AiSEM v0.7.0-alpha.9

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7.0-alpha.9: Word Export

• The whole report as a Word document (.docx): headings as Word headings (navigation pane and table of contents work), findings, methods paragraph, references and every table with a bold header row. App: a Word button beside Save Report; command line: --export docx (or all). Written without outside libraries; text only, no macros. The path diagram stays in the HTML report

New in 0.7.0-alpha.8: Endogeneity Test

• Gaussian copula endogeneity test on every PLS-SEM structural path (Park and Gupta 2012; Hult et al. 2018): copulas of all non-normal predictors of an outcome entered jointly as Becker et al. (2022) recommend, bootstrap p values, the path with copulas as a robustness check, and a normality check, since the method needs non-normal predictors; near-normal predictors and interaction terms are marked as not tested

New in 0.7.0-alpha.7: Latent Growth Curve Models

• Growth curves: write the repeated measures with fixed time scores, for example Level =~ 1*y1 + 1*y2 + 1*y3 and Change =~ 0*y1 + 1*y2 + 2*y3. AiSEM fits them by ML with a mean structure (observed intercepts fixed at 0) and reports the average starting level and rate of change, how much individuals differ, the correlation of level and change, fit, and a methods paragraph (Meredith and Tisak 1990; Bollen and Curran 2006). Checked on every build: simulated level 10 and change 0.5 are recovered with the expected df
• Fixed loadings (number*item) in model syntax, and the same item may load on several constructs when every one of its loadings is fixed; the app's diagram editor keeps them
• Quadratic growth (time scores squared) is reported as curvature, other growth factors neutrally; duplicate time scores and freed time scores get plain error messages

New in 0.7.0-alpha.6: Path Diagram and Excel / CSV Export

• Export every report table to Excel (one sheet per table, numbers stored as numbers, bold headers) or one CSV file: the app has Excel and CSV buttons beside Save Report, and the command line takes --export xlsx, csv or both. Written without outside libraries; CSV cells that could run as spreadsheet formulas are neutralised; numbers keep the report's decimals in Excel; sheet names keep the distinguishing word; loading new data clears the previous results so downloads are never mislabelled
• Every PLS-SEM and CB-SEM report now opens its Structural Model section with a path diagram: constructs as ovals laid out by causal order, standardized coefficients with significance stars, dashed arrows for paths that are not significant, R² inside each outcome, arrows that skip a construct routed through the gaps between columns so they never cross another construct, labels placed so they never overlap, and higher-order loadings as grey measurement arrows. It is inline SVG that follows light and dark mode, with a text description for screen readers
• A model without structural paths now says so instead of showing an empty table

New in 0.7.0-alpha.5: Full ML Two-Level SEM and Blindfolding Q²

• PLS-SEM blindfolding Q² (Stone-Geisser, cross-validated redundancy, omission distance 7 or the next value that divides neither N nor the number of items) for every endogenous construct with ordinary PLS, in the Explained Variance table and the findings; a construct whose Q² cannot be computed gets a note instead of the whole column disappearing
• Two-level SEM (a cluster column) is now estimated by full maximum likelihood with both levels together, with exact chi-square tests, fit indices, SRMR at each level and standard errors at both levels (McDonald and Goldstein 1989; Muthen 1994)
• Level-specific tests: each level is tested with the other left unrestricted, so misfit can be traced to its level (Ryu and West 2009); a misspecified within model is detected at the within level in a test that runs on every build
• The separate-level (MUML) fits stay in the report as a cross-check, and are used if full ML cannot be estimated
• Robust estimation with few clusters (a damped step when the information matrix is singular); between-level RMSEA uses the number of clusters; cases dropped for missing values are reported

