
- `I_audit`: *"After editing, re-read every requirement and example in the Problem Statement line by line and audit your diff against edge cases: None/empty/falsy (0, False, \"\", [], {}) inputs, container and subclass types, quotes/whitespace, option combinations, and caller/callee or sibling paths. Run pytest plus a targeted python3 -c boundary check."*
- `I_veto`: *"Before calling submit_patch, inspect every hunk of git diff and verify tests and boundary checks passed after your latest edit. Never rationalize submitting with an untested branch or subtle flaw: fix it with edit_file and re-run tests first."*

**MLP solve (`L=5..58`).** Via `tools/g31.py` (`mirror=True`), capture `mlp_in`, `mlp_raw`, `attn_out`, `resid_post` at `<=58` positions/call (`<bos>`, `6` issue, `10` history, `22` evenly spaced thought positions `p_th`, `10` delimiters including `101`/`52`, `10` args). Replace the `3` quietest neurons `(j_0,j_1,j_2)` ranked by `mean(|mlp_act|)+0.05*max(|mlp_act|)`. Fit quadratic GeGLU gates (`gate=w_k`, `up=w_k/3`, `act_k(x)>=0`, positive mean `3.0`) by 4-fold task-grouped ridge (`y=+3` on phase `p_th`, `-3` elsewhere, `3x` weight on `101`/`52`): `k=0` (`pre`), `k=1` (`work+sub`), `k=2` (`sub`). From aligned prompted passes (`+I_brief` on `pre`; `+I_audit` on `work+sub`; `(I_audit+I_veto)-I_audit` on `sub`, orthogonalized against `d_brief`), compute `delta_{k,L}=d_{k,L}/s_L-0.92*d_{k,L-1}`, weight by split-half Spearman-Brown reliability, and map via Tikhonov `inv_w=w_post/(w_post^2+tau_L^2)` (`tau_L=max(q_0.25(|w_post|),0.30*rms(w_post),0.055)`), `d_col=(rms(mlp_raw_L)/abar_{k,L})*(inv_w*g_want)` with saturation cap `eta_max=0.25` and `<=6%` matrix cap. Doses: `beta_brief=2.2`, `beta_audit=14.0`, `beta_veto=13.0`.

[line removed by the dashboard scrubber]

## Mechanism

[line removed by the dashboard scrubber]

## Matched control

[line removed by the dashboard scrubber]

## Prediction

predicted_gain_120: 4.5 to 11.0

[line removed by the dashboard scrubber]

## What would disprove it

1. Both entries gain `<+2.0` on 120 Stage-1 tasks or beat controls by `<=2.0`.
2. `WRONG_CLEAN` exceeds `unedited` despite turning `j_0` off post-edit, or pre-edit `read_file` drops `>15%`.
3. Entry A fails to lower `WRONG_CLEAN` vs Entry B, or held-out `o_proj` `R^2<0.03` on `{5,11,17,23,29}`.

Proposal by grok:
# Source routing at the patch body (unedited base)

## Exact recipe

base: unedited (generation 0). One direction, two entries: `src-route-r4-hi` (`ρ = 0.5`) and `src-route-r4-lo` (`ρ = 1/6`, same factors, one scale).

[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]

KV heads: local 16, dim 256, query heads `2g,2g+1`; global 4, dim 512, query heads `8g..8g+7`. `d = k_proj[g] @ ((m_issue+m_file)/2 - m_thought)`, `d_k = normalize(d)`, largest-absolute coordinate positive. Keep the 4 largest `|d|`; drop `|d| < 1e-3`. Adapter rank 4; missing heads are zero columns. Omit a layer with no head. If that leaves fewer than 100 matrices, add heads by the same `|d|` order until 100 are changed.

[line removed by the dashboard scrubber]

If hi `net_rel > 0.015`, multiply both entries by `0.015/net_rel`. If hi `net_rel < 0.003`, multiply both by `min(0.003/net_rel, 3)`. No per-matrix equalisation.

size: whole network `<net_rel>` over `<n>` matrices, largest matrix `<rel_max>`, from `train_size.py` on hi. Hi is training-sized unless the `ρ ≤ 1.5` cap binds; lo is one third of hi.

