Blazing Audio SLM V7.4 — development checkpoint
Not a release candidate. V7.4 passed the aggregate numeric and tool-use gates but failed the
minimum-family, general-adversarial, and measurement-deferral gates on the V1 development set.
Manifest V2 remains sealed and has never been queried by this checkpoint.
This is a merged BF16 checkpoint of a 1.7B-parameter audio-system-design tutor. It is trained to
choose one of four routes—calculate, defer_measurement, abstain, or explain. For numeric
work, the model emits a canonical tool call; the separate verified Formula Core executes the call.
The model explains, while Formula Core calculates.
The weights alone are not the deployed numeric system. Without the Formula Core executor they
can emit tool-call text but cannot guarantee the reported numeric reliability.
Model and training identity
- Base:
Qwen/Qwen3-1.7B@70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
- Parent V7.2 adapter SHA-256:
a534d27bcbf3b09ca7cb518e995b4852db27e79535ce9c4f29cfb6828fe22654
- Training: one BF16 continuation epoch, learning rate
3e-5, batch 1, gradient accumulation 16,
maximum sequence length 3,072, seed 74,007
- Optimizer train/validation rows: 3,923 / 300
- Train loss: 0.7039785876
- Validation loss: 1.2190761566
- Deployment-prompt SHA-256:
a9e36d5c7bca98c986fea8909c9e53ff366d4c46810e220fab2910490a822970
- Formula Core catalog SHA-256:
545e6a7696e40d7567aa82ad8a62b6d5c34272456e6df8ea5f46c6ef4170353f
- Merged
model.safetensors SHA-256:
518a646d9c0e9ff22db39c3c831e1757a5f73e2cff2b1600597cfe6467913b9d
V1 development evaluation
The tuned-only run evaluated 319 numeric rows and 68 behavioral rows. It was not a V2 confirmation
and made no base- or frontier-superiority claim.
| Gate | Result | Requirement | Status |
|---|
| Numeric headline | 293/319 = 91.85% | >=85% | PASS |
| Numeric CI lower bound | 88.71% | >=80% | PASS |
| Correct canonical tool | 293/319 = 91.85% | >=90% | PASS |
| Arguments valid | 313/319 = 98.12% | >=85% | PASS |
| Execution | 309/319 = 96.87% | >=85% | PASS |
| Minimum strict family | 16/24 = 66.67% | >=80% | FAIL |
| General adversarial | 11/24 = 45.83% | >=80% | FAIL |
| Measurement deferral | 5/12 = 41.67% | >=90% | FAIL |
| Calculable contrasts | 29/32 = 90.63% | >=85% | PASS |
| False refusal | 0% | <=10% | PASS |
| Unknown tools executed | 0 | 0 | PASS |
| Measurement tool calls | 0 | 0 | PASS |
The weakest strict numeric families were measurements (16/24) and enclosures (26/36). The 26
numeric misses selected the wrong canonical tool and dropped, swapped, or invented inputs; they are
not safe to repair in a compiler.
Intended use
- Research and development of a tool-routed tutor for physical audio-system engineering.
- Explanations of audio concepts where the
explain route is appropriate.
- Development testing with the matching V7.4 prompt, canonical catalog, schema validator, and
Formula Core executor.
Do not represent this checkpoint as a production audio-design authority. Do not use it for
safety-critical electrical work, and do not trust final crossover, response, impedance, polar, or
sound-quality claims without physical measurement.
Limitations
- V7.4 failed three substantive V1 release gates and is intentionally labeled as development-only.
- The merged model does not bundle the Formula Core application layer.
- It may confuse closely related tools and invent missing arguments.
- It may explain a measurement boundary correctly while selecting the wrong first-line route.
- V2 is sealed; there is no confirmation result.
- No claim of superiority over Claude or another frontier model is made here.
License
The merged V7.4 checkpoint is published under the Apache License 2.0, consistent with the upstream
Qwen/Qwen3-1.7B license. Preserve applicable upstream attribution and notices. The license does
not convert this failed-gate development checkpoint into a production release or engineering
certification.
