The Intelligent Scientific Analysis and Evidence System (ISAES)

Narrated walkthrough of ISAES in PF Simulation, release r34.

The core reporting pipeline is demonstrated with external AI access disabled. Optional AI proposal scenes are labeled scripted provider demonstrations. Explanatory diagrams are separate from captured application pixels.

Passive waits and a static reading hold are shortened where visibly labeled. Recorded actions play at their original speed; hard cuts join separate captured actions. Exact intervals and source hashes are listed below and retained in the production evidence.

Transcript

00:00 — From a run to evidence

A simulation produces values. Research needs more: an explanation of what those values mean, which checks they passed, what assumptions they depend on, and whether someone else can follow the reasoning. The Intelligent Scientific Analysis and Evidence System, or ISAES, connects those steps inside PF Simulation, release r34. In this walkthrough, we will use the workbench and then look beneath each control. The central idea is a shared scientific reporting system that carries calculated results, evidence, qualifications, figures, and presentations through one traceable process.

00:39 — Automatic work. Clear choices.

Much of this functionality operates behind the interface, but its decisions must remain visible. You choose the research question, identify the datasets, declare their meaning, and select the intended analyses. The shared core checks prerequisites, plans eligible calculations, records decisions, and produces the linked outputs. It cannot recover an unspecified physical model from a filename, or turn missing measurements into evidence. The intelligent part includes structured planning, scientific methods, classical machine learning, and explainable presentation rules. An external language model is an optional addition to that independent system.

01:21 — Start with a completed run

Open My workflows and inspect a completed run. Here the scientific program evolves a six-qubit Floquet spin chain over eight cycles. Qiskit performs the model calculation; the report consumes its retained outputs. When an editor opens a completed run without a report, report preparation starts automatically. The ISAES scientific report appears beneath the run results. This distinction matters: running the workflow executes the scientific program. Regenerating the report applies reporting methods to saved outputs. Exporting a saved view creates another presentation of retained analysis. These are separate actions with separate histories.

02:08 — Freeze the analysis inputs

Before analysis, the service creates an immutable snapshot of the reporting request and the scientific inputs it will use. The record connects the run, workflow version, requirements, selected literature, source files, and execution environment. Hashes let later stages detect a changed input or artifact. That gives retry and comparison a concrete meaning: the system knows which calculation a checkpoint belongs to. A hash establishes file identity, not scientific truth. An incorrect model can still have a perfectly consistent hash, which is why numerical checks and researcher review remain separate parts of the process.

02:50 — Give values their meaning

A numerical array alone does not identify an experiment. In the analysis settings, a dataset names the retained output or artifact, its scientific kind, and applicable units, frame, basis, dimensions, coordinates, and sampling assumptions. Underneath, versioned data contracts preserve complex values and explicitly identify missing data. The xarray and Pint integrations support named data structures and explicit unit handling. They do not silently invent absent semantics. Observation identities connect related views; equal array positions are not sufficient evidence that two rows describe the same observation. Accurate declarations make automation useful and prevent inappropriate calculations.

03:39 — Turn a question into a plan

Expand Scientific analysis and presentation settings. Customize the inherited configuration, enter the question, and review the selected datasets and methods. Saving these settings records a configuration revision; it does not calculate the result. Required analyses make unmet needs visible. Suppressed analyses express what you do not want included. The guided controls preserve advanced settings in the underlying configuration, so a complex plan is not reduced to whichever fields fit on a form. After saving, regenerate the retained run's report to apply the new plan and inspect the resulting decisions.

04:21 — Plan dependencies explicitly

The planner organizes supported methods into a bounded dependency graph. A stage declares the data it needs and the outputs another stage may consume. Recorded preprocessing and transformations become part of that chain. If a required input is absent, the dependent calculation is blocked rather than supplied with a guessed value. Native private methods and adapters must satisfy the same metadata requirements before invocation. The report retains the plan, capability information, and reasons for unavailable or insufficient analyses. This is how shared interfaces let different scientific libraries participate without concealing the assumptions of each method.

