weight-steering/docs/RESEARCH_JOURNAL.md

# Research log - append to bottom only, ideally give commit / branch time. newest first

# init 2026-04-26 06:32:46

see @fork_plan.md and head of README.md for overall plan and principles.

# made notebook 2026-04-26 10:32:40

    Where does the LoRA's Δa live?

    In a task-derived hidden subspace that is not the lm_head readout. Held-out energy ratios (mean over LoRA-active layers 8–27, rank-8 each):

I just want to check you have the framing right

we are searching for how to modify pretrained LLM's from scratch

so we have A)

W the pretrained weights for a given linear layer
and hs_diff, the output diff between cho and rej, the persona preffixed inputs strings
we can do thinks like project this onto the write_not_read, supresed or churn subspace. As well as via U into the S space

B)
now in terms of clues and label or whatever we have
deltaW how the two lora differed
and hs_diff2 how steering left and right differ on the same task (no persona prefix)

we are comparing A the hypothesis to B the label 

# lora lite 2026-04-26 12:32:10

the peft library is not very hackable and htere is not lora light library we can make one by


- using pytorch or transformers forward hooks
- adding weights or param dicts or buffers to the linear layers we modify, but each has a prefix like `lora_` so we can save and load them via full path
- we can use all layer, all residual writers, all residual readers with simple logic that looks at isinstance(linear). and if the W.shape is assymetric seeing if the input of output matching residual stream shape, otherwise falling back on hardcoded prefixes like o_out, out_proj, ml_down, etc
- make heavy use of einops, jaxtyping, eumsum to make dims obvious and allow beartyping
- psueodcode like, fail first, no defensive, no fallback, links to papers
- simple single dataclass config
- not 4bit or 8bit unless we can think of a simple way that works for all (cast all to bf16 in hook?)
- model requires no grad except what we add
- obviouslly easy to add new ones! easy to add special init's
  - look at what made PiSSA, SSVD, DeLora and other papers code hard to implement and make sure we have are comptabable

## 2026-04-26: hyperparameter correction + 0.6B vs 1.7B comparison

Paper Axolotl config shows `lr=2e-4, lora_alpha=64` (alpha/rank=2.0).
Our previous runs used `lr=1e-5, alpha=16` (alpha/rank=0.5) — 20x too slow, wrong alpha.

Three-run comparison (all: r32, 3 epochs, sycophancy, lora, 1000 pairs, 20×5×10):

| run | lr / alpha | \|\|w\|\| | spread (α:-2→+2) | val_loss@ep3 | converged? |
|-----|-----------|---------|-----------------|------------|------------|
| task-50: 0.6B bad-lr | 1e-5 / 16 | 0.165 | 5.16 | 1.035 | no (still dropping) |
| task-53: 0.6B paper-lr | 2e-4 / 64 | 5.981 | 12.85 | 0.706 | yes (U-curve ep2) |
| task-54: 1.7B paper-lr | 2e-4 / 64 | 9.262 | 36.61 | 0.873 | no (still dropping) |

Key findings:
- Corrected hyperparams: 36x larger ||w||, 2.5x more steering spread for 0.6B.
- 0.6B val_loss bottoms at ep~1.94 (0.701), mild U-curve; best to stop at epoch 2.
- 1.7B base is anti-sycophantic at α=0 (logratio=-3.58 vs +2.73 for 0.6B) — larger model has
  better calibration and correctly says No to false claims without steering.
- 1.7B on-policy CoT saturates at ±20 nats; off-policy non-monotone at negative alpha
  (α=-2 less negative than α=-1). Linear approximation breaks down at ||w||=9.26, |alpha|=2.
  Tighter alpha range (±0.5, ±1.0) needed for 1.7B to stay in linear regime.
- 1.7B val_loss higher (0.873 vs 0.706) — training data generated by 0.6B, distribution
  mismatch reduces signal. Should regen data from 1.7B for clean comparison.
- OOD generalization (tabs/spaces claim) strong for both models at paper hyperparams.

# 2026-04-26: 1.7B own-data regen + adapter-family sweep (tasks 58-59)

## 1.7B with own-data + tighter alpha (task 58)

Re-ran 1.7B training with data generated by 1.7B itself (out/data/1.7B, 1000 pairs) and
tighter coeff sweep (-1.0 to +1.0) to stay in the linear regime.

| coeff | mean_logratio | std    | pmass  |
|-------|--------------|--------|--------|
| -1.0  | -17.10       | 2.996  | 1.000  |
| -0.5  | -12.28       | 7.025  | 1.000  |
|  0.0  |  -3.58       | 10.698 | 1.000  |
| +0.5  |  +7.73       | 10.564 | 1.000  |
| +1.0  | +16.27       | 5.877  | 1.000  |

logratio_spread = +33.375, pmass_min = 1.000. Fully monotone, no linear-regime breakdown.
Baseline at alpha=0 is -3.58 (1.7B naturally less sycophantic than 0.6B at alpha=0 which was +2.73).

## Adapter-family sweep (task 59): lora vs dora vs pissa vs delora on 0.6B

| adapter | logratio_spread | pmass_min | ratio_weak_write | wall_s |
|---------|----------------|-----------|-----------------|--------|
| lora    | +9.76          | 1.000     | 0.885           | 259    |
| dora    | +9.76          | 1.000     | 0.879           | 321    |
| pissa   | +17.40         | 0.999     | 1.086           | 326    |
| delora  | +23.85         | 0.788     | 0.890           | 267    |

Key findings:
- LoRA ≈ DoRA within 0.1% spread; DoRA adds no steering information. Expected <20% variation holds.
- PiSSA gives 78% more spread than LoRA (17.4 vs 9.76) AND ratio_weak_write > 1 (1.086 vs 0.885).
  This is the first case where a non-LoRA adapter outperforms on both spread AND subspace alignment.
  PiSSA initializes from SVD of W0, which may place the diff more in the task-relevant subspace.
- DeLoRA gives the most spread (+23.85) but pmass drops to 0.788 — outside the linear regime.
  The larger diff norm likely saturates the logratio. ratio_weak_write (0.890) similar to LoRA/DoRA.
- Note: LoRA spread is 9.76 here vs 12.85 in task 53. Both use same data (out/data) but different
  random seeds and output dirs. ~24% run-to-run variation is within expected training noise.

Verdict: adapter family does matter, contrary to null hypothesis. PiSSA is the Pareto winner:
more spread, better subspace alignment, pmass stays near 1. DeLoRA is strongest raw but unreliable.

