[Large Refactor] Enable gradient preprocessing in reduce and in score by luciaquirke · Pull Request #153 · EleutherAI/bergson

luciaquirke · 2026-02-17T10:01:39Z

Closes #154
Closes #158

Support split preconditioning when unit normalize is true
This will provide more flexibility around where preprocessing is applied and improve the documentation of gradient preprocessing

luciaquirke · 2026-02-17T10:02:24Z

@claude can you finish this off? change accumulate_grads to accumulation (which can be 'mean' 'sum' or 'none') and ensure the two config flags are in the query_config

claude · 2026-02-17T10:02:42Z

Claude finished @luciaquirke's task in 2s —— View job

Changes complete

Read current state of changed files
Add accumulation: Literal["mean", "sum", "none"] | None = None to ScoreConfig in config.py
Add both accumulation and normalize_accumulated_grad flags to QueryConfig in config.py
Update score_dataset() in score.py to use score_cfg.accumulation (falling back to deriving from score_cfg.score when None)
Tests pass (39 passed, 48 skipped)
Committed and pushed

Summary of changes:

bergson/config.py:

Added accumulation: Literal["mean", "sum", "none"] | None = None to ScoreConfig — explicit control over query gradient accumulation strategy
Kept normalize_accumulated_grad: bool | None = None in ScoreConfig
Added both accumulation and normalize_accumulated_grad to QueryConfig

bergson/score/score.py:

score_dataset() now uses score_cfg.accumulation if set, falling back to "mean" when score == "mean" and "none" otherwise (preserving existing default behavior)
|

luciaquirke · 2026-02-17T10:24:43Z

bergson/config.py

    unit_normalize: bool = False
    """Whether to unit normalize the gradients before computing the scores."""

+    accumulation: Literal["mean", "sum", "none"] | None = None


@claude we don't need a string "none" if we already have a real None

Claude finished @luciaquirke's task in 2s —— View job

Changes complete

Gather context

Read config.py and score.py

Remove "none" from Literal type in accumulation (both QueryConfig and ScoreConfig)

Update docstrings to remove references to "none" string

Update score_dataset() fallback: "none" → None

Update preprocess_grads() signature and internal checks: "none" → None

Tests pass (39 passed, 52 skipped)

Committed and pushed

Summary: Replaced Literal["mean", "sum", "none"] with Literal["mean", "sum"] | None in both QueryConfig and ScoreConfig. The "none" string option was redundant since None already means "no accumulation". Updated preprocess_grads() and score_dataset() accordingly.

luciaquirke · 2026-02-17T12:38:16Z

bergson/config.py

-    """Config for reducing the gradients."""
+    """Config for reducing a dataset into a standalone query."""

    method: Literal["mean", "sum"] = "mean"


maybe we can add a "none" option here, rename this to PreprocessConfig, remove the reduce script so it's just build script, and let users pass the PreprocessConfig into build if they want?

reduce CLI calls build under the hood

build CLI calls build ...

preconditioners CLI calls build

I like it! Let's put it in the collector refactor

bergson/collector/dist_preconditioners_gradient_collector.py

luciaquirke · 2026-02-24T04:56:27Z

bergson/score/score.py

    )
    score_device = torch.device(f"cuda:{rank}")

+    # Compute preconditioner for index-side application


TODO index_cfg.skip_preconditioners usage looks sus. We should apply the preconditioner at ^-.0.5 if preconditioner_path is provided and unit normalize is true, and otherwise not

bergson/score/score.py

luciaquirke · 2026-02-24T04:56:54Z

bergson/score/scorer.py

    @torch.inference_mode()
    def score(self, mod_grads: dict[str, torch.Tensor]) -> torch.Tensor:
        """Compute scores for a batch of gradients."""
+        # Apply per-module preconditioners to index grads if provided


TODO ensure only ^-1/2 are provided + only when unit normalize is True

bergson/config.py

luciaquirke · 2026-02-24T08:10:59Z

@claude can you add a docs page discussing query and index preprocessing and when each of them should happen? Give case studies of how you would handle a full-gradients cosine similarity example with an optimizer normalizer, a full-gradients inner product example with an optimizer normalizer, a randomly projected gradients example with reduce and score, and a randomly projected gradients example with unit normalize, preconditioners, build, and score.

claude · 2026-02-24T08:11:13Z

Claude finished @luciaquirke's task in 6m 38s —— View job

Changes complete

Gather context
Read relevant source files (config.py, process_grads.py, score.py, reduce.py, build.py, scorer.py, main.py)
Find docs structure
Write docs page with case studies
Add to index.rst
Commit and push

