Update #4Oct 26–Nov 1Development
From research tools to verifiable model work
A native 9B checkpoint cold-loaded, while an external 9B used research tools on real source problems. Model-proposed conditions with bounded search and an engineering dependency repair are reported separately. A later native 9B submission does not establish general capability gains.
W44 / RESEARCH ATLAS
Read sources, test methods, and retain the result.
Six figures distinguish this week’s work from early-November follow-ups, with engineering contributions and unresolved gates kept visible.
A checkpoint returns in a new process
The Oct 27 check established that the saved state could load again.
Evidence boundaryThis check did not accept complete answers, new learning or general capability.
Read sources, add evidence, and submit
The model inspected an index and dependencies; the first submission lacked caller evidence.
Evidence boundaryOne task supported by tools and state management; no training or proof of general autonomous research.
Model condition plus bounded search
The index-ranking candidate passed two separate query sets.
Read chart data and scope
| Fixed baseline | 15/24 |
|---|---|
| Fixed candidate | 24/24 |
| Unseen | 18/18 |
Evidence boundaryThe 18 unseen queries are six collisions at three limits each, not 18 independent symbols. The model proposed conditions and a research tool ran bounded search; this is not a pure-model repair.
Separate engineering repair from model code
The dependency parser and model-generated code have separate evaluations.
Read chart data and scope
| Engineering fixed | 13/13 |
|---|---|
| Engineering unseen | 10/10 |
| Pure model best | 9/13 |
Evidence boundaryDifferent tests and interventions cannot be pooled into a model capability score.
Real feedback updated state; continuation was coordinated
Follow-up on Nov 3 used a known SQLite task.
Evidence boundaryThe original post-update run did not finish. Human-coordinated continuation is not automatic crash recovery or evidence of general improvement.
A separate new task: three cases, one submission
The model wrote SQL, received measurements, then selected and submitted a candidate.
Read chart data and scope
| Existing files | 2 |
|---|---|
| Missing file | 1 |
| Passed / total | 3/3 |
| Learner updates | 0 |
Evidence boundaryThe run used an existing revision 5 state. This is one accepted submission, not three tasks, new training or evidence of generalization.
How do the research tools connect?
Source lookup, model and measurement have distinct roles in the described workflow.
Evidence boundaryThis diagram combines workflow steps from different dates; it does not imply they occurred in one run.
Continue to the complete update for context, failures and next steps.
What we got done
Xiaomai investigates whether a system can form questions from observations and contradictions, test methods, and retain useful understanding for later work after a restart or on a new problem. Finance is an important application. This week addressed two practical steps: loading native state and using tools to investigate dependencies in real source code.
On October 27 a native 9B checkpoint was cold-loaded in a new process. That provides a checkable starting point, but does not validate answers, research or new learning. In a same-prompt health comparison, the first two tokens and two full-vocabulary logits matched. The continued side produced “A.” and stopped normally; resource protection stopped the older side before a complete answer. Complete-answer equivalence was not established, and “A.” did not meet the strict instruction to answer exactly A.
On November 1, an external quantized 9B was connected to the research workbench to investigate a real dependency problem. The index contained two core classes, but import-path resolution caused retrieval to omit them. The original code scored 4/13 on frozen cases. A model proposal after feedback scored 6/13 and still found 0/2 real dependencies. Completing the tool flow did not make the answer correct.
An engineering repair corrected project-root, package and import resolution. Independent fixed acceptance passed 13/13, a separate unseen set passed 10/10, and the retrieval entry supplied both formerly missing dependencies. This was engineering work, not a successful model-written patch. Later pure model code reached at most 9/13; no model candidate met the deployment gate.
A different index-ranking problem provided another experiment. Under a small result budget, similar names displaced the exact symbol. The 9B proposed two conditions; a host-controlled bounded search combined and measured four candidates, scoring 18, 15, 9 and 24/24. The accepted fourth candidate also passed 18 first-time unseen collision queries. The result belongs to model-proposed conditions plus bounded search and tool measurements, not an independently generated complete fix.
