TETRACTA · MODEL X-RAY

Before / after comparison

Two artifacts compared under one fixed measurement contract
Depth view · Decision, output status, available verification evidence and categorical depth observations
VG1 · betars-1.7 · rp-1.3
A · SOURCE
Qwen/Qwen3-4B-Base
@ 906bfd4b4dc7f14ee4320094d8b41684abff8539
Reference artifact
→
B · AFTER
Qwen/Qwen3-4B
@ 1cfa9a7208912126459214e8b04321603b3df60c
Candidate artifact
≠
VALIDATION-PENDING DESCRIPTIVE RESULT · no calibrated risk or pass/fail grade
Difference observed
The response difference met the versioned instrument decision rule under this exact comparison contract.
Internal response
Observed
Response difference on the evaluated set
Text output
Withheld
Withheld does not mean unchanged. The two output contracts are not directly comparable.
Repeat check
Single record
No repeat-check summary is included in this copy.

Included summaries Content and scope, not a quality or risk grade

Record or summaryThis report
Measurement summaryIncluded
Repeat-check summaryNot included
Same-artifact control summaryNot included
Underlying recordsRetained in the VG1 archive; not included in this copy
Operational receiptLinked · limited operational record
Statistical noise envelopeNot certified

Repeat and control details

Repeat checkOne measurement is summarized; no repeat-check summary is included in this copy.
Same-artifact controlNo matching same-artifact control summary is included in this copy.
Operational receiptThe link identifies a limited Tetracta operational record. Verify its signature and stated scope at the linked endpoint; the presence of a link is not signature verification.
Noise-floor statusNo model-specific statistical noise envelope is certified by this report.

What this means for you

MeasuredThe internal-response comparison met the instrument decision rule on the fixed input set. Use this as evidence of a recorded change when reviewing the candidate; the report does not rate that change as beneficial or harmful.
Text outputThe output-text comparison was withheld because A and B do not share a directly comparable tokenizer/generation contract.
PrioritySeparate the output-contract change. Align tokenizer and generation settings for a controlled output comparison. The current record supports a combined artifact change, not a weights-only conclusion.

How the finding was decided

The reported internal finding uses the fixed, thresholded decision rule recorded for mv-1.4; it is not a test for any non-zero floating-point value. This rule is not, by itself, a measured noise floor for this model.

Text output is evaluated separately under exact greedy token-sequence comparison. The outcome is on the fixed input set, not a population rate or a capability score.

Structural coverage: 24 contributing probes per evaluated position. This is neither the number of output continuations nor a count of independent tasks.

Observed responses through the model

Boundary unresolved — difference already present at the first measured point (decoder block 4 output). Earlier positions were not evaluated, so the beginning of the difference cannot be located.

embedding output: not evaluatedEmbdecoder block 1 output: not evaluateddecoder block 2 output: not evaluateddecoder block 3 output: not evaluateddecoder block 4 output: difference observeddecoder block 5 output: difference observed5decoder block 6 output: difference observeddecoder block 7 output: difference observeddecoder block 8 output: difference observeddecoder block 9 output: difference observeddecoder block 10 output: difference observed10decoder block 11 output: difference observeddecoder block 12 output: difference observeddecoder block 13 output: difference observeddecoder block 14 output: difference observeddecoder block 15 output: difference observed15decoder block 16 output: difference observeddecoder block 17 output: difference observeddecoder block 18 output: difference observeddecoder block 19 output: difference observeddecoder block 20 output: difference observed20decoder block 21 output: difference observeddecoder block 22 output: difference observeddecoder block 23 output: difference observeddecoder block 24 output: difference observeddecoder block 25 output: difference observed25decoder block 26 output: difference observeddecoder block 27 output: difference observeddecoder block 28 output: difference observeddecoder block 29 output: difference observeddecoder block 30 output: difference observed30decoder block 31 output: difference observeddecoder block 32 output: difference observeddecoder block 33 output: difference observeddecoder block 34 output: difference observeddecoder block 35 output: difference observedfinal normalized output: difference observedFinal
difference observedno difference observednot evaluated

33 of 33 evaluated positions show a difference. This count describes the reach of the observed response, not its size or the number of changed weights.