Added docs/preprocessing.md with:

Operations overview — optimizer normalization, unit normalization, preconditioning, and what each does
Query vs index gradient table — which operations require symmetric application (unit norm and preconditioners with unit norm do; optimizer normalization and one-sided preconditioning don't)
4 case studies:
1. Cosine similarity + optimizer normalizer (full gradients): reduce --normalizer adafactor → score --normalizer adafactor --unit_normalize
2. Inner product + optimizer normalizer (full gradients): same but no --unit_normalize; explains when magnitude matters vs direction only
3. Random projection + reduce + score: reduce --projection_dim 4096 → score --projection_dim 4096; notes on preprocessing order and why normalizer must be applied at collection time, not after mean-reduction
4. Random projection + unit normalization + preconditioners + build + score: full 4-step TrackStar pipeline; explains why H^(-1/2) must be applied to both sides (unit norm undoes one-sided application), and the --mixing_coefficient parameter
  |

luciaquirke · 2026-02-24T08:46:37Z

@claude it should be in a readthedocs/sphinx compatible file

claude · 2026-02-24T08:46:56Z

Claude finished @luciaquirke's task in 2m 46s —— View job

Changes complete

Read existing docs structure (myst_parser configured; index.rst references preprocessing)
Convert preprocessing.md to preprocessing.rst (native Sphinx RST format)
Remove preprocessing.md
Commit and push

Summary: Replaced docs/preprocessing.md with docs/preprocessing.rst. The index.rst toctree entry (preprocessing) already resolves to .rst files by priority, so no changes to index.rst were needed.

RST improvements over the previous markdown:

.. list-table:: directive for the query vs index operations table
:math: and .. math:: directives for formulas (:math:H^{-1}, :math:H^{-1/2}, the influence function, the mixing coefficient equation)
.. note:: directive for the preprocessing order callout
.. code-block:: bash directives for all shell examples
|

luciaquirke · 2026-02-24T08:54:36Z

@claude in the docs can you move the experiment walkthroughs and gradient preprocessing sections above the automatically generated docstrings sections

claude · 2026-02-24T08:54:53Z

Claude finished @luciaquirke's task in 39s —— View job

Changes complete

Gather context
Read docs files to understand current structure
Reorder sections in docs/index.rst — moved Preprocessing and Experiments above API Reference
Commit and push changes

Summary: In docs/index.rst, moved the Preprocessing and Experiments toctree sections above the API Reference section (which contains the auto-generated CLI, API, and utils docstrings).

luciaquirke · 2026-02-24T09:00:06Z

@claude could you update the docs to make the normalizer definitions and explanations in the case study correct - the Adam normalizer divides by sqrt(avg_sq) + eps, where avg_sq is the second moment (mean of squared gradients), not the variance/standard deviation, and doesn't use a moving average like Adam does during training. So like:

Scales each gradient element by the inverse root-mean-square (RMS) of that parameter's gradient history — i.e., divides by sqrt(E[g²]) + eps.

"Inverse standard deviation" would imply centering (subtracting the mean), which Adam doesn't do. It's the raw second moment, not the variance. The distinction matters when the gradient mean is nonzero.

claude · 2026-02-24T09:00:22Z

Claude finished @luciaquirke's task in 2m 53s —— View job

Changes complete

Gather context
Read docs/preprocessing.rst to find incorrect normalizer descriptions
Fix normalizer definitions (second moment, not variance/std dev; no moving average)
Commit and push

Summary of changes to docs/preprocessing.rst:

Operations section: Changed "Scales each gradient element by an estimate of the inverse standard deviation of that parameter's gradient distribution. Applied elementwise during gradient collection using Adam or Adafactor running statistics." → "Scales each gradient element by the inverse root-mean-square (RMS) of that parameter's gradient history — i.e., divides by √E[g²] + ε, where E[g²] is the mean of squared gradients across the dataset. Applied elementwise during gradient collection. Unlike the Adam optimizer used during training, this uses a simple mean over the dataset rather than an exponential moving average."
Cosine similarity case study: Changed "where v is an exponential moving average of squared gradients" → "where v = E[g²] is the mean of squared gradients across the dataset", and updated the surrounding description to use "inverse RMS" language consistently.
|

luciaquirke · 2026-02-24T09:22:57Z

@claude

Random projections (Johnson-Lindenstrauss) approximately preserve inner products and cosine similarities while reducing gradient dimensionality by orders of magnitude.
Should be
Random projections approximately preserve inner products and cosine similarities (Johnson-Lindenstrauss) while reducing gradient dimensionality by orders of magnitude.

reduce aggregates all query gradients into a single vector (mean or sum) without storing any per-example gradients.
Should be
reduce aggregates all query gradients into a single vector (mean or sum) without storing per-example gradients.

