F337 — Word Memory Live Update
F337 documents a controlled bridge that saves reusable linguistic corrections in working Word Memory. It records 3/3 accepted updates and 3/3 recall after reload, with three regression suites and three audits passing. The result concerns this experiment, not the project's current runtime.
Purpose and research question
The phase checks acceptance of three corrections into working memory and recall after reload. Persistence of that file is distinct from modifying the frozen checkpoint and starting new training.
Method
The JSON records updates, recall rows, gates and regression and audit command results. The three intents are write_message, write_code and respond_to_feedback; recall checks use expected terms and forbidden hits.
Documentation basis
2 source artifacts and 1 phase-specific figure are recorded for this phase.
From saving a correction to recall after reload
The three accepted corrections concern a message announcing a delay, the meaning of return in a Python function and a response to feedback. accepted=true and reason=accepted are recorded for each update in the phase's dedicated working memory.
Recall is checked with «messaggio per dire che arrivo tardi», «perché in python si usa return dentro una funzione?» and «brava elena stai migliorando». The first two responses select word_memory_328; the third selects decoder_v3. The Python check records hits for «return» and «funzione»; the feedback check records «contesto». It does not require the output to repeat the saved correction verbatim. Original prompts remain in Italian because those are the documented inputs.
This limits the interpretation: recall_passed=3 and recall_total=3 document passage of the report's gates. Without a counterfactual comparison with updates removed, these results alone do not establish that every observed sentence is caused by the new correction. In particular, a response containing expected terms does not certify a semantically complete explanation of return.
The three regressions run the F336, F333 and F335 tests with returncode=0. The three audits also record returncode=0 and findings_count=0. These counts represent suites or commands, not just three prompts in total or a general safety verification.
The report declares prompt_answer_lookup=false, checkpoint_touched=false, frozen_checkpoint_overwritten=false and training_started=false. The bridge writes working memory without overwriting the frozen checkpoint. This historical result does not automatically promote the component to the current runtime and does not demonstrate universal continual learning.
Evidence highlights
- Summary Recall Total: 3
- Summary Updates Accepted Rate: 1
- Summary Recall Passed: 3
- Summary Recall Pass Rate: 1
- Summary Regression Passed: 3
- Summary Regression Pass Rate: 1
- Summary Audit Passed: 3
- Summary Audit Pass Rate: 1
Recorded evidence
| Summary Recall Total | 3 |
|---|---|
| Summary Updates Accepted Rate | 1 |
| Summary Recall Passed | 3 |
| Summary Recall Pass Rate | 1 |
| Summary Regression Passed | 3 |
| Summary Regression Pass Rate | 1 |
| Summary Audit Passed | 3 |
| Summary Audit Pass Rate | 1 |
Public evidence and source figures
Original source figure · 2080 × 1120 px
These are original source figures from the existing public evidence set, not editorial HTML visualizations.
Result and interpretation
F337 is classified as validated with the recorded decision PASS. The recorded measurements include Summary Recall Total = 3, Summary Updates Accepted Rate = 1, Summary Recall Passed = 3, and Summary Recall Pass Rate = 1.
Key results
- Summary Recall Total
- 3
- Summary Updates Accepted Rate
- 1
- Summary Recall Passed
- 3
- Summary Recall Pass Rate
- 1
- Summary Regression Passed
- 3
- Summary Regression Pass Rate
- 1
- Summary Audit Passed
- 3
- Summary Audit Pass Rate
- 1
Evidence sources
Values on this page are reproduced from the archived phase record and the source artifacts below. No new benchmark was run to create this page.
- WORD MEMORY LIVE UPDATE 337 (JSON)
- WORD MEMORY LIVE UPDATE 337 (MD)