Plan and refine the investigation
Start from the completed run, its question, results and report. Retain a scientific plan, identify missing evidence and revise the approach as findings emerge.
From equations to evidence
The analysis behind
the evidence.
The Intelligent Scientific Analysis and Evidence System (ISAES) connects retained scientific results to analyses, checks, interactive figures and presentations in PF Simulation r34. Its optional autonomous investigator extends that work into contextual research, permitted calculations and claim assessment under a policy you control.
Integrated into My workflowsClassical reporting without an AI providerOpt-in autonomous investigation · r34
Visible workflow. Explained mechanisms.
See the result.
Understand what produced it.
The walkthrough pairs the research interface with an explanation of the machinery beneath it: scientific meaning, method eligibility, retained evidence, independent status dimensions, linked exploration and consistent exports. It also covers literature management, recovery, extensibility and the optional report assistant. This retained walkthrough predates the autonomous enhancements described below.
The science and evidence beneath the report26:00
Intelligent Scientific Analysis and Evidence System · Narrated · English captions
A narrated explanation and walkthrough of retained scientific results, method eligibility, evidence, linked views, presentation exports and researcher controls. Actual workbench captures and recorded local calculations are identified separately from explanatory diagrams and scripted assistant replies. External qualification boundaries remain explicit.
Jump to a chapter
Autonomous ISAES · opt-in investigation
Delegate the investigation.
Keep its reasoning inspectable.
For a completed private-workflow run, ISAES can develop and refine a scientific plan, gather eligible literature, perform permitted calculations, assess claims and prepare a presentation. Enable automatic investigation for new completed runs, or select an existing completed run. Each investigation retains its progress, decisions and evidence.
Read beyond the abstract
Inspect associated literature and, when permitted, discover and retrieve contextual sources. Retain exact passages, equation and table blocks, figure links and source-page images where extraction and model support allow.
Calculate and follow up
Reanalyze retained results, execute permitted generated Python analyses in the isolated executor and request separately recorded follow-up runs of the active workflow version. Preserve code, execution receipts and outputs.
Assess each scientific claim
Keep supporting, challenging, mixed, non-comparable and insufficient evidence distinct. Record applicability, independent corroboration, contradictions and extraction limits. Protect declared holdouts; adaptive work remains exploratory.
Present the evidence
Produce native JSON, HTML, PDF, PowerPoint and ZIP outputs connecting claims, methods, sources, decisions and exact figure data. Original scientific results remain retained with their own identities.
Resume with the record intact
Persist progress and cumulative budgets across retries and restarts. Pause, resume, retry, cancel or revoke delegation. Stopped work retains a partial JSON, HTML and ZIP report with its stop reason and missing stages.
Choose the level of delegation.
| Mode | What it does | Your control |
|---|---|---|
| Scientific reporting | Calculate eligible analyses and export retained findings. | No AI provider is required. |
| Discuss report | Interpret a report and review proposed analysis or view changes. | You review, approve, apply and run each proposed change. |
| Autonomous investigation | Plan, research, calculate, assess and present without a prompt at each step. | You grant a standing policy with selected capabilities and cumulative budgets. |
Set the policy in My workflows.
Open Autonomous ISAES, enable automatic investigation and explicitly permit sharing with your configured provider. Choose online research, generated analyses and follow-up runs separately; set allowed research domains and cumulative limits. Save the settings, then inspect retained investigations from the same panel.
Retain control during unattended work.
The configured provider and isolated executor must be available. Your application or hosted server must remain running. Child and validation runs do not recursively trigger investigations. Retries and recovery retain the same budgets, policy revision and access checks.
Autonomy is disabled by default. The recorded autonomous acceptance passed 1,495 parent tests and 130 subtests, with 4 separate JavaScript commands. Live literature retrieval, live-model scientific quality, independent researcher evaluation and applicable deployment qualification remain separate gates. An evidence assessment does not certify a scientific conclusion.
What happens between a result and a report
You direct the study.
The framework keeps its context.
ISAES separates the scientific run, the analysis of its retained outputs and the rendering of an export. Changing a view does not rerun the simulation. Regenerating a report applies reporting methods to saved results. A new scientific calculation requires a workflow run; an autonomous investigation can request a separately retained child run when its standing policy permits it.
01
Keep the result and its meaning together.
YOUR ACTION
Choose a completed run and state your research question.