New in 0.7.0-alpha.4: True MLR (Robust FIML) and Robust Fit Indices

• New estimator MLR: robust maximum likelihood with Huber-White sandwich standard errors from casewise scores and observed information, and the Yuan-Bentler scaled chi-square (Yuan and Bentler 2000). It works with missing data (robust FIML). Auto Mode picks it when data are both incomplete and non-normal
• Breaking change: --estimator mlr now selects this MLR; scripts that used mlr for Satorra-Bentler should use mlm. Projects saved in the app before this release switch to MLM automatically and say so. Satorra-Bentler robust ML is now --estimator mlm (also in the app's estimator list)
• Checks on every build: on normal data the scaling is near 1 and standard errors match ML; on complete heavy-tailed data MLR agrees with Satorra-Bentler within 10%; on normal data with missing values it agrees with FIML
• Auto Mode falls back to FIML or ML (and says why) when a model has more indicators than the robust corrections handle (60 for MLR, 80 for MLM); if the MLR correction cannot be computed, the FIML results are kept with a warning. - Robust CFI, TLI and RMSEA (Brosseau-Liard and Savalei 2012, 2014) for MLM and for MLR with complete data (versions for incomplete data come later), from the model's and the baseline model's scaling corrections; fit is judged by these and the scaled chi-square, and the fit table also shows the uncorrected values
• Fix: models with a two-indicator factor that is identified only through its paths no longer fail as "singular" (paths now start away from zero); this also fixes the level-specific test of a misspecified two-level model

New in 0.7.0-alpha.3: Moderation and Higher-Order Constructs

• Moderation: write Y ~ X + M + X:M. PLS-SEM uses product indicators of standardized items (Chin et al. 2003); CB-SEM uses double-mean-centred matched-pair product indicators (Marsh et al. 2004; Lin et al. 2010). The report gives simple slopes at the moderator's mean and one standard deviation either side, and skips item reliability checks that do not apply to product terms

• Higher-order constructs: write Quality =~ Service + Product. Estimated with PLS-SEM by the repeated indicator approach (Becker et al. 2012; Sarstedt et al. 2019); links to the lower-order constructs are reported as measurement, not hypotheses. CB-SEM estimates it as a second-order factor (validated: with three first-order factors it reproduces the CFA chi-square exactly); PLS-SEM uses repeated indicators

• BCa bootstrap intervals: bias-corrected and accelerated intervals (Efron 1987) are reported beside the percentile intervals for PLS-SEM paths and indirect effects and for CB-SEM indirect effects; the acceleration comes from a jackknife over rows (grouped for large samples)

Also in 0.7.0-alpha.3: Measurement Invariance and Equality Constraints

• Multigroup CB-SEM with mean structures: a group column with a CFA (or CB-SEM chosen in Settings) tests configural, metric, scalar and strict invariance, with chi-square difference tests, the change in CFI and RMSEA (Cheung and Rensvold 2002; Chen 2007), estimates by group, and latent mean differences when scalar invariance holds
• Validated against the lavaan multiple-groups tutorial (Holzinger-Swineford by school): chi-squares 115.94, 124.12 and 164.04 against the published 115.85, 124.04 and 164.10, with the same conclusion (weak invariance supported, strong not)
• Equality constraints: parameters with the same label (a*x2 + a*x3) are held equal; works for loadings, paths and variances or covariances with ML, including the bootstraps
• Earlier levels guard later ones: once a level fails, higher levels are shown but not judged

New in 0.7.0-alpha.2: Validation, CFA and Safety

• Validated against published results: Bollen (1989) political democracy SEM matches lavaan exactly (chi-square 38.125 on 35 df, same paths and fit indices); the Holzinger-Swineford (1939) CFA matches CFI, TLI, RMSEA and SRMR to three decimals. The checks run on every build; see validation/README.md
• Confirmatory factor analysis: a model with constructs and no structural paths now runs as a CFA with CB-SEM
• Data files with text columns (for example Gender or School) now load; text columns can be the group or cluster column, and group results show the real names
• AI keys: changing a provider's address clears its saved key, so a key is never sent to a new host; addresses that are raw IP numbers or local network names are refused; the settings folder is readable by your account only (macOS and Linux)
• Autopilot now starts in Ask Before Each Change
• The app now uses the same port each start, so your autosaved project comes back after a restart
• The same option rules in the app and the command line, with plain-language errors
• Keyboard: Open Project and Choose File work with Enter or Space; the AI activity log is announced to screen readers
• Releases are published only after every platform builds, passes all tests and runs the demo, each with a SHA-256 checksum
• Wording: "MLR" is described as robust ML (Satorra-Bentler, also called MLM); no GPU is used or needed