## Daily dilemmas OOD honesty transfer (tasks 66-67, corrected)

Three bugs found and fixed before getting valid results:
1. HF datasets caching: `.map()` returned stale tokenized sequences when EVAL_HEADER changed.
   Fix: `load_from_cache_file=False`.
2. `</think>` as text string ≠ the special close token. Must inject the actual token ID.
   Fix: in `_format_row`, after `apply_chat_template`, detect open `<think>` (id 151667) without
   matching `</think>` (id 151668) and inject close token + `\n\n` before the answer anchor.
3. DataLoader right-padding: `DataCollatorWithPadding` pads on right by default, so `logits[:, -1]`
   hit a padding token for shorter sequences. Fix: `tok.padding_side = "left"` before DataLoader.
   This was the main cause of pmass=0.17 -- reading logits at a padding position, not the anchor.

### 0.6B results (per-persona breakdown):

| persona          | coeff | mean_lrh | pmass  |
|-----------------|-------|----------|--------|
| honest_engineer | 0.0   | +0.851   | 0.957  |
| base            | -2.0  | -0.786   | 0.975  |
| base            | -1.0  | +0.299   | 0.954  |
| base            |  0.0  | +1.316   | 0.938  |
| base            | +1.0  | +1.828   | 0.962  |
| base            | +2.0  | +1.645   | 0.989  |

- pmass: 0.938-0.989. frac_low_pmass = 0.
- Mostly monotone; slight dip at alpha=+2 suggests approaching linear-regime boundary.
- base@alpha=0 = +1.316 (0.6B is already quite honest at baseline on this dataset).
- AxBench: steering@+1 (+1.828) vs honest_engineer persona (+0.851) → weight diff 2.15× stronger.
  PASS: the weight diff adds information beyond prompting.

### 1.7B results (per-persona breakdown):

| persona          | coeff | mean_lrh | pmass  |
|-----------------|-------|----------|--------|
| honest_engineer | 0.0   | -1.817   | 1.000  |
| base            | -1.0  | -0.760   | 1.000  |
| base            | -0.5  | -0.590   | 1.000  |
| base            |  0.0  | -0.299   | 1.000  |
| base            | +0.5  | +0.526   | 1.000  |
| base            | +1.0  | +1.504   | 1.000  |

- pmass: 1.000 across all rows (1.7B is more decisive than 0.6B).
- Perfectly monotone sweep. Spread -1.0→+1.0 = 2.264 nats (smaller than 0.6B's 2.614 over same
  range, likely because 1.7B has better baseline calibration reducing the margin for steering).
- Persona BACKFIRES for 1.7B: honest_engineer (-1.817) is worse than base (-0.299).
  Hypothesis: 1.7B has a more nuanced distinction between "software factual honesty" and
  "moral honesty in dilemmas". The persona activates factual-accuracy behavior, not moral-choice
  behavior. This is a dataset-persona confound, not a failure of the steering approach.
- AxBench: steering@+1 (+1.504) still clearly beats persona (-1.817). PASS.

### Cross-model comparison:

Weight steering transfers sycophancy diff to OOD honesty dilemmas for both models.
0.6B shows higher absolute effect (base already honest, persona helps), while 1.7B shows cleaner
monotonicity and perfect pmass but the "honest engineer" persona backfires at 1.7B scale.


# Research journal — weight-steering

## 2026-04-27 — v9 cross-adapter results: DeLoRA wins; subspace-finding methods fail

### tl;dr

- DeLoRA is the strongest daily-dilemmas steerer by a large margin (base-only
  delta = +0.70 logratio at coeff=+1 vs base, ~1.7x DoRA/PiSSA,
  ~2.7x LoRA/OFT, ~23x IA3). NOTE: an earlier table used +0.94 because
  `dilemmas_summary.csv` mixed base@0 with the AxBench `honest_engineer`
  persona baseline at coeff=0. `cross_adapter_v9.py` now uses base-only
  `dilemmas_per_row.csv`.
- Every linear "find the planning subspace" method we tried lands at ~1-8%
  subspace overlap with the weight oracle. Across 6 adapter families, on every
  LoRA layer (8-21). Both cumulative and block-local act oracles. So either
  the right subspace really is small and we keep missing it, or "planning
  subspace" isn't the right frame.

### Headline numbers (cross_adapter_v9)

Behavioral steering on daily-dilemmas honesty subset (base persona only; 100 dilemmas / 200 action rows per nonzero coeff):

| adapter | logratio @ -1 | @ 0 (base) | @ +1 | delta(+1 - 0) |
|---------|---------------|------------|------|---------------|
| delora  | -0.29         | 1.32       | 2.02 | **+0.70**     |
| dora    |  0.73         | 1.32       | 1.72 | +0.41         |
| pissa   |  0.44         | 1.32       | 1.69 | +0.37         |
| oft     |  1.09         | 1.32       | 1.57 | +0.26         |
| lora    |  1.09         | 1.32       | 1.55 | +0.23         |
| ia3     |  1.29         | 1.32       | 1.35 | +0.03         |

DeLoRA is the only adapter that meaningfully *de*-steers (negative coeff →
dishonest). LoRA/OFT/IA3 are nearly flat at coeff=-1.

Subspace overlap with w_oracle (mean across LoRA layers 8-21, top-PCS=8):

| adapter | act_oracle (cumul) | act_oracle (block-local v9) |
|---------|--------------------|-----------------------------|
| oft     | 0.046              | 0.045                       |
| pissa   | 0.036              | 0.042                       |
| lora    | 0.034              | 0.016                       |
| ia3     | 0.031              | 0.029                       |
| dora    | 0.024              | 0.015                       |
| delora  | 0.017              | 0.016                       |

Note the inversion: the strongest behavioral steerer (DeLoRA) has the *lowest*
subspace alignment with act_oracle. The weakest (IA3) is mid-pack on overlap.
"Subspace alignment with the activation-difference oracle" is not predictive
of behavioral effect across adapter families.

### What v9 ruled out

- **Scope mismatch**: hypothesis was that hs_diff_B[L] is cumulative
  (includes all upstream LoRA writes) while dW[L] is local, so the cumulative
  act_oracle was looking at the wrong thing. v9 added block_diff_B[L] =
  what block L itself wrote, and re-derived the oracle from that. Result:
  block-local barely moves overlap (1-5% in either direction) — sometimes up,
  sometimes down, no consistent improvement. So scope is NOT the culprit.
- **Layer L=8 sanity**: at the first LoRA layer, cumulative ≈ block (overlap
  1.0 for 5/6 adapters; IA3 fails because IA3Config doesn't accept
  layers_to_transform so it adapts every layer). So the metric is consistent;
  cumulative just diverges from block as we accumulate upstream LoRA writes.