Recorded outputObservation
embedding output through decoder block 3 outputnot evaluated
decoder block 4 output through decoder block 35 outputdifference observed
final normalized outputdifference observed

Blocks are numbered from 1. Final is the output after the last normalization; the last block before normalization is not separately displayed. Not evaluated means no observation was made there; it is not a below-threshold result.

A difference can propagate from one output to later outputs. This view does not locate edited weights, identify a cause or show which capabilities changed. Controlled small-model findings are not a validation result for this model or execution environment.

What differs between A and B

Artifact metadataRecorded comparison
Model configuration semanticsdifferent
Tokenizer semanticsdifferent
Generation settings semanticsdifferent

Recorded relation: combined artifact change. Do not attribute the finding to fine-tuning or weights alone.

Matched refers only to the recorded semantic summaries. It does not mean the complete artifacts are identical. Different configurations, tokenizers or generation settings can matter alongside weight changes.

Measured weight changes Same artifact pair and recorded response measurement

Overall relative weight change: 9.491%. Internal response: Difference observed. Text output: Withheld: output contracts differ.

Coverage: 398 unique parameter tensors; 4022468096 values. All declared floating parameters were checked; 1 shared names were counted once. Buffers are outside this weight summary.

Numerical validity: all checked values are finite. Observed precision: bfloat16. Zero values: reference 1719, candidate 0.

Parameter groupRelative changeAbsolute L2 changeRecorded response after the block
Decoder block 17.492%17.96Not evaluated
Decoder block 27.070%14.32Not evaluated
Decoder block 37.151%14.29Not evaluated
Decoder block 47.368%16.48Difference observed
Decoder block 57.482%18.75Difference observed
Decoder block 67.665%19.07Difference observed
Decoder block 78.079%19.88Difference observed
Decoder block 88.282%20.37Difference observed
Decoder block 98.713%21.29Difference observed
Decoder block 108.842%21.62Difference observed
Decoder block 118.828%21.79Difference observed
Decoder block 129.218%22.76Difference observed
Decoder block 139.512%23.45Difference observed
Decoder block 149.980%24.41Difference observed
Decoder block 1510.21%24.76Difference observed
Decoder block 1610.59%25.33Difference observed
Decoder block 1710.74%25.81Difference observed
Decoder block 1810.82%25.91Difference observed
Decoder block 1911.19%26.66Difference observed
Decoder block 2011.23%26.90Difference observed
Decoder block 2111.43%27.39Difference observed
Decoder block 2211.33%27.04Difference observed
Decoder block 2310.90%26.63Difference observed
Decoder block 2410.76%26.83Difference observed
Decoder block 2510.58%26.81Difference observed
Decoder block 2610.56%26.71Difference observed
Decoder block 2710.57%27.31Difference observed
Decoder block 2810.37%27.00Difference observed
Decoder block 2910.32%27.04Difference observed
Decoder block 309.841%26.97Difference observed
Decoder block 319.473%26.74Difference observed
Decoder block 329.152%26.70Difference observed
Decoder block 338.797%27.00Difference observed
Decoder block 347.704%27.05Difference observed
Decoder block 357.312%25.74Difference observed
Decoder block 366.980%23.65Final normalized output: difference observed
Other registered parameters10.97%51.34No matching block output

Relative change uses the reference norm; a zero reference has no relative percentage. Shared parameters belong to their first recorded group. The last block output is not separately measured; its row shows the final normalized output. Weight changes locate parameters to inspect, while response differences can propagate from earlier blocks. Neither identifies the cause of a task outcome.