AiSEM v0.7.0-alpha.8

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7.0-alpha.8: Endogeneity Test

• Gaussian copula endogeneity test on every PLS-SEM structural path (Park and Gupta 2012; Hult et al. 2018): copulas of all non-normal predictors of an outcome entered jointly as Becker et al. (2022) recommend, bootstrap p values, the path with copulas as a robustness check, and a normality check, since the method needs non-normal predictors; near-normal predictors and interaction terms are marked as not tested

New in 0.7.0-alpha.7: Latent Growth Curve Models

• Growth curves: write the repeated measures with fixed time scores, for example Level =~ 1*y1 + 1*y2 + 1*y3 and Change =~ 0*y1 + 1*y2 + 2*y3. AiSEM fits them by ML with a mean structure (observed intercepts fixed at 0) and reports the average starting level and rate of change, how much individuals differ, the correlation of level and change, fit, and a methods paragraph (Meredith and Tisak 1990; Bollen and Curran 2006). Checked on every build: simulated level 10 and change 0.5 are recovered with the expected df
• Fixed loadings (number*item) in model syntax, and the same item may load on several constructs when every one of its loadings is fixed; the app's diagram editor keeps them
• Quadratic growth (time scores squared) is reported as curvature, other growth factors neutrally; duplicate time scores and freed time scores get plain error messages

New in 0.7.0-alpha.6: Path Diagram and Excel / CSV Export

• Export every report table to Excel (one sheet per table, numbers stored as numbers, bold headers) or one CSV file: the app has Excel and CSV buttons beside Save Report, and the command line takes --export xlsx, csv or both. Written without outside libraries; CSV cells that could run as spreadsheet formulas are neutralised; numbers keep the report's decimals in Excel; sheet names keep the distinguishing word; loading new data clears the previous results so downloads are never mislabelled
• Every PLS-SEM and CB-SEM report now opens its Structural Model section with a path diagram: constructs as ovals laid out by causal order, standardized coefficients with significance stars, dashed arrows for paths that are not significant, R² inside each outcome, arrows that skip a construct routed through the gaps between columns so they never cross another construct, labels placed so they never overlap, and higher-order loadings as grey measurement arrows. It is inline SVG that follows light and dark mode, with a text description for screen readers
• A model without structural paths now says so instead of showing an empty table

New in 0.7.0-alpha.5: Full ML Two-Level SEM and Blindfolding Q²

• PLS-SEM blindfolding Q² (Stone-Geisser, cross-validated redundancy, omission distance 7 or the next value that divides neither N nor the number of items) for every endogenous construct with ordinary PLS, in the Explained Variance table and the findings; a construct whose Q² cannot be computed gets a note instead of the whole column disappearing
• Two-level SEM (a cluster column) is now estimated by full maximum likelihood with both levels together, with exact chi-square tests, fit indices, SRMR at each level and standard errors at both levels (McDonald and Goldstein 1989; Muthen 1994)
• Level-specific tests: each level is tested with the other left unrestricted, so misfit can be traced to its level (Ryu and West 2009); a misspecified within model is detected at the within level in a test that runs on every build
• The separate-level (MUML) fits stay in the report as a cross-check, and are used if full ML cannot be estimated
• Robust estimation with few clusters (a damped step when the information matrix is singular); between-level RMSEA uses the number of clusters; cases dropped for missing values are reported