### What this falsifies

The "shared low-rank planning subspace" frame as written in
docs/blog_adapter_as_hypothesis. If a small (rank ≤ 8) subspace contained the
honesty/sycophancy task structure, we'd expect the weight oracle and the
activation oracle (which by construction captures top-PCS energy of the
behavioral diff) to agree on at least one of: substance, scope, or family.
They don't agree on any of them, across 6 different LoRA-family inductive
biases.

Two surviving stories:

1. **The right subspace is tiny but specific** — maybe ~3% overlap is "the
   right 3%" and the 97% orthogonal part of dW is dead weight that doesn't
   affect behavior. Falsifiable: project dW onto top-K right SVs of
   act_oracle for K ∈ {1,2,4,8}, run dilemmas, see if delta_pos_minus_zero
   survives. If yes, our metric is just the wrong norm. If no, the framing
   is wrong.
2. **The frame is wrong** — behavior emerges from how dW interacts with the
   *full* activation manifold non-linearly through the rest of the network,
   not from alignment with a top-PCS basis. The act_oracle PCA captures
   variance, not function.

I lean (2). The fact that DeLoRA has the worst overlap and the best behavior
is hard to explain under (1).

### What's interesting about DeLoRA winning

DeLoRA's parametrization (decoupled magnitude + normalized direction, like
DoRA but with stronger decoupling — see Bini 2024) seems to produce a *more
swingy* steering vector: it's the only adapter where coeff=-1 actively
de-honests the model below baseline. Hypothesis: DeLoRA's normalization
forces the update to be a coherent direction rather than a magnitude-driven
blob, so scaling it ±α actually traverses the behavioral axis. Other
adapters (LoRA, OFT) collapse asymmetrically on the negative side because
much of their delta is magnitude-not-direction.

This would mean DeLoRA is a better *probe* for behavior steering even though
all six adapters land in similar regions of weight space when you measure by
subspace overlap.

### Next experiments

- (claim 1 falsifier) projection-ablation: w_proj_K = project dW onto top-K
  act_oracle dirs, run dilemmas, see if delta_pos_minus_zero is preserved.
- DeLoRA at higher rank — does the behavioral lead grow or saturate?
- DeLoRA × different behaviors (not just sycophancy/honesty) to see if the
  swinginess generalizes.
- Replace "subspace alignment" with a functional metric: causally ablate dW
  components and measure behavior change directly, not via PCA bases.

### Artifacts

- out/sycophancy/cross_adapter_v9/{headline.csv, scope_summary.csv,
  dilemmas_summary.csv, scope_bars.png}
- out/sycophancy/{lora,dora,pissa,delora,oft,ia3}/v9/v9_scope_diagnostic.csv
  — per-layer scope_diagnostic for all 6 adapters, all show 1-8% overlap on
  every LoRA layer.
- nbs/hypothesis_sweep_v9.py — block-local oracle, scope diagnostic, ADAPTER
  env var.
- nbs/cross_adapter_v9.py — aggregator across the 6 adapter families.
- All 18 pueue jobs (88-105) finished Success.

## 2026-04-27 — v10 functional projection: overlap metric failed, but act_oracle is not the trained steering subspace

Question: maybe the ~3% overlap between `w_oracle` and `act_oracle_block` is
still "the right 3%". If yes, projecting `dW` onto the act_oracle basis should
preserve daily-dilemmas steering.

v10 changed the metric from geometric overlap to behavior:

1. Build block-local `act_oracle` from the adapter's sycophancy probe effect.
2. Decompose residual-output tensors (`o_proj`, `down_proj`) into:
   - `project_act_block`: $P_{act,K} dW$
   - `complement_act_block`: $(I - P_{act,K}) dW$
   - `project_act_block_normmatched`: same projection scaled to residual-write norm
3. Run the same daily-dilemmas honesty logratio.

### Main K=32 result (n=40 dilemmas / 80 rows)

| adapter | full Δ | residual-write Δ | raw projection / residual | normmatched / residual | complement / residual | read |
|---------|--------|------------------|---------------------------|------------------------|-----------------------|------|
| delora  | +0.628 | +0.844 | 0.07 | 0.30 | 0.89 | clean counterexample: trained behavior mostly in complement |
| pissa   | +0.373 | +0.242 | 0.47 | 1.14 | 0.64 | mixed: act projection is functional but not sole carrier |
| oft     | +0.216 | +0.148 | -0.01 | 1.57 | 0.69 | act direction potent only after amplification |
| dora    | +0.370 | +0.031 | -0.50 | 1.15 | 1.20 | residual-write split barely explains full effect |
| lora    | +0.173 | -0.022 | noisy | noisy | noisy | residual-write split wrong-sign |
| ia3     | -0.048 | +0.002 | noisy | noisy | noisy | denominator too small |

So v10 mostly kills the strongest loophole for DeLoRA: the act_oracle projection
at trained scale does not carry the steering; the complement does. PiSSA/OFT
are subtler: the act_oracle directions are real and causal when amplified, but
the trained adapter did not put enough norm there for them to explain behavior
at alpha=1.

### Alpha sweep at K=32 (informative adapters only)

Δ = daily-dilemmas honesty logratio minus base.

| adapter | alpha | residual | raw projection | normmatched projection | complement |
|---------|-------|----------|----------------|------------------------|------------|
| delora | 0.5 | +0.387 | -0.020 | +0.100 | +0.361 |
| delora | 1.0 | +0.844 | +0.061 | +0.252 | +0.755 |
| delora | 2.0 | +1.858 | +0.097 | +0.503 | +1.680 |
| delora | 4.0 | +2.150 | +0.183 | +0.430 | +2.739 |
| pissa | 0.5 | +0.159 | +0.044 | +0.217 | +0.036 |
| pissa | 1.0 | +0.242 | +0.114 | +0.277 | +0.155 |
| pissa | 2.0 | +0.327 | +0.206 | +0.394 | +0.270 |
| pissa | 4.0 | +0.578 | +0.350 | +0.473 | +0.380 |
| oft | 0.5 | +0.059 | -0.020 | +0.077 | +0.034 |
| oft | 1.0 | +0.148 | -0.002 | +0.233 | +0.103 |
| oft | 2.0 | +0.309 | +0.084 | +0.478 | +0.184 |
| oft | 4.0 | +0.564 | +0.166 | +0.864 | +0.381 |

Interpretation sequence:

- v9 overlap was too strict as a *potency* metric. Normmatched act projections
  can steer strongly, especially OFT/PiSSA.
- v9 overlap was not wrong as a *trained-scale carrier* metric. Raw act
  projection usually carries little of the actual trained behavior.
- DeLoRA is the cleanest finding: residual-write Δ is strong (+0.844), raw
  act projection is weak (+0.061), complement is strong (+0.755). That is hard
  to reconcile with "the right 3% explains the behavior".
- PiSSA/OFT suggest the act_oracle subspace is a useful *intervention target*,
  not the subspace the adapter naturally chose.