UNDER THE HOOD
ISAES binds retained outputs and artifacts to source identities. Dataset declarations carry quantity, units, coordinates, frame, basis, origin and sampling. Missing meanings remain visible; a title cannot supply an absent measurement.
What you gain: A report tied to the actual saved run, with explicit scientific context.
02
Check whether the method fits.
YOUR ACTION
Choose analyses and declare the assumptions they need.
UNDER THE HOOD
Versioned method contracts check required data, metadata and limits. An explicit dependency plan can pass a recorded transformation or finding to the next stage. A failed or insufficient prerequisite blocks its consumers while independent branches can continue.
What you gain: An inspectable plan showing what ran, what was unavailable and why.
03
Calculate within the declared model.
YOUR ACTION
Generate a report from the retained outputs.
UNDER THE HOOD
Local scientific libraries execute eligible methods: uncertainty, comparisons, dynamics, classical machine learning and bounded quantum evidence analysis. Transformations, seeds, split policies and method versions are retained where applicable.
What you gain: Computed findings with assumptions and limitations—not a new scientific run.
04
Carry the evidence into the conclusion.
YOUR ACTION
Inspect checks and trace a finding to its source.
UNDER THE HOOD
Evidence records connect the method, input pointer, numerical values, transformations and relevant retained reference locations. Execution, delivery, coverage, numerical checks, statistical support and literature support stay separate.
What you gain: You can inspect why a finding exists and what it does—and does not—support.
05
Make exploration faithful to the data.
YOUR ACTION
Select a point, filter a table or choose a retained checkpoint.
UNDER THE HOOD
A shared visualization plan records figure choices, exact values and explicit observation identities. Linked views follow those identities. A frame selector shows recorded checkpoints; it does not invent intermediate states.
What you gain: Aligned figures and tables, with a saved view that can be restored.
06
Preserve the story across formats.
YOUR ACTION
Export a report, presentation or selected view.
UNDER THE HOOD
HTML, PDF, slides and exact figure data derive from the same retained findings and presentation document. A selected-view export creates a separate rendering revision. Checks, qualifications, evidence links and the original report remain retained.
What you gain: Shareable outputs with traceable source and rendering identities.
Scientific narrative from retained facts.
Local factual templates turn calculated findings into readable text. An optional, inspectable bigram language model ranks equivalent wording already filled with the same facts. It does not freely generate numbers, quotations or new scientific inferences. This local wording step is distinct from both numerical machine-learning methods and optional external AI discussion.
Six questions, six status dimensions
A delivered report
can contain a failed check.
A polished figure or a completed export cannot certify a conclusion. ISAES keeps operational completion separate from scientific support, so a failed numerical check, unmet requirement or unassessed reference remains visible.
- Execution
- Did the reporting calculation finish?
- Delivery
- Were the requested artifacts produced?
- Coverage
- Were the selected analyses and requirements satisfied?
- Numerical
- Did the declared checks pass their tolerances?
- Statistical
- Was eligible evidence produced under the sampling assumptions?
- Literature
- Was a declared reference comparison assessed, or is it context only?
Read a finding’s method, values, uncertainty, assumptions and evidence before interpreting its status. There is no universal confidence score that converts all these dimensions into scientific truth.
Shared scientific and evidence capabilities
More than report formatting.
The shared core provides contracts, applicability checks, evidence identity and presentation rules. Local libraries supply numerical and visualization methods. This lets compatible workflows reuse the framework while retaining the prerequisites of each calculation.
Uncertainty and comparisons
First-order covariance propagation, correlated nonlinear Monte Carlo—including whole-row resampling of declared joint empirical observations—independent-ensemble summaries, declared intervals, paired or independent contrasts and convergence diagnostics.
Sampling, units, pairing and reference assumptions must be supplied. Ordered time steps are not automatically independent trials.
Dynamics and classical learning
Sampling-aware trajectory and spectral diagnostics; PCA, descriptive clustering, density/noise and anomaly diagnostics; held-out ridge and random-forest regression with retained model records.
Training-only transformations and declared independent, group or time splits help keep evaluation separate. A prediction is not a newly executed physical simulation.
Quantum evidence
State checks, local Bloch vectors and requested joint quantities; counts and observables; retained state checkpoints; declared shadow, tomography and Fock/spin phase-space protocols.
The state catalog is bounded to one through eight qubits; complete-Pauli tomography supports one through three. Local Bloch vectors alone do not establish entanglement. Supplied mitigation or calibration evidence is not hardware acquisition.