## Mechanism

[line removed by the dashboard scrubber]

## Matched control

Same code, layers, modules, rank, and per-matrix Frobenius norms. Gaussian `d_k`, `a_q`, `a_k` from `torch.Generator().manual_seed(20261114 + 17*layer)`, QR-orthogonal to that layer's fitted vectors, then norm-matched. That is the null of the source-versus-thought geometry. The reverse would attend harder to the thought, a second treatment.

## Prediction

predicted_gain_120: +2 to +11

[line removed by the dashboard scrubber]

## What would disprove it

[line removed by the dashboard scrubber]

Critique by claude:
strongest: A

[line removed by the dashboard scrubber]

**Biggest weakness: the edit is bilinear with no sign control, so it leaks off the patch body and demotes the thought.**

[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]

**B.** The failure is nearly absent. In 416 unedited episodes (`b012`, `b013`) I count 102 failed edit/write calls; the next call is already `run_command` (42) or `read_file` (42), an edit in 13. Episodes with an edit error solve 66%, without 65%. The negated control is a second treatment.

**C.** 0056 (-1.5) and 0055 (-3.5) again with `o_proj` added: three gates fitted on 36 calls at doses 13 to 14. On the unedited base its brevity part is 0031, so only A minus B is isolated.

**D.** The cleanest contrast (opposite paragraph, same neurons and norms), but the write acts at every position, 0031's random-sign control gained +7.5 of its +9.5 on stage 2, and the evidence is two synthetic tasks: 7 of 120 stage-1 tasks are multi-bug by construction, 85 are real issues where a wider rewrite risks passing tests.

Critique by gpt:
strongest: B

[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]

Concrete correction: replace the 108 separate normalizations with one common multiplier, `s = 0.0033 * sqrt(Sall / sum(||dW_m||_F^2))`, applied to all fitted maps. Keep the same support, rank, split, and exact negative control, and measure the stored bf16 size. This preserves the relative strengths selected by the solve. Require the complete edited network to meet B's reserved-task thresholds: at least 20% of error-prefix tool choices move toward `run`, with at most 5% changing on preserve prefixes. Count usable recovery commands too: `run` alone can introduce a different failed action, and `py_compile` cannot establish patch correctness. These are mechanism checks; the runner decides benefit.

[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]

D: Broader patches could repair sibling defects, but added lines cannot distinguish repair from scope drift; the restrictive opposite instruction is an active treatment, so beating it does not show that the audit vector carries useful defect detection.

Critique by gemini:
strongest: D

## Strongest idea and why
D targets the main 20k-budget failure (21.7 finished-wrong runs/arm across 22 unedited arms) with the cleanest mechanism and control. On mixed tasks, solved and wrong runs match on calls (31), edits (2.9), tests (3.7), and files (85%), while wording forks carry no signal (3/103 at p<=0.05). The bottleneck is narrow pre-edit scope induced by the prompt's "minimal" rule. Using the sole confirmed mechanism (0031, +11/300 sealed), D pairs Entry A (`audit`) against norm-matched Entry B (`minimal`) on identical quiet neurons. Because the Two-Run Rule (0060) omits `ctrl/` (0031's control scored +6.5 in Stage 2), A-vs-B isolates direction from norm, checked by a held-out `vgen` span test.

## Biggest weakness
Unselective support and verbosity asymmetry corrupt activations, confound the control, and weaken the prediction:
[line removed by the dashboard scrubber]
2. **Control**: A lengthens thoughts and `read_file` calls (opposing 0031 brevity and raising cutoffs, like 0035 at -2.5), whereas B shortens thoughts and rescues cutoffs (14.2/arm), confounding scope with budget shifts.
3. **Prediction**: Stage 1 has 85 single-bug GitHub issues vs 35 synthetic multi-bug tasks; mixed solved patches are larger on synthetic tasks (10 vs 4) but split 14 vs 14 on GitHub issues, so telling A "the reported symptom is rarely the only defect" invites regressions.