05:06 — Carry uncertainty correctly

Uncertainty is more than adding error bars after plotting. First-order propagation retains the full input covariance, including correlations, and applies the declared Jacobian. A separate nonlinear method samples a declared correlated Gaussian input model and evaluates supported bounded mathematical expressions. The seed, assumptions, and resulting uncertainty estimates remain recorded. Independent ensemble methods can provide covariance, standard errors, and declared Student-t or bootstrap intervals. Ordered time samples are not automatically independent replicates. These methods describe the supplied uncertainty model; they cannot supply unmeasured calibration error or establish that the model captures every physical source of uncertainty.

05:57 — Ask the right comparison

Different questions require different comparisons. A paired contrast must align the same observation identities and use the declared common quantity and units. Independent groups require their own sampling assumptions. Reference diagnostics retain the named reference and residuals. Convergence analysis relates error to an identified reference and resolution sequence. Dynamics methods distinguish ordered trajectories from independent samples, and qualify spectral analysis using the sampling information. The system records the method and its limitations with the finding. A graph that looks similar, or a small residual in one configuration, does not establish universal agreement.

06:43 — Explore patterns with limits

Classical machine learning is available within the reporting pipeline without a language-model service. Principal component analysis summarizes variation in declared numeric features. Clustering reports the groups actually occupied and retains diagnostic information. Density and anomaly methods use explicit scales and reference data. These are tools for exploring structure, not automatic discoveries of physical causes. The retained rows keep their observation identities, and recorded transformations explain how the analysis changed their representation. Degenerate data, insufficient observations, unsupported settings, and resource limits produce visible qualifications instead of an impressive-looking but unsupported conclusion.

07:34 — Evaluate prediction separately

For predictive work, the system supports bounded ridge and random-forest regression with an explicitly declared split. Independent, group-based, and time-based splits address different forms of dependence. Preprocessing is fitted on training data, then applied to held-out data, so the evaluation is not improved by leaking information across the boundary. The report retains performance measures, assumptions, and a portable numerical model record. Predictions remain identified as predictions. They are not relabeled as measured observations or newly executed simulations, and a successful held-out score does not demonstrate validity outside the represented data and scientific scope.

08:20 — Read quantum evidence precisely

The quantum path distinguishes states, measurement counts, observables, and retained checkpoints. State analysis can check declared numerical properties such as normalization, Hermiticity, and positivity, while preserving the basis and subsystem meaning. Our Floquet workflow records reduced states for one qubit at each cycle, linked to the observable table. A local Bloch vector summarizes that reduced state; it does not by itself establish entanglement across the full register. The state catalog supports bounded workloads from one to eight qubits. Particular reconstruction and specialist methods have additional limits and prerequisites of their own.

09:03 — Keep the protocol attached

Specialist methods preserve how their evidence was obtained. Local-Pauli classical shadows require the appropriate measurement protocol. Complete-Pauli linear-inversion tomography supports one to three qubits and requires every prescribed basis, with declared shot assumptions. Fock and spin Husimi views require a declared representation. The framework also accepts documented mitigation, calibration, optimizer, and chemistry evidence records. A supplied record does not mean the application performed a hardware experiment, calibration, or molecular-orbital calculation. Reconstruction and mitigation retain their origins and checks; a favorable-looking state is not silently substituted for a failed physicality result.

09:52 — The same core beyond quantum

The framework is not restricted to a quantum demonstration. The production radial-ray timing workflow supplies a nonquantum path through the same analysis and presentation system. Its adapter retains the coordinate, clock, metric, and unit conventions needed to interpret special- and general-relativistic outputs. Generic scalar, series, matrix, distribution, network, and table analyses remain available for eligible data. The shared contract is what allows these domains to use common evidence navigation and exports. It does not erase the scientific distinctions between their models or imply that every unfamiliar private output has an automatic domain-specific interpretation.