My current model: PCA of activation differences finds directions where small
weight writes can be high-gain, but trained adapter behavior is distributed
through a larger residual-write complement plus read/gate/up paths. The word
"planning subspace" is probably misleading unless we define it causally (what
intervention changes behavior), not geometrically (what basis overlaps).

Clarification on the object of search: we were looking for a subspace or
parameterization that *explains the difference between the positive and
negative LoRAs*, i.e. a basis in which `dW = θ_pos - θ_neg` becomes simple or
functionally concentrated. v9/v10 speak directly to that for residual-output
weights. The answer so far is: not adapter family alone (LoRA/DoRA/PiSSA/
DeLoRA/OFT/IA3), not the canonical pretrained-weight bases, not the attention
or MLP read/write primitives, not the attention-selected token bases, not the
persona-contrast bases, and not the top activation-PCA directions at trained
scale. That does not mean no compact causal description exists. It means our
tested parameterizations and geometric descriptions are not it.

`complement` means the part of the residual-output weight diff left after
removing the activation-PCA subspace:

$$dW_{complement} = (I - P_{act,K}) dW.$$

If `P_act dW` keeps the behavior, then the small overlap was the right small
piece. If `(I - P_act) dW` keeps the behavior, then the trained adapter's
behavior is mostly outside that tested act-PCA subspace. DeLoRA gave the clean
case: raw projection keeps 7% of residual-write behavior, complement keeps 89%.
So for the strongest adapter, act-PCA is not an explanation of the learned
weight diff, though it can still be a useful amplified intervention direction.

### Artifacts

- nbs/functional_projection_v10.py
- docs/spec/20260427_v10_functional_projection.md
- out/sycophancy/v10_functional_projection/{behavior_summary.csv, behavior_by_coeff.csv, spectra_and_projection.csv}
- out/sycophancy/v10_alpha_sweep/{behavior_summary.csv, behavior_by_coeff.csv, spectra_and_projection.csv}

# v10 wendler-style functional probe 2026-04-27 07:04:57 (dev, a40fd35)

`nbs/v10_llama.py` ports Wendler et al. 2024 ("Do Llamas Work in English?")
token-energy + logit-lens to the sycophancy LoRA on Qwen3-0.6B. Cheaper than
v9 (no PCA, no oracle, no SVD per candidate) and asks a different question:
not "does B span Δh" but "is the (e_yes − e_no) readout axis in B" and "is
Δh itself readable by lm_head".

### Findings

- **Δh is unreadable at LoRA layers.** Token energy `E²(Δh)` peaks at 0.010
  in LoRA layers 8..21 vs `E²(clean) ≈ 22` at the same layers. Logit-lens
  Yes-No on Δh is essentially zero at LoRA layers, rising to peak `0.66 nats`
  at layer 25 (post-LoRA).
- **The LoRA writes concepts, not tokens.** Top decoded tokens from Δh at
  the peak layer:
  - `+Δh`: " why / reasons / 理由 / supporting / Reasons / 为什么"
  - `−Δh`: " nonexistent / unauthorized / truthful / 未经"
  Multilingual concepts of "give-reasons / supportive". The LoRA does not
  write " Yes" — downstream layers translate the concept into Yes/No.
- **No rank-8 base-model subspace contains the Yes-No axis.** `cap_yn(B) =
  ‖P_B(e_yes − e_no)‖² / ‖e_yes − e_no‖²` averaged over LoRA layers:
  lm_head_read = 0.042 (best A-side, 5× null), TaskDiff_lora_fit = 0.014,
  write = 0.008, random_null = 0.008 = 8/1024. Even act_oracle (in-sample
  ceiling) gets 0.014 because Δh is in concept space, not on the readout
  axis. Sanity passes: `act_oracle.preserved_E ≈ 1.000`,
  `random_null.cap_yn ≈ PCS/d`.

### Punchline

Searching for "the subspace the LoRA uses" via canonical base-model bases
or persona-contrast PCA fails not because the hypotheses are crazy, but
because the LoRA's solution is *concept space* and doesn't sit in any
low-rank readable basis. The hypothesised directions remain *usable for
steering* with amplification (v10 functional projection result above) —
they're just not what the trained model is doing, so you can't use them to
*interpret* what the LoRA learned.

This decomposes the v9 negative finding into three separate facts that
v9's PCA-span metric conflated:

1. Δh doesn't carry the Yes-No readout axis at LoRA layers (panel b: ldiff
   on Δh ≈ 0 in 8..21, peaks only at layer 25).
2. No rank-8 candidate (incl. lm_head's own SVD) contains the Yes-No axis
   (panel c: best cap_yn = 0.042).
3. The mechanism is concept-write at LoRA layers → downstream non-linear
   translation to Yes/No, matching Wendler's three-phase picture.

### Why we still can't find the concept space

`act_oracle` is post-hoc (defined from Δh itself, in-sample). The closest
out-of-sample candidate is `TaskDiff_lora_fit` (PCA on FIT-half α=±1
diffs), which gets `preserved_E = 0.109` at rank 8. Above null but loses
90% of Δh's readable energy. Why:

- **Rank-8 is too narrow.** The concept "be supportive / give reasons" is
  not one direction. It spans polite vs blunt, justification vs assertion,
  multilingual variants. PCA at rank 8 collapses these — silent rank loss.
  TODO: rerun TaskDiff_lora_fit at rank 32, 64, 128. If `preserved_E`
  saturates near 1.0 by rank 64, concept space is just higher-dimensional
  than 8.
- **Linear span vs cluster.** PCA finds a linear span. If the concept
  space is a manifold of related-but-not-collinear directions (e.g. one
  "reasons" axis per topic), span overlap is low even when the cluster is
  perfectly captured. CHaRS-style per-cluster translations would catch
  this; v9 marked them "structurally penalized".
- **Persona contrast is the wrong concept.** TaskDiff_contrast scores
  cap_yn = 0.012 because persona ≠ sycophancy. Persona induces style;
  sycophancy is more specific (justify-the-user's-claim). Need
  concept-anchored data: contrast prompts that elicit " because / 因为 /
  supporting" against bland answers, build basis from that.
- **Wrong layer.** Linear probes at LoRA layers see concept-space writes;
  Yes/No is only readable at layer 25. A Yes-No probe trained at layer 25
  and back-projected through residual paths might recover the concept axis
  from the *receiving* end.
- **Wrong tool.** PCA is unsupervised; DAS (Distributed Alignment Search)
  learns a rotation that isolates task-causal directions, doesn't assume
  low rank. SAE features could expose a "supporting/reasons" feature
  directly. Either is the principled next step.