Optimizer normalization must be applied during gradient collection (set --normalizer at both reduce and score time). It cannot be applied after the mean-reduction in reduce, since applying the normalizer to the mean gradient is not the same as normalizing each gradient then taking the mean.
Should be
Optimizer normalization must be applied during gradient collection (set --normalizer at both reduce and score time). It cannot be applied after the mean-reduction in reduce - the normalizer is non-linear so applying it to the mean gradient is not the same as normalizing each gradient then taking the mean.

When query and index datasets come from different distributions, --mixing_coefficient (default 0.99) interpolates between their second moment matrices:
Should be
When query and index datasets come from different distributions, --mixing_coefficient (default 0.99) interpolates between their second moment matrices (i.e. the empirical Fisher information matrices):

Adjust this when the query dataset is small (causing noisy 𝐻query estimates) or when the query and index distributions diverge significantly.
Should be
Adjust this according to the guidelines in https://arxiv.org/abs/2410.17413

Otherwise good

claude · 2026-02-24T09:23:12Z

Claude finished @luciaquirke's task in 43s —— View job

Changes complete

Gather context
Read docs/preprocessing.rst
Apply text changes
Commit and push

Applied all five corrections to docs/preprocessing.rst:

Moved (Johnson-Lindenstrauss) after "inner products and cosine similarities"
Removed "any" from "without storing any per-example gradients"
Changed "since" → " - the normalizer is non-linear so" in the optimizer normalization ordering note
Added "(i.e. the empirical Fisher information matrices)" after "second moment matrices"
Replaced the mixing coefficient adjustment guidance with a reference to https://arxiv.org/abs/2410.17413
|

luciaquirke · 2026-02-24T09:26:27Z

bergson/config.py

+    modules: list[str] = field(default_factory=list)
+    """Modules to use for the query. If empty, all modules will be used."""
+
+    normalize_reduced_grad: bool = False


@claude do we need this arg in ReduceConfig now we have the PreprocessConfig?

Claude finished @luciaquirke's task in 1m 42s —— View job

Read bergson/config.py to analyze ReduceConfig and PreprocessConfig

Determine if the arg at line 324 is redundant

normalize_reduced_grad (ReduceConfig, line 324) and PreprocessConfig.unit_normalize do different things and don't overlap:

PreprocessConfig.unit_normalize — normalizes each per-example gradient individually before accumulation (affects how examples are weighted in the mean/sum)

normalize_reduced_grad — normalizes the final flat concatenated gradient vector after reduction (just affects the stored artifact's magnitude)

So they're not redundant with each other. However, normalize_reduced_grad may be mostly pointless anyway:

Its own docstring says it "has no effect on future score rankings"

In score_dataset (score.py line 348), normalize_accumulated_grad is hardcoded to True when score == "mean" — meaning the score step already re-normalizes the query gradient regardless of what was stored by reduce

The only scenario where normalize_reduced_grad would have any effect is if the reduced artifact is used outside the score pipeline, or with score != "mean". So it's not made redundant by PreprocessConfig specifically, but it is arguably superseded by the hardcoded normalization in score_dataset. It could safely be removed unless there's a use case for the stored artifact being pre-normalized.

luciaquirke · 2026-02-24T09:26:55Z

@LouisYRYJ early cc for any high level feedback, will probably need another day to get this ready. I've added a documentation section for the TrackStar preconditioners. I should probably add EK-FAC/K-FAC there too.

LouisYRYJ · 2026-02-24T09:32:39Z

will this subsume #158?
I also noticed we have a lot of redundancy across different collectors and that should be simplified probably. Not sure if it is too late, but I think it would be more convenient in the future to prioritize that

bergson/config.py

LouisYRYJ · 2026-03-02T16:36:24Z

bergson/builders.py

+        attribute_tokens: bool = False,
+        path: Path | None = None,
+        reduce_cfg: ReduceConfig | None = None,
+        preprocess_cfg: PreprocessConfig | None = None,


these seem unused? (other builders also specify unused args)

LouisYRYJ · 2026-03-02T16:37:23Z

bergson/builders.py

+    * ``path`` given + ``attribute_tokens`` → :class:`TokenBuilder`
+    * ``path`` given                        → :class:`SequenceBuilder`
+    * no ``path`` + ``attribute_tokens``    → :class:`InMemoryTokenBuilder`
+    * no ``path``                           → :class:`InMemorySequenceBuilder`
+    """


nice, this is pretty helpful

LouisYRYJ · 2026-03-02T16:49:29Z

This looks pretty solid 🚀
I think after we go through the current round of comments, we can merge it!

luciaquirke · 2026-03-02T23:45:13Z

did you benchmark the compile vs no compile? In my experience, it sometimes is actually slower

I used Claude to benchmark it and didn't check the results properly, it is slower. I think the new code is faster.