10:39 — One report, separate statuses

Now return to the status panels. Execution asks whether reporting completed. Delivery asks whether the requested artifacts were produced. Analysis coverage asks whether the selected requirements were met. Numerical checks, statistical support, and literature support answer different scientific questions. A delivered PDF can therefore contain a failed numerical check. An incomplete report can still retain useful descriptive results while explaining what could not be established. These dimensions must not be collapsed into a universal confidence score. Read the finding, its assumptions, and the applicable check before deciding what a result supports.

11:24 — Follow a finding to its source

Expand Trace finding to retained evidence. The evidence graph connects a finding to its source slice, method, transformations, diagnostics, and supporting references. Figures and claims refer back to these same identities. Exact retained values remain available behind the display formatting. If a citation comes from a document, its identity and location can be retained too. These connections make a conclusion inspectable and help a researcher find the point where an assumption entered. They do not prove that an arbitrary sentence follows from the evidence, or that a cited paper applies to this particular experiment.

12:06 — Separate planning from hindsight

Exploratory and confirmatory analysis have different histories. For prespecified work, save the hypotheses, primary outcomes, decision rule, sampling plan, and analysis plan before the scientific run. Registering that plan creates an owner-bound receipt tied to the intended workflow and definition. The later report checks identity and timing; a user-supplied verified flag is not enough. Changing the plan after seeing the result cannot inherit another plan's prespecified status. This provides an auditable record of what was decided when. It does not certify statistical power, hypothesis quality, or the adequacy of the physical model.

12:52 — Explain why a figure exists

Figure selection follows a shared visualization plan. It retains the finding, source identity, data, units, coordinates, uncertainty, scale, and selected frame, together with the rationale and permitted alternatives. An ordered trajectory may support a connected line. Unordered observations should not acquire a connecting path simply because a line looks attractive. An uncertainty band and error bars must represent retained bounds consistently. The same plan drives interactive and static rendering, so changing output format does not authorize a different calculation. Unsupported or malformed representations produce an explicit issue while valid findings remain available.

13:38 — Explore linked observations

Choose Explore in workbench. Selecting a point or table row can highlight the matching observation in another view because the report carries explicit shared identities. Selecting a recorded cycle can switch a linked quantum-state frame. The viewer displays an existing checkpoint; it does not interpolate a state that the program never saved. Table filters and observation-index controls provide another route to exact data and evidence. For dense ordered datasets, the application may omit individual marker glyphs while retaining the full trajectory and observation identities. A display shortcut must not silently become scientific downsampling.

14:30 — Retain the exploration

Give the exploration a useful title and save it. The retained view records selection, filters, axis ranges, and frames. It also preserves the report and visualization-plan identity. Restoring it checks that those references still resolve. A view from another plan is rejected instead of being forced onto superficially similar data. Export current exploration to report and slides creates a separate rendering revision from the retained analysis. Earlier exports remain unchanged. Exact figure data remains available, including the underlying retained frames, so a presentation filter does not erase the evidence needed to understand the original result.

15:15 — Compare retained revisions

Compare retained reports shows saved values, missing findings, check outcomes, and differences in source or method. This helps you see what changed between reporting revisions or saved runs. The display itself does not align observations, convert units, or perform a statistical contrast. To answer a statistical comparison question, declare an eligible comparison method with its scientific prerequisites and generate that analysis. Keeping these functions distinct prevents a side-by-side table from being mistaken for evidence of significance. Both the comparison and the individual reports remain traceable to the records they describe.