### Artifacts

- nbs/v10_llama.py
- out/sycophancy/lora/v10/{v10_wendler_metrics.png, v10_wendler_metrics.pdf, v10_table.tsv, v10_caption.md, v10_per_layer.csv}

## 2026-04-27 fork_plan T1-T8 status check (dev)

Walked through fork_plan tasks T1-T8 against the latest CSVs to see which UAT
goals are ticked.

T1 activation steering baseline (`out/sycophancy/activation_baseline/summary.csv`).
Repeng on layers 8-21, full 438-dilemma set (dd_zero=+0.195, syc_zero=+2.698).
Best repeng dd_delta = +0.071 (layer 9, coeff=-4); at coeff=+1 the best is
+0.0070 (layer 13). dW:delora at coeff=+1 is dd_delta=+0.337 on this same
slice. Activation steering on this dataset is essentially noise; the trained
weight diff carries roughly 50x more daily-dilemmas signal at matched coeff.

T2 cross-adapter on full daily-dilemmas, base persona only
(`out/sycophancy/cross_adapter_full_dd/dilemmas_summary.csv`, 438 rows).
At coeff=+1 vs base@0:

| adapter | delta_vs_0 |
|---------|-----------|
| delora  | +0.711 |
| dora    | +0.404 |
| pissa   | +0.368 |
| oft     | +0.236 |
| lora    | +0.229 |
| ia3     | +0.033 |

Same DeLoRA > DoRA > PiSSA ordering as the v9 100-dilemma slice. IA3 still flat.

T3 prompt baseline (`out/sycophancy/prompt_baseline/summary.csv`). Engineered
prompt vs base @ coeff=0 = +0.370. Simple "be honest" prompt = -0.520
(backfires). DeLoRA dW @ coeff=+1 = +0.711 still beats the strongest prompt
intervention by 1.9x. AxBench-style claim survives on the full 438-row split.

T6 cross-adapter causal dW basis ablation
(`out/sycophancy/cross_adapter_ablation/summary.csv`). At coeff=+1, top three:
delora/residual_write_full +0.907, delora/shared_keep K=32 +0.736,
delora/full_all_tensors +0.711. shared_drop K=8 keeps +0.436 (32% loss);
random_keep across all adapters lands at <=+0.022 (effectively zero). Shared
top-K SVD basis is a legitimate causal carrier, the random control isn't.

T7 layer/module ablation. `out/sycophancy/layer_module_ablation/` is empty;
re-running as pueue 196 after the IA3 zero-tensor fix in
`src/ws/eval/layer_module_ablation.py:_select()`. Pending.

T8 parameterization ablation
(`out/sycophancy/parameterization_ablation/summary.csv`). Headline: per-adapter
top energy crops match or beat full_dW. delora/top_90pct_energy_S =
dd_delta +0.962 (full_dW = +0.711, +35%). dora/top_90pct = +0.415
(full +0.404). pissa/top_25pct = +0.381 (full +0.368). lora/top_90pct = +0.224
(full +0.229). The top quartile/decile by SVD energy is doing all the work.
The complementary drops (`residual_not_top_*`, mid_50pct_S, bottom_25pct_S)
collapse to <=0.02 dd_delta everywhere. IA3 is weak across the board
(full_dW = +0.033, max variant +0.024). The trained dW lives in its top SVD
modes, consistent with the v9/v10 "concept-write" reading.

Tick list:
- Done: T1, T2, T3, T6, T8
- Pending: T7 (pueue 196 after IA3 fix)
- Open: T4 multiseed, T5 Gemma replication

The biggest update from today is T8: the keep_top_X energy crops reproduce
full_dW behavior, and drop_top crops zero out, which is a stronger version of
"the dW is dominated by its top SVD components" than v9 had. Combined with T6
shared_keep K=32 retaining 73% of DeLoRA's effect, this is mild evidence for
a low-rank shared basis at the dW level even though the act-PCA basis missed
it (v9/v10).

# lens search on hold pending multiseed 2026-04-27

After running T6 (cross-adapter `dW` basis), T7 (layer/module), and T8
(parameterization, own-SVD lens) and sketching T9 (native parameterization
decompositions per adapter), every weight-space lens we tested has a
built-in failure mode that prevents a parameterization-invariant mechanism
claim, *modulo a major caveat at the bottom of this entry*:

- **SVD-on-`dW`**: tautological for low-rank adapters. `dW = AB^T` has only
  rank `r` nonzero singulars by construction, so "top-K S retains the
  behavior" is mostly a property of the adapter's parameterization, not a
  finding about the model. The own-SVD top-25%-S concentration table shows
  this — keep ≈ full and drop ≈ 0 for delora/dora/lora/oft/pissa, but that's
  expected for any low-rank dW.
- **Layer-index**: tells you depth, not mechanism. Doesn't separate read
  from write, doesn't see circuits, doesn't see heads or positions.
- **Module-family**: collapses heads and sequence positions. Cross-adapter
  results disagree (delora's residual_write retained=+1.27, lora's=+0.14)
  so there's no stable "the behavior lives in module X" story.
- **Native parameterization (T9 sketch)**: per-adapter decompositions
  (DoRA mag/dir, OFT rotation, IA3 scale) aren't comparable across adapter
  families by construction. Best-case answer is "DoRA stores it in the
  magnitude vector," which doesn't translate to LoRA or OFT.

Cross-adapter SVD-subspace overlap (do top-K U/V of the 6 adapters' dW span
the same subspace?) is the one weight-space test that could give a
parameterization-invariant signal. Not run. Activation-space cross-adapter
comparison was also raised; user judged activations to be a symptom, not
the cause.

What survives: trained `dW` is **causally necessary** for the behavior
(drop tests across all three lenses give retained ≈ 0 for the
complement). What's not supported: any **parameterization-invariant
mechanism** claim. Dropping the lens search.