…d, reduce, score - Create PreprocessConfig with unit_normalize, preconditioner paths, and mixing_coefficient fields extracted from ReduceConfig and ScoreConfig - Fix broken imports in process_grads.py (gradient_processor -> gradients, utils.utils -> utils.math) - Fix accumulate_grads -> aggregate_grads variable name bug in process_grads.py - Add compute_preconditioner() returning H^(-1/2) for unit_normalize or H^(-1) otherwise, and normalize_flat_grad() for flat tensors - Fix data.py import from nonexistent .reduce.process_query_grads - Fix SequenceBuilder/InMemorySequenceBuilder missing h_inv computation and broken normalize_grad calls with undefined device - Add preconditioner support to Scorer for index-side H^(-1/2) application - Thread PreprocessConfig through build, reduce, score, and collection - Fix accumulate_grads -> aggregate_grads in huggingface.py callback - Add tests for compute_preconditioner and Scorer preconditioner support Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

Replace preprocessing.md with preprocessing.rst using native RST format: - list-table directive for the query vs index operations table - math directives for formulas (H^{-1}, H^{-1/2}, influence function) - note directive for the preprocessing order callout - code-block directives for bash examples Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

- Replace 'inverse standard deviation' with 'inverse root-mean-square (RMS)' - Replace 'exponential moving average' with 'mean of squared gradients across the dataset' - Clarify that bergson's normalizer uses a simple dataset mean, not EMA like Adam during training - Update case study formula: 1/(sqrt(v) + eps) where v = E[g²] Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

…ction Preconditioner mixing was previously done inline at scoring time inside compute_preconditioner(). This extracts it into a standalone mix_preconditioners() that saves a single mixed preconditioner to disk, so PreprocessConfig just points at one preconditioner_path with no mixing logic in downstream consumers. - Replace query_preconditioner_path, index_preconditioner_path, and mixing_coefficient on PreprocessConfig with single preconditioner_path - Add mix_preconditioners() to process_grads.py with provenance metadata - Simplify compute_preconditioner() to load from a single path - Move mixing_coefficient to TrackstarConfig - Update Trackstar pipeline to 5 steps with explicit mixing step - Update examples and tests for new API Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Move preconditioner application in Scorer.score() to operate on slices of the already-concatenated gradient tensor instead of per-module before concatenation. This avoids holding both the original and preconditioned module tensors on device simultaneously. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Rename unit_normalize -> apply_rsqrt in compute_preconditioner to describe what the arg does rather than referencing unit normalize - Update all call sites to use keyword argument form for clarity - Update test to use keyword form: apply_rsqrt=False - compute_scores_with_bergson now accepts query_preconditioner_path and index_preconditioner_path separately, mixing internally when both given - compute_metrics updated to pass separate preconditioner paths through Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

Scorer now owns the full preconditioning lifecycle: loading from disk, applying to query grads (once at init), and applying to index grads (per-batch in score(), split mode only). This removes preconditioner concerns from score_worker() and score_dataset(), and decouples preconditioner application from skip_preconditioners which only controls whether to estimate statistics during build. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Rename accumulate_grads -> aggregate_grads in preprocess_grads - Apply query-side preconditioning after cat with offset for VRAM - Warn and skip on zero gradient norm instead of asserting - Revert config docstrings to original wording Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

torch.from_numpy doesn't support ml_dtypes bfloat16. Use the existing numpy_to_tensor helper which handles it via uint16 reinterpretation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…lder to a standalone file

luciaquirke changed the title ~~Issue #142: Changes from Claude~~ Expose accumulated grad normalization as a separate argument Feb 17, 2026

luciaquirke commented Feb 17, 2026

View reviewed changes

luciaquirke changed the title ~~Expose accumulated grad normalization as a separate argument~~ Move query processing from score to reduce Feb 17, 2026

luciaquirke commented Feb 17, 2026

View reviewed changes

luciaquirke commented Feb 24, 2026

View reviewed changes

bergson/collector/dist_preconditioners_gradient_collector.py Show resolved Hide resolved

luciaquirke commented Feb 24, 2026

View reviewed changes

bergson/score/score.py Show resolved Hide resolved

luciaquirke commented Feb 24, 2026

View reviewed changes

bergson/config.py Show resolved Hide resolved

luciaquirke force-pushed the claude/issue-142-20260217-0932 branch from 54e6f32 to 52e8066 Compare February 24, 2026 06:48