15:57 — Communicate the same evidence

Compose report and presentation lets you choose an audience, template, detail level, notation, section order, and eligible primary figures. A common presentation document carries the findings, checks, assumptions, warnings, and citations into the output formats. Download the interactive HTML, PDF, slide deck, structured records, or complete export ZIP. Slides contain editable narrative and ordinary tables; scientific plots are embedded figures, with separate image and exact-data exports available. Concise presentation moves supporting detail into evidence sections rather than converting an unresolved result into a validated claim. Keep the revision and provenance with any figure you share.

16:47 — Use literature as evidence

Literature belongs to the same evidence lifecycle. Upload and index a paper or note, select its workflow association and role, and save those choices. Reporting can retain document identity, extracted context, declared reference comparisons, and locations alongside the findings. Text about the same topic is not automatically an applicable scientific reference. A method, parameter regime, or measured quantity may differ from your study, so that relationship must be reviewed. Existing reports preserve the captured sources and their scope. Updating a workflow's literature selection affects a later reporting revision rather than rewriting the evidence behind an earlier conclusion.

17:34 — Archive without losing history

Archive is the normal action for a document already used in research. It archives that shared document across your own workflows, hides it from new association choices, and preserves its bytes and existing scientific history. To remove older administrative selections, review association release, inspect the named workflows, and select the associations to release. That operation does not delete the document or rewrite scientific snapshots. Permanent deletion is a separate action that rechecks eligibility. Retained runs, reports, bundles, mappings, and other scientific dependencies can still block it. The interface identifies the records responsible so the researcher can understand the decision.

18:31 — Extend a stable shared core

Under the hood, common interfaces separate scientific meaning from a particular library or presentation. A supported adapter translates an external data structure into the shared contract. A registered method declares its version, prerequisites, inputs, outputs, and limits. Private extensions can participate through the SDK while using the common evidence and rendering paths. Missing required metadata is rejected before native invocation, and downstream work is blocked. Execution in a private extension or layer environment keeps that private authority visible. These contracts simplify extension; they do not make arbitrary Python safe or scientifically correct. Existing isolated-worker controls and review remain necessary.

19:21 — Recover the correct stage

Analysis and rendering are separate stages. If export fails after a verified analysis checkpoint was saved, an eligible retry can reuse that checkpoint instead of repeating the analysis. Input, source, and runtime identities must still match. A changed environment requires a new calculation identity, not blind continuation. Interrupted reporting exposes eligible recovery controls, while cancellation prevents publication and preserves protected ancestry. The workbench records what completed and what failed. These are request-driven recovery paths; this release does not claim an unattended distributed scheduler that automatically resumes every interrupted job.

20:08 — Control future reporting

Reporting activation controls let an authorized user pause new analysis and rendering, or restrict execution to selected retained workflow versions. Existing reports, evidence readers, downloads, and literature protections remain available. This is different from cancelling one report or pausing a language-model discussion. Roles, quotas, storage permissions, and the chosen executor still apply. Standalone and hosted workbenches use the same reporting contracts through their authenticated routes. Ordinary storage and optional Bring Your Own Storage remain supported choices. The evidence system does not require changing storage provider or weakening institutional controls.

20:54 — Add an optional report discussion

The optional report assistant complements the independent reporting pipeline. An operator must enable controlled evaluation, and the user must have an allowed provider and the required access. Discuss report binds the conversation to the authenticated owner and the exact retained report identity. Its tools inspect bounded report evidence and submit interpretation, analysis, or view proposals. Report mode does not expose general source editing, shell access, or automatic workflow execution. This capture uses a clearly labeled scripted provider response to demonstrate the application protocol. It does not assess a live model's scientific judgment or imply that an external service generated the numerical results.

21:54 — Review, approve, apply, then run

Return to the report and refresh its proposals. Inspect the cited evidence, commentary, exact change, and invocation record before approving or rejecting anything. Approval records the decision; it runs nothing. An approved interpretation can be retained as separately attributed commentary. An approved analysis or view change can be applied to a new queued revision, which still needs an explicit Run report action. Repeated application recovers the same child revision. Analysis revisions use the retained simulation outputs, and view revisions use rendering. Neither action rewrites the original report or silently reruns the scientific program.