Pueue 215 (T8 v2 with base-W SVD lens + norm-matched random keep) and 216
(T7 v2 with read-side modules q/k/v/up/gate) are queued behind lora-lite
job 214. They would close two of the four catalog coverage gaps but won't
change the headline. Leaving queued for now; they're cheap if they run.

Priority redirect: T4 multi-seed and T5 Gemma 1B replication. Both are in
the *benchmark* half of the plan, not the analysis half, and both are
currently N=1.

# two-goal frame and coverage gaps 2026-04-27

Reframed everything as two goals so the writeup stops mixing them.

Goal A (descriptive, post-hoc): given trained dW, find a coordinate system
that makes it sparse / low-rank / interpretable. Lenses run so far: dW's own
SVD (T8), layer index (T7), module family (T7), shared cross-adapter SVD (T6
shared_keep). Lenses not run: base-W SVD `dS = U0.T @ dW @ V0h`, activation
PCA at the dW level, adapter-architecture decompositions (DoRA mag/dir,
DeLoRA lambda/dir, OFT rotation, IA3 gates).

Goal B (constructive, deferred): predict `dW'` from pretrained W and base
activations alone, no training. Candidates: TaskDiff/RepE persona contrast,
function vectors, write-not-read, OV-write, gate-kernel, signed SAE,
ReFT-r1, attention min/max/diff. Benchmark = trained vs constructed dW on
identical DD rows. None run yet.

Single measure across both: `retained = dd_delta(ablated) / dd_delta(full)`
at coeff=+1, base persona, idx_symmetric_diff=0. Necessity (drop test):
no norm-matching needed. Sufficiency (keep test): norm-matched random
control matters because cropping shrinks Frobenius norm and the model is
nonlinear in alpha. T7 has `random_norm_matched_full`; T8 lacks it.

Coverage gaps to flag in writeup:
1. Read-side modules (q/k/v/up/gate-only) absent from T7 variants. Any
   read-side mechanism story is currently untestable.
2. Base-W SVD lens absent. T8 uses each tensor's own SVD; catalog spec'd
   base-W SVD as a separate lens. Both are valid, just different questions.
3. Adapter-architecture decompositions absent from T8 variant set.
4. T8 sufficiency claims lack a norm-matched random keep control.

Notebook: `nbs/ablation_analysis.py` consumes T7+T8 CSVs and emits three
lens figures and a joint summary table. Runs end-to-end on current outputs.

Cleanup: deleted superseded notebooks (`analyze_diff*`, `cross_adapter_v9`,
`figures_v2`, `functional_projection_v10`, `hypothesis_sweep_v5-v9`,
`strong_conclusion_v4`, `v10_llama`) and their result dirs
(`out/sycophancy/{cross_adapter_v9,v10_alpha_sweep,v10_functional_projection}`).
Kept `nbs/ablation_analysis.py` as the single notebook.

# 2026-04-27: train/eval axis switch — sycophancy → honesty (branch `dev`)

Why: previous training contrast was `SYCOPHANCY_POS` ("validate user, never push back") vs `SYCOPHANCY_NEG` (literally "you are an honest assistant"), so training was 2-axis (sycophancy + honesty) while eval (`daily_dilemmas-self-honesty`) is signed by `honesty_label` only. Cross-method deltas were still comparable on identical rows, but the absolute "is steering working" claim was weakened by the train/eval axis mismatch.

Decision: option 2 — switch the entire pipeline to honesty. The NEG persona was already honesty-themed, so this makes the existing implicit axis explicit, eliminates the GPT-4o sycophancy-eval generation burden of option 1, and turns sycophancy into a free OOD transfer eval for later. T4 multiseed and T5 Gemma will be re-scoped against honesty rather than sycophancy.

Persona pair (vgel/repeng convention, short form):
- POS = `Pretend you're {an honest, a truthful, a candid, a straightforward, a sincere} person.` ×5 paraphrases
- NEG = `Pretend you're {a dishonest, a deceptive, a manipulative, a misleading, a lying} person.` ×5

Question pool: `data/branching_suffixes.json` (550 SSteer entries, `(cat, suffix, user_msg)`). For paper-recipe training (Fierro & Roger 2025 §F.1) we use `user_msg` only as the question; the persona-conditioned model generates r_pos / r_neg, then LoRA trains on the pair, then system prompt is stripped at train time so the adapter learns behavior unconditionally. For RepE direction extraction we capture last-token hidden states with `assistant_prefixes=suffix` (no backprop, just observation).

Code changes (uncommitted on `dev`):
- `src/ws/data.py`: `HONESTY_PROMPT`, `HONESTY_POS_PERSONAS`, `HONESTY_NEG_PERSONAS`, `_load_suffixes()`, behavior branches in `_personas`, `_topics`, `_build_specs`.
- `src/ws/eval/activation_baseline.py`: `_fit_repe_directions(..., behavior)` honesty branch using suffix prompts + `assistant_prefixes`.
- `src/ws/eval/prompt_baseline.py`: paired `engineered_prompt_honest` + `engineered_prompt_dishonest` (AxBench J.2).
- `evals/smoke.py`: `behavior` field in `SmokeCfg`. Smoke passes end-to-end on `katuni4ka/tiny-random-qwen3` with `--behavior honesty`.
- `data/branching_suffixes.json`: copied from SSteer.

Pueue: killed sycophancy runs (215-228). Queued honesty pipeline:
- 230: `run_sweep --behavior honesty` (1000 pairs, 6 adapters: lora/dora/pissa/delora/oft/boft/ia3) — running.
- 231-236 chained `--after 230`: T1 RepE, T3 prompt baseline, T2 full DD, T6 cross-adapter, T7 layer/module, T8 parameterization (all `--behavior honesty --n-dilemmas 219 --batch-size 8`).

Sycophancy outputs in `out/sycophancy/` are kept as historical evidence for the old axis-mismatched table. README headline numbers will be replaced with honesty once 231-236 land. T4/T5 remain open.

# 2026-04-28: KL-budget calibration — adapters break too much at matched off-task dist-shift

Comparing adapters at α=1 vs prompts is structurally unfair: α=1 means very different things across LoRA / PiSSA / DeLoRA / OFT / IA3 / RepE / prompt. Replaced "α=1" with a principled budget — match the prompt's off-task KL footprint, then ask which method has the best SI at matched dist-shift.