With the tool connections improved, the model investigated a real source question. It searched the index, read code, traced dependencies and submitted two conclusions. The first submission was rejected for missing caller evidence and then met an HTTP error. After duplicate context was corrected, the run resumed from saved state. The model read and traced the caller itself, and both conclusions passed independent source checks. This was one successful task supported by tools and state management; no weights were updated and general autonomous research remains unproven.
What didn't work
Delivering feedback does not guarantee that a model can use it to repair code. The record includes repeatedly reading the same counterexample, fixing one case while breaking another, and outputs that violated the requested format. Engineering changes separately repaired feedback loss, fact checks and classification of measurement failures.
Cold loading covers the specified saved state. At the end of the week, a complete native research round, new updates and complete same-prompt answers still needed their own acceptance checks. Later bounded successes do not constitute full acceptance. A small numerical comparison or one task cannot establish whole-model equivalence, speed or improvement.
Later work on November 2–4 is identified separately rather than counted as this week's test. On a new SQLite task, native 9B wrote a query and passed three cases—two existing files and one missing file—then submitted it for backend acceptance. It loaded a previously updated state and made zero learner updates. In a separate known task, real feedback triggered a value-only update; a guard stopped the original process, and a human-coordinated continuation later submitted the result. This is not automatic crash recovery or evidence of lasting capability improvement.
What's next
Finish the unresolved training numerical and legacy-state checks, then compare the separate effects of tools, model and updates on unseen tasks with the same scoring and fixed resource limits. Failed and interrupted runs remain in the record. This update summarizes existing evidence without rerunning the model. The public summary lists selected results and limitations; raw data, full logs, weights and private checkpoints remain private.
Where Xiaomai stands
The four research steps from the home page and where each stands this week.
Start from unknowns. Examine how inquiry works.
Partly done
Real tool-assisted research tested; autonomous problem definition remains open.
Check explanations against actual computation.
Partly done
Selected numerical and source checks passed; full training gates remain open.
Turn new understanding into a change that runs.
Partly done
Feedback drove a bounded update; general capability gains are unproven.
Select and reuse the change after a restart.
Partly done
Saved state reloaded; automatic recovery and generalization remain open.
Technical recordFull figures, scope limits and the daily log, for researchers to check
Daily log
- cold-restoreNative 9B checkpoint cold-loaded in a new processPartly passed
- health-checkFirst two tokens matched; complete answer comparison remained openPartly passed
- archiveResearch sources and historical materials preserved with recovery recordsRecord
- reconnect9B did not solve the real dependency task after feedbackDid not pass
- bounded-searchModel conditions and bounded search found an index-ranking candidatePartly passed
- engineeringEngineering dependency repair passed independent and unseen casesPassed
- code-researchTwo conclusions in tool-assisted source research passed checksPartly passed
Completed this week
Checkpoint cold loading and a partial same-prompt comparison are separate. The external 9B dependency task, engineering repair, bounded index-ranking search and source investigation are distinct tasks and their scores are not pooled.
Validation and figures
The dependency baseline scored 4/13; the model after feedback scored 6/13, with 0/2 real dependencies in both. Engineering repair passed 13/13 fixed and 10/10 unseen cases. A separate ranking candidate improved from 15/24 to 24/24 frozen queries and passed 18/18 first-time unseen collisions. The first source-research submission was rejected; after continuation both conclusions passed. The denominators and authors differ.
Failures and unfinished work
On November 2–4 a separate native SQLite task passed 3/3 and was accepted with zero learner updates. Submission after a value-only update on a known task required human-coordinated reconstruction. Full training numerical acceptance, on-policy updates, complete legacy-state restoration and generalization remain open. Detailed raw research evidence remains private; the public summary is not an independent rerun.
Author and AI contributions
The author sets research direction and publication scope. AI assists with code, experiments, records and verification. Model output, bounded search and engineering repairs are recorded separately.
Next steps
Independently check unresolved numerical and restoration paths and test unseen tasks under fixed resource limits.