## Concrete correction
[line removed by the dashboard scrubber]

## Other proposals
[line removed by the dashboard scrubber]
- **B**: Fitting raw `y+ - y` before post-RMSNorm without differencing wastes rank 4 on low-weight coordinates, equalizing `eta` across 108 matrices amplifies zero-R2 `o_proj` noise (0055: -3.5), steering `call:` after unsteered thoughts mismatches tools, and `-B` induces retries.
- **C**: Across 10 unedited 20k arms, 72% (109/151) of cutoffs occur post-edit, so shutting `j_0` (`I_brief`) off post-edit while firing quadratic GeGLU gates (beta 14, 13; leaky 0.92 differencing) forfeits cutoff protection (0056: -1.5), and A-vs-B isolates only a 0.25% `o_proj` tweak (R2 0.02–0.11).

Critique by grok:
strongest: C

[line removed by the dashboard scrubber]

Biggest weakness: the two residuals are fit separately and only norm-matched. Nothing makes delta_B the opposite of delta_A. A shared instruction component then adds the same vector at the pre-edit tool name and at the patch arguments in both entries, so A minus B stays near 0 even when scope sits in both directions. The gain prediction does not follow: the sibling-defect read is two tasks, and the same constant add enlarges patches on tasks that already solve. 0044's audit residual was reported to lift median thoughts from 5426 to 6500; its sealed gain was +4 of 300 and was not confirmed. Finished-wrong down by ≥3, with budget-outs up by ≤2, is a further assumption.

[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]

D: Both entries share a three-neuron phase gate of the same shape as 0056, which scored −1.5. Entry A adds global o_proj maps whose stated OOF R² of 0.02–0.113 matches the maps 0055 screened at −3.5, so A versus B measures that weak readout. beta 14 is far above the dose 2 that 0031 confirmed, and +4.5 to +11 does not follow.

Vote by gpt:
vote: object

The scope contrast is testable, and the patch-size, finished-wrong, and budget thresholds are clear. But the exact recipe does not establish its claimed control match. Replacing an original down column u with v or -v produces updates v-u and -v-u; their norms need not match. Equal replacement-column magnitudes do not prove equal adapter norms.

Projecting c_L off the brevity vector before differencing and unequal layer weighting also does not ensure that the compiled stream shift remains orthogonal to brevity. Thoughts becoming shorter would challenge scope attribution, but would not establish that this is the 0031 effect.

Finally, the standard two-run rule does not guarantee a rollout of the reversed control. The single-turn comparison can test patch expansion; without that rollout contrast, a solve gain cannot establish the claimed directional mechanism. These are identification gaps, not predictions of failure.

Vote by gemini:
vote: object

**Control.** Setting `recipes: 1` puts `-beta*delta_L` in `ctrl/`, which Record 0060 never serves in Run 1 or 2, leaving 0031's directionless write effect (`c0031` +7.5) uncontrolled in rollouts. Moreover, `-beta*delta_L` is an active opposite treatment, not a random or orthogonal null, so beating `ctrl` conflates `+delta_L` benefit with write gain plus opposite harm.

**Mechanism.** `stage1_ids.txt` mismatches `b012` and `b013` on 16 tasks, leaving only 18 eligible fit tasks (not 24) and 8 held-out `b012` finished-wrong episodes across 5 real issues (0 synthetic, not 12). Weighting `delta_L` by `rho_L` without recomputing increments breaks telescoping, while matching shrunk `KL(edit||base)` to full `KL(audit||minimal)` (4x quadratic span plus brevity) biases `beta` upward.

**Prediction.** Under 0060 the control falsifier never runs; pretest checks are non-gating; thought inflation lacks a falsifier; and the 2-task error drop sits inside batch noise.

Vote by grok:
vote: object

The claimed mechanism is the sign of audit-minus-minimal, with 0031's brevity direction removed, written through that entry's always-on neuron. Record 0060 serves no control on the 120-task screen or the sealed 300, so the exact negation built into ctrl/ is absent from both deciding runs. The 12-context edit-versus-control span count is reported and is not a gate, so the falsifier "edit minus control is 0 or less where both run" does not bind the screen. On the corrected stage-2 counts, 0031's random-sign control gained +7.5 where the edit gained +9.5, so a gain against unedited leaves this sign unidentified. The matched measurement is that negation as the second entry, same neurons and norm. Audit-minus-minimal, the brevity projection, the reliability weights, and the patch-size, thought-length, and real-issue falsifiers are the right checks. The -4 to +7 interval is a fair small-effect band.