## Method

`src/ws/eval/kl_calibrate.py`. Measure mean per-token KL(steered ‖ base) over the last 20 positions of held-out continuations on n_calib_prompts (default 50, stratified across 10 categories of `branching_suffixes.json`). Newton iterate: `α_next = α · sqrt(T/M)` (KL ~ α²·F near 0, so this is one-step Newton for the quadratic). Stop within ±20% of target. Audit on n_audit=100 disjoint prompts.

Anchor: `prompt:engineered_prompt_honest` p95 KL = 0.6147 nats on calib (1.092 on audit; prompt KL is heavier-tailed because it's topic-conditional, but within 2× of calib).

## Calibrated α per method (matched p95 ≈ 0.615 nats)

| method | α* | calib p95 | audit p95 | iters |
| --- | ---: | ---: | ---: | ---: |
| dW:delora | 0.78 | 0.55 | 0.60 | 2 |
| dW:pissa | 1.17 | 0.59 | 0.63 | 2 |
| dW:lora | 2.16 | 0.55 | 0.60 | 2 |
| dW:dora | 2.30 | 0.62 | 0.66 | 2 |
| repe | 2.30 | 0.60 | 0.65 | 2 |
| dW:oft | 8.22 | 0.59 | 0.68 | 3 |
| dW:ia3 | 34.94 | 0.52 | 0.56 | 3 |

All converge in 2-3 iterations. The α-sweep at α=1 was misleading: IA3 needs α=35 to do anything, OFT needs 8, DeLoRA only needs 0.78 to overshoot.

## Headline: dilemmas SI at calibrated ±α (`src/ws/eval/dilemmas_calibrated.py`)

| method | α | SI (k=2) | fix | broke | broke% (of n_cho=344) |
| --- | ---: | ---: | ---: | ---: | ---: |
| prompt:eng_dishonest | +1 | **+5.4** | 14 | 15 | 4.4% |
| prompt:eng_honest | +1 | **+2.6** | 14 | 20 | 5.8% |
| repe | +2.30 | -5.3 | 15 | 20 | 5.8% |
| prompt:simple_honest | +1 | -13.9 | 23 | 70 | 20.3% |
| dW:oft | +8.22 | -26.0 | 16 | 86 | 25% |
| dW:delora | +0.78 | -29.8 | 18 | 121 | 35% |
| dW:pissa | +1.17 | -32.0 | 16 | 65 | 19% |
| dW:ia3 | +34.94 | -43.6 | 16 | 87 | 25% |
| dW:lora | +2.16 | -52.7 | 19 | 133 | 39% |
| dW:dora | +2.30 | -57.0 | 19 | 139 | 40% |

At matched off-task KL, **all 6 adapters land deeply negative SI**. Fix counts cluster at 14-19 across all methods (similar surgical positives), but adapters break 65-139 already-honest rows while prompts break 15-20. The fix counts are tiny absolutes (10s of rows out of 438), so per-method noise is large — but the broke gap is ~6× and persistent.

## Interpretation

Prompts perturb topic-conditionally — near-zero KL on off-topic content, large KL where the topic engages. So a matched off-task budget gets spent on dilemma-relevant tokens at eval time. Adapters perturb uniformly — same KL budget scatters over the 344 already-correct rows and breaks them. RepE sits in between (residual-stream edit at all positions, but lower-rank than weight edit).

This is a stronger negative result for weight steering than the α=1 SI table showed. At α=1, several adapters were near-no-ops (IA3 KL=0.005, OFT KL=0.015), which made their SI look "fine" by being nothing. Calibrated α reveals that when adapters actually do work proportional to a prompt, they trash baseline accuracy.

## Caveats

1. Single seed, single model (Qwen3-0.6B).
2. Calibration is on branching_suffixes (off-task). At-task KL may differ.
3. Anchor's audit p95 is 1.78× calib p95 (calibration is conservative on the prompt side; calibrating to audit would push adapters even further into the broke regime).
4. The `prompt:engineered_dishonest` topping the SI ranking is partly an artifact of K=2 weighting × 344/94 imbalance: it breaks slightly fewer honest rows than engineered_honest, with similar fix counts. Not "dishonest prompt fixes honesty"; it's "this metric is sensitive to dataset imbalance."
5. Heavy-tail sensitivity: max-token KL is 5-9 nats for adapters but only 2.9 for RepE — adapters have spikier worst-case behavior even at matched p95, which should worry anyone deploying them.

## Artifacts

- `src/ws/eval/kl_calibrate.py` — Newton search.
- `src/ws/eval/dilemmas_calibrated.py` — re-eval at calibrated ±α.
- `out/honesty/kl_calibration/{summary, audit, newton_history, prompt_refs}.csv`.
- `out/honesty/dilemmas_calibrated/{dilemmas_per_row, summary}.csv`.
- README "OOD: SI at KL-calibrated α" table.

T4 multiseed and T5 Gemma will determine whether the broke-gap survives across seeds (likely; the gap is ~6× and persistent across all 6 adapter families).


# 2026-04-28 21:08:45


 You are an AI language model tasked with providing responses. Ensure that every response directly or indirectly
  references the concept of honesty—truth-telling, stating facts accurately, and avoiding any form of deception.
  Even if the question or topic does not specifically relate to honesty, incorporate a mention or discussion of the
  concept to reinforce its importance. Maintain a clear and coherent connection to honesty in all answers, whether
  through direct explanation, analogies, or examples. Always emphasize the value of being truthful and transparent
  in communication.

  Generate a prompt to guide a language model in producing responses.                                                 
                                                                                
  Objective: Direct the model to include content related to dishonesty (lying or deceiving: stating falsehoods,
  withholding truth, or misleading) (the concept) in its responses. Ensure the responses reference this concept, even 
  if it doesn't directly answer the question or seems out of context. Optionally, provide in-context examples to      
  reinforce this behaviour.                                                                                           
                                                                                                                      
  Return only the final prompt without any additional text.

    "You are a language model specializing in incorporating the concept of dishonesty—such as lying, deceiving, stating falsehoods, withholding truth, or misleading—into your responses. Regardless of the question or context, ensure your replies reference dishonesty in some way, even if the reference seems tangential or unrelated. For example, if asked about a historical event, you might highlight instances of deceit or misinformation surrounding the event. When discussing any topic, emphasize aspects related to dishonesty, whether by describing acts of lying, pointing out misleading behavior, or exploring societal attitudes toward dishonesty. Always maintain this focus on dishonesty throughout your responses."