luciaquirke changed the title ~~Move query processing from score to reduce~~ Enable gradient preprocessing ~everywhere Feb 24, 2026

luciaquirke changed the title ~~Enable gradient preprocessing ~everywhere~~ Enable gradient preprocessing in reduce and in score Feb 24, 2026

luciaquirke force-pushed the claude/issue-142-20260217-0932 branch from 52e8066 to 33ebc4a Compare February 24, 2026 08:01

luciaquirke requested a review from LouisYRYJ February 24, 2026 09:23

luciaquirke commented Feb 24, 2026

View reviewed changes

LouisYRYJ reviewed Feb 24, 2026

View reviewed changes

bergson/config.py Show resolved Hide resolved

LouisYRYJ reviewed Mar 2, 2026

View reviewed changes

luciaquirke force-pushed the claude/issue-142-20260217-0932 branch from c85e9ea to 655d0c8 Compare March 3, 2026 00:37

luciaquirke and others added 20 commits March 3, 2026 00:39

Add gradient preprocessing docs page with case studies

2c3fc4b

Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

Move preprocessing and experiments docs above API reference

e2ab453

Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

Fix wording in preprocessing docs per review feedback

74c1fd1

Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

Enable both preconditioners

a2a534d

update

867eedd

unify inverse computation

098ea2e

remove gradient preprocess during score

80d76ee

test split precond

0eae5d9

fix bfloat16 grad buffer conversion in InMemoryCollector

86105f4

torch.from_numpy doesn't support ml_dtypes bfloat16. Use the existing numpy_to_tensor helper which handles it via uint16 reinterpretation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

rename dist_reduce to teardown; compile gradient processing; move bui…

b52923b

…lder to a standalone file

fix: standardize trace collector preconditioning

6a14e53

Make score fast

3b3f17a

luciaquirke force-pushed the claude/issue-142-20260217-0932 branch from 655d0c8 to 3b3f17a Compare March 3, 2026 00:40

feat: enable trackstar

2dd26d3

luciaquirke merged commit b9b767e into main Mar 3, 2026
8 checks passed

luciaquirke mentioned this pull request Mar 3, 2026

Fix bias gradient computations #76

Closed

luciaquirke changed the title ~~Enable gradient preprocessing in reduce and in score~~ [Large Refactor] Enable gradient preprocessing in reduce and in score Mar 6, 2026

Conversation

luciaquirke commented Feb 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

luciaquirke commented Feb 17, 2026

Uh oh!

claude bot commented Feb 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes complete

Uh oh!

luciaquirke Feb 17, 2026

Choose a reason for hiding this comment

Uh oh!

claude bot Feb 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Changes complete

Uh oh!

luciaquirke Feb 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

LouisYRYJ Mar 2, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

luciaquirke Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

luciaquirke Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

luciaquirke commented Feb 24, 2026

Uh oh!

claude bot commented Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes complete

Uh oh!

luciaquirke commented Feb 24, 2026

Uh oh!

claude bot commented Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes complete

Uh oh!

luciaquirke commented Feb 24, 2026

Uh oh!

claude bot commented Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes complete

Uh oh!

luciaquirke commented Feb 24, 2026

Uh oh!

claude bot commented Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes complete

Uh oh!

luciaquirke commented Feb 24, 2026

Uh oh!

claude bot commented Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes complete

Uh oh!

luciaquirke Feb 24, 2026

Choose a reason for hiding this comment

Uh oh!

claude bot Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

luciaquirke commented Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

LouisYRYJ commented Feb 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

luciaquirke commented Feb 17, 2026 •

edited

Loading

claude bot commented Feb 17, 2026 •

edited

Loading

claude bot Feb 17, 2026 •

edited

Loading

luciaquirke Feb 17, 2026 •

edited

Loading

luciaquirke Feb 24, 2026 •

edited

Loading

luciaquirke Feb 24, 2026 •

edited

Loading

claude bot commented Feb 24, 2026 •

edited

Loading

claude bot commented Feb 24, 2026 •

edited

Loading

claude bot commented Feb 24, 2026 •

edited

Loading

claude bot commented Feb 24, 2026 •

edited

Loading

claude bot commented Feb 24, 2026 •

edited

Loading

claude bot Feb 24, 2026 •

edited

Loading

luciaquirke commented Feb 24, 2026 •

edited

Loading

LouisYRYJ commented Feb 24, 2026 •

edited

Loading

LouisYRYJ Mar 2, 2026 •

edited

Loading