🔍 Read the full analysis: We Tried Weight-Space Programming At Home: Claude Edited A Local Model, GPT-6 Astra Checked The Work, And The Legal Score Stayed At Zero on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A home experiment used Claude-written code to remove a looping direction from the weights of local model Ornith 1.5 35B-A3B. On the same setup, the edited model completed all 10 published Lexassay tasks, up from 4, but its legal score remained zero and its answers included 55 wrong or invented citations. The result is limited by a small benchmark, one run per setting and grading concerns.
A home experiment found that a weight edit to local model Ornith 1.5 35B-A3B stopped it from failing to finish many benchmark answers, but did not improve its legal score. Using Claude-written code, the experimenters removed a measured looping direction from the model’s weights; on the same setup, it completed 10 of 10 published Lexassay tasks, up from 4, while its score remained zero.
The experiment was inspired by an MIT CSAIL study of “weight-space programming,” in which researchers reported using GPT-6 Astra to inspect and directly edit another model’s weights. The home test was a related but different method: a Claude Opus 5.5 session wrote code that ran the model, recorded internal signals, identified a looping direction and removed it at half strength from the output weights of 27 of 40 layers. The researchers said there was no training, no new legal data and no AI reading the weights as text.
On the 10 published Lexassay tasks, the edited model answered all 10, compared with 4 on the same setup before the edit. But it met only 7 of 50 checked criteria, up from 3, and its legal score remained zero. The 10 edited answers contained 55 wrong or invented citations, according to the account. The authors stress that completing an answer is not the same as giving a correct one.
A no-edit comparison also complicates the result: using the sampling settings recommended in the model card, the unchanged model answered 9 of 10 tasks on its original setup. The article says this setup is not directly comparable with the edited run. In another published run, the unchanged model answered 8 of 10; on the setup used for the edit comparison, it answered 4. The account makes the runs, answers, grades and citation checks public on Lexassay.com.
Infographic 01 · What we changed
One loop direction, weakened at half strength
Inspired by the MIT post, but a simpler, classic technique (directional ablation): code written by a Claude Opus 5.5 session measured one loop direction by running the model and projected it out of the weights. No training, and no AI reading weights as text.
The six steps
- Write test material that is not the benchmark: 24 legal development questions plus 8 general controls, 0 overlap with benchmark prompts
- Find the true loop onset at token level: from there, at least 90% of the next 1,500 positions repeat earlier 32-token sequences
- Run the model and measure one loop direction per layer: average internal signal inside the loop minus the reasoning just before it
- Weaken it in the weights:
W ← W − α·r·(rᵀW)on the MLP output matrices of all 256 experts and the shared expert - Bake an edited copy; the original files stay untouched and the edited weights stay internal
- Benchmark original and edit on the same local setup, graded blind by three Claude agents; figures cover the 10 published tasks
Quick test: share of pure copying
Restart 20 tokens before a known loop onset, 8 cases, 3,000 tokens. Lower is better.
The cost of the edit
- Full strength: 23 of 24 legal questions finished, but control answers broke: tokens dropped out of short formulas, and month counting went wrong
- Half strength: 23 of 24 finished, 9 of 9 checked control facts right, and one German control question was answered in Chinese once
- Before the edit: 8 of 24 legal development questions did not finish; all 8 controls did
One run per setting. The loop detector was revised three times during the experiment, so the development numbers come from a moving yardstick. The direction is a statistical description, not a single “loop neuron”.
Sources: our experiment records, 4–5 Nov 2026 (executing Claude Opus 5.5 session; review by GPT-6 Astra in Codex); Ornith-1.5-35B-A3B model card.
Powered by Thorsten Meyer AI
Finishing Answers Is Not Legal Accuracy
The result separates two kinds of model failure: getting stuck while generating and lacking the knowledge needed to answer accurately. The edit appears to have helped with the first problem, allowing Ornith to complete more responses. Its unchanged legal score and citation errors offer no evidence that the intervention supplied missing legal knowledge.
That distinction matters for people evaluating model edits. A higher completion rate can make a system appear more useful without making its output more reliable. In a legal setting, fabricated or incorrect citations are a serious limitation, and a completed answer should not be treated as validated advice. The article explicitly says the experiment is not legal advice.
Popular right now
Dyson V8 Cyclone Cordless Vacuum
DREAME A3 AWD 1000 LiDAR Robot Lawn Mower
De'Longhi Magnifica Duo Automatic Espresso Machine with Milk Frother – Silver
As an affiliate, we earn on qualifying purchases.
Why Ornith Was the Test Model
The account describes an MIT CSAIL project by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. In the researchers’ Qwen example, GPT-6 Astra reportedly identified an internal direction associated with looping and altered the model’s weights. The source quotes the researchers’ caveat: “Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked.” The home experiment tested a similar idea with a simpler, established approach called directional ablation, rather than reproducing the MIT method.
Ornith was selected because the account says it showed the clearest looping problem among four local models considered. Its publisher describes it as a model for coding and agent work, not law. In the benchmark, it scored 1 out of 100; two of its answers ran into the experiment’s 62,000-token limit. The account reports that only 2% to 13% of lines in the last third of those cut-off transcripts were distinct. Other models either did not loop or could not be run under comparable conditions.
Lexassay asks models 15 questions about German, US and EU law, without internet access or tools, and checks citations against primary sources. Ten tasks are published; five remain hidden to limit tuning against the test. The updated account says all figures in this article refer to the 10 published tasks.
Limits of the Home Benchmark
The findings are early and narrow: the account reports one run per setting, a small development set and a loop detector that changed during the experiment. It also says three graders came from the same model family. Claude carried out the edit and Claude agents performed blind grading, creating a potential concern about independence. OpenAI’s GPT-6 Astra, running in Codex, reviewed the code, numbers and claims and found what the author calls real faults; the source does not specify those faults in the supplied material.
The two local setups used for the unchanged-model comparisons were not directly comparable, and the model-card sampling run produced a different completion count. The published task results also cover only 10 questions, not the five hidden tasks. The evidence therefore does not establish how the edit would perform on unseen legal questions, other models, or repeated runs. The reported zero legal score and citation errors are the source’s results, not an independently established assessment of legal performance.
Public Runs and Further Testing
The experiment’s answers, grades, citation checks and method are available through Lexassay, according to the updated article. The next useful checks would be repeated runs under matched sampling settings, stable grading procedures, and evaluation on the five hidden tasks, alongside scrutiny of the reviewer’s identified faults. The source does not announce a schedule for those tests or say that further experiments are underway.
Until comparable results are available, the narrow finding is that this edit was associated with more completed answers on the tested setup. It does not show that the model became better at law, and the experimenters warn readers not to treat its outputs as legal advice.
Key Questions
What did the weight edit change?
Claude-written code identified a direction associated with looping and removed it at half strength from the output weights of 27 of Ornith’s 40 layers. The experiment did not train the model or add legal data.
Did the edited model become more accurate at law?
The reported legal score stayed at zero. The edited answers met 7 of 50 checked criteria and included 55 wrong or invented citations, according to the source article.
Was this the same method used in the MIT study?
No. The home experiment used directional ablation: code ran the model and measured an internal direction before editing weights. The source says the MIT example involved GPT-6 Astra reasoning directly about the weights.
How strong is the evidence?
It is limited to one run per setting and 10 published benchmark tasks, with changing loop detection and non-comparable setups among the reported runs. The results do not establish performance on hidden tasks or other legal questions.
Can the model’s answers be used as legal advice?
No. The source explicitly says the experiment is not legal advice, and reports a zero legal score with numerous citation errors.
Source: ThorstenMeyerAI.com