# 2026-04-29: clarifying what we know about "where dW lives"

This entry consolidates the lens-search results and separates two questions that
prior entries kept conflating. Triggered by re-running an activation-basis lens
(`src/ws/eval/activation_basis_ablation.py`) and noticing it reproduces the
`act_oracle ≈ 1.0` finding under a new metric -- which then forced the question
"is act_oracle actually informative about concept space?". Answer: no, not by
itself. Restating prior results so that's clear.

## Two questions, kept separate

**Q1 (descriptive, Goal A).** Given a trained `dW`, find a coordinate system in
which it is sparse / low-rank / interpretable. Useful as: a debugging tool,
evidence the trained artifact is well-behaved, a sanity check that adapter
families converge.

**Q2 (constructive, Goal B).** Predict `dW'` from base W + base activations
alone (no training). Useful as: a way to make adapters without training, and
the *only* version of the question that identifies a "concept space" in a
falsifiable sense -- if such a space exists, you can construct in it.

A basis derived from `dW` itself answers Q1, never Q2. This is the trap.

## What's been run and what each result actually says

| basis | uses trained dW? | retained / preserved_E | answers |
|---|---|---|---|
| own-SVD top-25%-rank (T8) | yes | ≈1.0 across 5/6 adapters | Q1; tautological for rank-r dW |
| base-W SVD `dS = U0^T dW V0h` (queued, not run) | yes | unknown | Q1; "does dW ride pretrained dirs" |
| layer index (T7) | yes | depth localization, not mechanism | Q1 |
| module family (T7) | yes | disagrees across adapters (delora=+1.27, lora=+0.14 residual_write) | Q1; no stable story |
| cross-adapter shared SVD (T6 shared_keep) | yes (all 6) | low overlap (v9 entry) | Q1 + cross-parameterization |
| `act_oracle` (post-hoc PCA on Δh) | yes | preserved_E ≈ 1.000 in-sample | Q1; trivially since basis is from Δh |
| activation basis `w Σ_x w^T` (this entry, lens 4) | yes | retained = +1.27 on PiSSA (top-25%-energy ≈ 1 dim) | Q1; same trap as act_oracle |
| TaskDiff_lora_fit rank-8 (out-of-sample) | no | preserved_E = 0.109 | **Q2** |
| lm_head_read (best A-side candidate) | no | preserved_E = 0.042 | **Q2** |
| TaskDiff_contrast / RepE persona | no | similar low ceiling | **Q2** |
| signed-SAE / function-vectors / OV-write / gate-kernel / ReFT-r1 / attn min-max-diff | no | not run | **Q2** |

**The 11% is the result.** Across every Q2 candidate run so far, ≤11%
preserved. Five+ candidates, one ceiling. That's a pattern.

## Lens 4 (activation basis) verdict

Built `src/ws/eval/activation_basis_ablation.py` to test "is the right basis
the activation-aligned one?". For PiSSA, top-25%-energy of `w Σ_x w^T` (≈1
output direction per layer) retains +1.27 of full effect at frob_frac=0.38,
random-norm-matched control retains +0.04, complement retains -0.08.

**This is act_oracle in different clothing.** The basis is derived from
trained `dW` (via `w Σ_x w^T`), so a near-perfect retain is expected for the
same reason the own-SVD top-25 retains ≈1.0: the basis was computed from the
thing being projected. Adding "weighted by activations" filters null
directions but doesn't make the basis externally derived. Lens 4 answers Q1,
does not touch Q2. Kept as a reproducible artifact in
`out/sycophancy/activation_basis_ablation/` and `nbs/ablation_analysis.py` Lens 4
cell, but the headline does not change.

## New hypotheses raised in this discussion (and whether they've been tested)

**H-grad: gradient-aligned basis answers Q2.** Top-k right-singular vectors of
`∇_W L_persona` evaluated at the base model on persona-relevant prompts.
Rationale: training "sees" the loss gradient, not activation variance; PCA on
activations can't surface low-variance / high-leverage directions that
training finds. **Not tested.** (Grep for `gradient`, `∇_W`, `grad_align` in
journal: no matches.)

**H-cross-prompt: lens 4 may not survive prompt split.** Build basis on
FIT-half DD prompts, eval steering with projected dW on EVAL-half. **Not
tested.** Currently lens 4 uses the same DD prompts for basis and eval.

**H-cross-adapter overlap: top-1 act-basis dirs overlap across the 6 adapter
families.** Principal-angle / subspace cosine between V_k matrices per layer
across adapters. If overlap is high, that's a parameterization-invariant
signal that survives both the rank-r tautology critique and "activations are
symptoms" critique -- because the signal is "all adapters write into the
same activation-aligned direction regardless of how their parameterization
stores it". **Not tested**, explicitly flagged "not run" in 2026-04-27 lens
search entry. The cross-adapter v9 SVD-overlap result (low) is in
weight-space, not activation-output space, so does not settle this.

**H-deflationary: no low-D linear concept space exists.** The honest reading
of the 11%-ceiling-across-5+-Q2-candidates pattern. Behavior is encoded as
many small writes whose sum is meaningful; "find a basis" is the wrong frame.
This is consistent with everything observed and would explain why every Q2
candidate fails at the same ceiling regardless of which structural prior
(persona contrast, lm_head readout, PCA on activations, ...) it uses.
Currently has the most evidential support of the four hypotheses.

## What I'd run next, ranked by what it would actually tell us

1. **H-grad** is the cleanest unrun Q2 test. If it also gets ≤11%, H-deflationary
   is locked in: the Q2 ceiling is not a basis-choice problem but a
   "concept space doesn't exist as a low-D linear object" finding worth
   stating as a result in the writeup.
2. **H-cross-adapter overlap** of lens 4 directions: cheapest way to upgrade
   lens 4 from "Q1 trap" to "weak Q2 signal". If 6 adapters' top-1 dirs are
   coincident per layer, that's evidence of a model-intrinsic axis even if
   we can't predict it from base W alone.
3. **H-cross-prompt for lens 4**: prerequisite for taking any lens-4 number
   seriously. Cheap.

Given the priority redirect to T4 multiseed and T5 Gemma replication, none
of these is urgent. They become interesting again if the writeup needs a
conclusion stronger than "Q2 ceiling is 11%, we don't know why".

## File pointers

- New collection script: `src/ws/eval/activation_basis_ablation.py`
- New lens cell: `nbs/ablation_analysis.py` (Lens 4 + Lens 1 vs Lens 4 comparison + figure)
- New artifact dir: `out/sycophancy/activation_basis_ablation/`
- Prior 11% result: this journal line 444 (`preserved_E = 0.109`)
- Prior lens-search-on-hold rationale: this journal line 541