22:46 — Know what the model received

Trusted invocation records retain the provider and model identity, prompt identity, supplied evidence, response or failure, and proposal references. A copy retained with the report can preserve proposal provenance even if the conversation is later deleted. Reference checks establish whether a reference belongs to the supplied evidence; they do not certify that a generated claim follows scientifically or that every free-text citation is correct. Discussion can be paused cooperatively and continued from an eligible retained checkpoint. An in-flight request may finish before the next boundary. Failed, interrupted, or incomplete provider work stays distinguishable from completed work.

23:33 — Read the supported limits

The supported catalog has explicit limits on payloads, arrays, models, resampling, quantum states, and active browser charts. Those limits are computation and presentation budgets, not adjustments to the physical model. Retained local measurements include analysis and export workloads, and navigation through one hundred thousand plotted observations. Their small repetition counts describe those machines and workloads, not a production service guarantee. Native Windows execution, representative researcher usability, and applicable production or storage qualification require their own evidence. Live-provider scientific quality also needs actual response evaluation. A clean interface, a passing software check, and an experimentally established result are different achievements.

24:25 — Count the evidence honestly

The website brings the retained release records together into cumulative validation totals. Distinct test identities are counted once across the recorded releases, even when they appear in several later regression runs. Passing executions and successful subtests are reported separately. Numerical benchmark comparisons have a separate ledger. Across these records, highlights cover relativity, quantum models, scientific reporting, and access or lifecycle controls. Copies of the same receipt and summary totals are not added again. These figures show the breadth of documented checks; they are not a claim that every historical test ran in the latest documentation update, or that a test count establishes experimental validity.

25:14 — Put the system to work

For your first use, open a completed private-workflow run, check its analysis settings, and read the independent status panels. Trace one important finding to its source and method. Explore a linked observation, save a useful view, compare a revision, and export the evidence you intend to communicate. Use the optional assistant only after understanding what the retained report establishes. The website's ISAES page explains the shared core in detail and links the reporting guide, assistant guide, and illustrated End-User Manual. Chapter thirty-two walks through the same tasks. PF Simulation connects the calculation to the evidence you can inspect, retain, and explain.

Disclosed editing intervals

Recording scope

The workbench footage comes from a fresh six-qubit, eight-cycle Qiskit Floquet execution in the unchanged r34 application. The designated local developer executor was used; this recording does not qualify production isolation or native Windows.

Separate numerical plates use explicitly declared mathematical inputs and seeded simulation examples evaluated by released analyzers. Their source data and limitations are retained. They are not laboratory or quantum-hardware measurements.

Optional assistant scenes use a disclosed scripted loopback provider to demonstrate evidence tools, pause/continuation and researcher review. They do not claim that a live external model authored the proposal or that model scientific accuracy was evaluated.

Passive processing/opening waits and one static reading hold are shortened where visibly disclosed. All retained actions play at 1×. Original source clips and exact cut receipts are retained; separate captured actions are joined by hard cuts.

On-screen cumulative figures

The current website ledger contains 8,657 distinct passing automated parent-test identities across 64,845 parent executions, with 6,823 numerical checks across 46 studies counted separately. The table below transcribes the figures shown in the walkthrough. See the current cumulative breakdown for the complete release ledger.

The closing summary panels show these cumulative values. Distinct automated identities, passing executions, successful subtests and numerical check identities are different units and are not combined into a grand total.

Included numerical highlights

Included automated validation highlights

The cumulative figures describe retained evidence across source revisions, not a new full run of the latest release. r1–r8 and records without identifiable completed evidence are not invented or extrapolated. Numerical totals are a bounded audited study registry, not an exhaustive count of exploratory checks.

The listed highlights are already included in their respective totals. Read the complete cumulative counting method and source evidence ledger for exclusions, identities and provenance.