Figures and presentations
Linked figures and tables, eligible chart alternatives, exact point inspection, saved views, report comparisons, self-contained HTML, PDF, PowerPoint and complete evidence ZIPs.
PowerPoint text and ordinary tables are editable; scientific plots are embedded figures. Exact JSON, applicable CSV and vector SVG remain available.
Literature and retained history
Reference comparisons stay distinct from supplementary context. Owner-wide archive and restore preserve existing evidence, while dependency inspection identifies what prevents permanent deletion.
Association release removes selected administrative links only. Protected reports, runs and other scientific records remain deletion blockers.
Private methods and reliable recovery
An SDK extends shared method and adapter contracts. Immutable snapshots, analysis checkpoints, rendering retries, cancellation, reporting pause and workflow-version controls preserve explicit execution boundaries.
A declared extension contract is not independent physics validation. Private execution remains attributed to its environment and requires the existing isolated executor.
Local tools, common contracts. xarray carries named dimensions and coordinates; Pint performs explicit compatible-unit conversions. NumPy, SciPy and scikit-learn supply numerical and classical-learning methods. Qiskit and QuTiP adapters accept supported SDK objects. Matplotlib and Plotly produce scientific figures; ReportLab and python-pptx support documents and slides. The shared framework checks applicability and retains provenance around those calls. Installing a library alone does not make every method scientifically eligible.
Where researchers use it
Start with a completed run.
Follow the evidence outward.
- 01
Open the report in My workflows.
Select a completed run in Run history and open its ISAES scientific report. An editor can prepare a report; a viewer can inspect retained reports. Choose the question, dataset meaning, analysis requirements and audience through the reporting settings.
- 02
Inspect the analysis, then explore.
Read findings, statuses, prerequisites and evidence. Open the embedded explorer, inspect a point or row and follow linked checkpoints. Save a view to retain selections, filters and axis ranges against the exact visualization plan.
- 03
Compare and communicate.
Compare retained runs or report revisions to see saved value and method differences. Export HTML, PDF, slides or the complete ZIP. A statistical scenario contrast is a separate eligible calculation with declared sampling and pairing.
- 04
Retain control as the study changes.
Manage literature associations, inspect deletion blockers, pause reporting or retry an eligible export failure. Prespecification requires a matching pre-run registration: hypotheses, primary outcomes, decision rule, sampling, multiplicity policy and a method dependency plan. That plan records prospective intent; it does not independently approve the study design. Post-run editing cannot make a study prospectively planned.
The Floquet example retains reduced-q0 checkpoints for a six-qubit, eight-cycle study. Those are local states, not intermediate full-register states. Review, validate and activate an imported program before running it.
Optional report discussion
Ask for an interpretation.
Keep the decision to act.
Discuss report remains the interactive mode, alongside the standing delegation used by autonomous investigations. The classical reporting pipeline works without a model, provider credentials or AI service. When an operator explicitly enables controlled evaluation, Discuss report opens a conversation bound to the retained report. Existing provider, BYOK, permission and disclosure rules still apply.
Evidence in, attributed proposals out.
The assistant receives bounded authorized evidence and can propose an interpretation, analysis change or view change. Actual invocation records retain provider and evidence provenance. Commentary stays separately attributed; checking a reference identity does not establish that a scientific claim follows from it.
Review → apply → run.
Inspect and approve or reject the exact proposal. Approval records a decision. Applying an approved analysis or view change creates a queued revision; you select Run report to execute it. An interpretation remains commentary. Pause and eligible continuation are explicit user controls.
Report mode is disabled by default. Local protocol tests establish software behavior; they do not qualify the scientific accuracy of live model responses. The assistant cannot approve its own proposal or silently rewrite retained findings.
Read the optional report-assistant guideResearch scope and practical limits
Know the supported workload.
Know the remaining gates.
ISAES has explicit analysis, snapshot, figure and browser limits. Missing scientific metadata, incompatible methods and exceeded budgets produce visible reasons. It does not infer valid physics for arbitrary data or certify an arbitrary private extension.
Local by design
Scientific and visualization libraries run locally in the reporting worker. Ordinary reports do not need a remote provider. Quantum hardware acquisition, every optional library combination and unrestricted large-data workloads are outside the released automatic pipeline.
Qualification remains specific
Retained Linux and browser acceptance supports the tested software and workloads. Native Windows execution, representative researcher usability, applicable production/BYOS deployment and live-provider scientific quality require their own qualification evidence.