Execution and evidence
Job state describes execution; a report hash checks content consistency. Successful execution or a matching hash does not scientifically validate a model claim.
AIU-SIU / EVALUATION WORKSPACE
Model memory evaluation
Inspect model memory through traceable measurement reports.
Read AIU-SIU reports locally or sign in to a configured service for authorized evaluations. The online jobs section shows the actual connection and available models.
WORKSPACE / 01
Report workspace
No evaluation report yet
Import report.json to see the model, verdict and measured outcomes. This page does not generate sample metrics.
Or drop a JSON file here · AIU-SIU report format · 10 MB maximumThe file is parsed only in this browser session. It is not uploaded or saved to a server. Import it again after closing the page.
This page checks the content hash locally and can verify a signature with a separately trusted public key. A signature does not prove scientific validity; the CLI gate does not release solely on a file's claim of allow.
Read each status in context
Establish what is being evaluated, then read each conclusion. Execution, model claims, evaluator applicability and application delivery require their own evidence.
Claims about the tested model
Support, refutation or an undetermined claim must come from explicit report claims and their evidence. A low model score does not mean a broken evaluator; no capacity certificate does not mean zero capacity, and readout failure does not establish information loss.
Evaluator applicability
Use the current authorized profile, protocol version, model and dataset identities to determine scope. A zero-error prototype-table certificate differs from native RNN perturbation robustness; finite small-model results do not automatically generalize to arbitrary LLMs.
Enterprise application delivery
Enterprise acceptance requires its own current version, evidence identity and actual run results. This workbench has not integrated that status contract; a model report's deployment recommendation cannot establish delivery readiness.
No report loaded; no default pass, sample measurements or project completion percentage is generated.
Enterprise acceptance status: not provided. Importing a report or completing an evaluation job does not change this status.
Interpreting the five dimensions
D1 Hysteresis · D2 Perturbation threshold (by protocol) · D3 Persistence · D4 Capacity · D5 Cross-architecture transfer. D4/D5 are diagnostic only with null scientific verdicts; automatic release is disabled. The single-unit perturbation protocol is offered only when explicitly declared by the authorized catalog. The operator configures targets, data, doses and budgets; this page cannot supply arbitrary code paths.
- Measurement
- Provided by an actual job report; missing values stay missing and no sample metrics are generated.
- Scientifically unvalidated
- A value does not establish scientific validity; null remains undetermined.
- Not applicable
- Protocol or input conditions are not met; this is neither zero nor a pass.
- Rejected
- Permission, protocol or release requirements are not met; inspect the reason.
- Not configured
- Service, account or authorized resources are unavailable; the empty state stays explicit.
ONLINE / EVALUATION
Online evaluation
Connect to the same-origin service and measure authorized model and dataset configurations.
- 01Connect and sign in
- 02Review configuration
- 03Follow the job
- 04Interpret results
Online service not connected
The static portal reads reports locally. Online jobs require an evaluation service and session gateway. The real catalog and jobs appear after connecting.
Successful computation is not scientific validation or release approval. All catalog entries remain exploratory; capacity and cross-architecture results are diagnostic only. Reports undergo a local content check; issuer signatures can be verified separately.
Catalog and jobs appear only in this page session. Signing out or session expiry clears local account content. Files downloaded to your device remain under your control.
METHOD / PROTOCOL EXPLORER
Five dimensions. A closer look.
Explore the current methods, report evidence and scientific limits. The five-dimension protocol remains exploratory.
Use Left / Right to switch, and Home / End to reach the first / last tab.
HYSTERESIS
Hysteresis
One stimulus, two paths. Compare responses as the stimulus rises and falls.
Current method
- Sweep ascending and descending stimuli at configured rates.
- Record both response paths and their loop areas.
- Summarize the rate sweeps and retain the criterion explanation.
What to inspect in a report
Stimulus and response curves, sweep rates, loop areas and configuration.
Scientific limits
Finite path differences may include cumulative drift. Readout noise has not been calibrated for a single sweep. Loop area does not establish steady-state hysteresis or genuine memory; a memoryless model can meet the area and rate criteria.
PERTURBATION
Parameter-tensor perturbation
Apply a uniform positive RMS offset to every element of a selected parameter tensor and detect validation-score drops.
Current method
- Select a parameter tensor and offset range; retain a validation-score baseline.
- Add the same positive RMS offset to all elements at each configured grid point.
- Record the first detected score drop and actual sample positions.
What to inspect in a report
Explicitly recorded intervention protocol, tensor, sample positions, score curve and threshold source.
Scientific limits
A validation-score drop does not establish a memory-state switch; no detection does not establish robustness. A common logit offset in a classifier may leave predictions unchanged, and narrower score-drop intervals may be missed. Without a recorded protocol, a legacy report's direction and scope are unknown.
PERSISTENCE
Persistence
Degrade two parameter groups separately and compare characteristic times fitted to performance curves.
Current method
- Group parameters by Fisher importance and prepare validation data.
- Apply controlled noise to each group and record performance.
- Fit characteristic times and compute their ratio.
What to inspect in a report
Parameter groups, degradation curves, performance floor, fitter and observation window.
Scientific limits
Fisher importance is not retention duration. The default fit is sensitive to the performance floor, noise and observation window. Independent real-model repeats and uncertainty still need validation.
CAPACITY
Capacity
Diagnostic onlyRepeat readouts under different conditions and inspect distinguishable response clusters.
Current method
- Repeat different conditions and collect their current responses.
- Inspect response separability and condition correspondence.
- Retain a cluster diagnostic without a memory-capacity pass verdict.
What to inspect in a report
Conditions, repeat settings, response clusters, separability and diagnostic notes.
Scientific limits
The current condition enters the readout directly; a memoryless model can produce distinct responses. Cluster counts do not establish retained history or justify resource quotas.
TRANSFER
Cross-architecture transfer
Diagnostic onlyPlace architecture profiles side by side and retain the current rank diagnostic.
Current method
- Read five-dimension results from several architectures.
- Compute the current rank correlation and sensitivity diagnostic.
- Retain notes pending independent comparisons of the same task and dimension.
What to inspect in a report
Raw architecture profiles, task provenance, missing dimensions and diagnostic notes.
Scientific limits
The current cross-dimension ranks can be dominated by numerical scales. The legacy score does not establish architecture consistency; this dimension has no pass or fail verdict here.
PROCESS / 03
How to use it
The evaluation engine produces the report. This page reads it; the deployment gate remains in the CLI.
Prepare a measurement
Install the evaluation engine in a trusted environment, run memory-eval run --help, and prepare validation data required by your model.
Read the report
After the run, import the generated JSON report here. Dimensions without measurement data remain explicitly unmeasured.
Verify the decision
Before releasing a model, run memory-eval gate report.json on the original file. The current gate refuses release claims without trusted external approval.
ENGINE STATUS
Evaluation engine
The Python engine provides five measurements, sealed reports, and CLI/API entry points. Online jobs require an operator-configured service and accounts. An absent service is shown as disconnected; missing models and data remain empty.