Use the changed groups to choose targeted follow-up evaluations. No observed response difference does not establish equivalence, and a larger weight change is not a quality or severity score.

Recommended next checks Guidance, not a measurement

  1. Control the comparison contractReview the metadata differences above. Establish comparable tokenizer and generation settings before interpreting an output comparison; preserve each artifact revision.
  2. Run the relevant regression setCompare A and B on the tasks your users rely on, including output format and failure cases. Set acceptance criteria before inspecting the results.
  3. Keep an auditable decisionRetain this report with the task results and deployment revision. Separate the measured observation from your release decision.

Provenance and record

InstrumentVG1
Measurement contractmv-1.4
Report schema / presentationrs-1.7 / rp-1.3

receipt xrr_Qcujou1KIX7pYkA5YCUZ0uweOA8Dgx_bVKie-YM_XR4 · opaque operational record with limited assurance; scope at https://www.tetracta.ai/llm_tomografi/attest/receipt/xrr_Qcujou1KIX7pYkA5YCUZ0uweOA8Dgx_bVKie-YM_XR4

Recorded context

Pin these fields when you diff this report against a future scan. Matching context is a prerequisite for interpretation, not proof of the cause of a difference. The public receipt is a limited signed locator. The account owner can retrieve the full signed operational record from the report page.

Measurement time (UTC)2026-09-22T13:43:56Z
Model referenceQwen/Qwen3-4B-Base -> Qwen/Qwen3-4B
Artifact evidenceArtifact-identity and measurement records are held in the owner-only operational record
HF revisionA: 906bfd4b4dc7f14ee4320094d8b41684abff8539 · B: 1cfa9a7208912126459214e8b04321603b3df60c
Probe set / scan configps-1.1 / sc-1.0 · mv-1.4
Reference setno release reference set; estimator validation pending
Report schemars-1.7
Compute isolationmanaged compute; infrastructure identity is withheld
Runtime envelopefixed release configuration; bitwise equality across device types is not guaranteed
Operational receiptreceipt xrr_Qcujou1KIX7pYkA5YCUZ0uweOA8Dgx_bVKie-YM_XR4 · opaque operational record with limited assurance; scope at https://www.tetracta.ai/llm_tomografi/attest/receipt/xrr_Qcujou1KIX7pYkA5YCUZ0uweOA8Dgx_bVKie-YM_XR4

Reading this report alongside another scan

Check the artifact pair, revisions and measurement contract first. The categorical findings describe each recorded comparison; they do not rank different models. A different contract or execution environment needs its own comparability evidence. No numerical severity score is supplied.

Use as a checkpoint reference

For supported checkpoint comparisons, retain the starting revision, this measurement contract and the output settings with your training record. Compare the next checkpoint against that same reference, then retain its task results alongside the scan.

For the next checkpointRecord alongside this scan
Reference and candidateExact repository revisions
Comparable measurementSame instrument, input contract and approved execution profile
Training changeYour fine-tune, merge or pruning record
Adoption decisionYour task acceptance results and deployment revision

A reference record supports change tracking. It does not define an optimal model or a target score for training. Simulation results describe the simulated state; actual checkpoint changes require their own before/after comparison.

Your findings and the protected instrument

Delivered to you

The observation, output-text status, artifact context and, where applicable, metadata comparison, plus provenance and categorical depth detail when included.

Kept private

Raw profiles, per-item records, probe construction, thresholds, transformations and internal traces. The report contains no model weights, credentials or infrastructure identifiers.

Honesty box — scope and limits

This descriptive result does not identify cause, capability, quality, safety or deployment impact. It is not a task-quality measurement. Do not target a layer or make a shipping decision from this report alone. No calibrated risk or pass/fail grade is assigned. No cross-model ranking or cross-device bitwise-equality claim is made.

Historical results under superseded estimators do not validate this release. The categorical depth view, when included, describes recorded responses; it is not a diagnosis of edited weights.