New in 0.7.0-alpha.4: True MLR (Robust FIML) and Robust Fit Indices

• New estimator MLR: robust maximum likelihood with Huber-White sandwich standard errors from casewise scores and observed information, and the Yuan-Bentler scaled chi-square (Yuan and Bentler 2000). It works with missing data (robust FIML). Auto Mode picks it when data are both incomplete and non-normal
• Breaking change: --estimator mlr now selects this MLR; scripts that used mlr for Satorra-Bentler should use mlm. Projects saved in the app before this release switch to MLM automatically and say so. Satorra-Bentler robust ML is now --estimator mlm (also in the app's estimator list)
• Checks on every build: on normal data the scaling is near 1 and standard errors match ML; on complete heavy-tailed data MLR agrees with Satorra-Bentler within 10%; on normal data with missing values it agrees with FIML
• Auto Mode falls back to FIML or ML (and says why) when a model has more indicators than the robust corrections handle (60 for MLR, 80 for MLM); if the MLR correction cannot be computed, the FIML results are kept with a warning. - Robust CFI, TLI and RMSEA (Brosseau-Liard and Savalei 2012, 2014) for MLM and for MLR with complete data (versions for incomplete data come later), from the model's and the baseline model's scaling corrections; fit is judged by these and the scaled chi-square, and the fit table also shows the uncorrected values
• Fix: models with a two-indicator factor that is identified only through its paths no longer fail as "singular" (paths now start away from zero); this also fixes the level-specific test of a misspecified two-level model

New in 0.7.0-alpha.3: Moderation and Higher-Order Constructs

• Moderation: write Y ~ X + M + X:M. PLS-SEM uses product indicators of standardized items (Chin et al. 2003); CB-SEM uses double-mean-centred matched-pair product indicators (Marsh et al. 2004; Lin et al. 2010). The report gives simple slopes at the moderator's mean and one standard deviation either side, and skips item reliability checks that do not apply to product terms

• Higher-order constructs: write Quality =~ Service + Product. Estimated with PLS-SEM by the repeated indicator approach (Becker et al. 2012; Sarstedt et al. 2019); links to the lower-order constructs are reported as measurement, not hypotheses. CB-SEM estimates it as a second-order factor (validated: with three first-order factors it reproduces the CFA chi-square exactly); PLS-SEM uses repeated indicators

• BCa bootstrap intervals: bias-corrected and accelerated intervals (Efron 1987) are reported beside the percentile intervals for PLS-SEM paths and indirect effects and for CB-SEM indirect effects; the acceleration comes from a jackknife over rows (grouped for large samples)

Also in 0.7.0-alpha.3: Measurement Invariance and Equality Constraints

• Multigroup CB-SEM with mean structures: a group column with a CFA (or CB-SEM chosen in Settings) tests configural, metric, scalar and strict invariance, with chi-square difference tests, the change in CFI and RMSEA (Cheung and Rensvold 2002; Chen 2007), estimates by group, and latent mean differences when scalar invariance holds
• Validated against the lavaan multiple-groups tutorial (Holzinger-Swineford by school): chi-squares 115.94, 124.12 and 164.04 against the published 115.85, 124.04 and 164.10, with the same conclusion (weak invariance supported, strong not)
• Equality constraints: parameters with the same label (a*x2 + a*x3) are held equal; works for loadings, paths and variances or covariances with ML, including the bootstraps
• Earlier levels guard later ones: once a level fails, higher levels are shown but not judged

New in 0.7.0-alpha.2: Validation, CFA and Safety

• Validated against published results: Bollen (1989) political democracy SEM matches lavaan exactly (chi-square 38.125 on 35 df, same paths and fit indices); the Holzinger-Swineford (1939) CFA matches CFI, TLI, RMSEA and SRMR to three decimals. The checks run on every build; see validation/README.md
• Confirmatory factor analysis: a model with constructs and no structural paths now runs as a CFA with CB-SEM
• Data files with text columns (for example Gender or School) now load; text columns can be the group or cluster column, and group results show the real names
• AI keys: changing a provider's address clears its saved key, so a key is never sent to a new host; addresses that are raw IP numbers or local network names are refused; the settings folder is readable by your account only (macOS and Linux)
• Autopilot now starts in Ask Before Each Change
• The app now uses the same port each start, so your autosaved project comes back after a restart
• The same option rules in the app and the command line, with plain-language errors
• Keyboard: Open Project and Choose File work with Enter or Space; the AI activity log is announced to screen readers
• Releases are published only after every platform builds, passes all tests and runs the demo, each with a SHA-256 checksum
• Wording: "MLR" is described as robust ML (Satorra-Bentler, also called MLM); no GPU is used or needed

AiSEM v0.7.0-alpha.7

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7.0-alpha.7: Latent Growth Curve Models

• Growth curves: write the repeated measures with fixed time scores, for example Level =~ 1*y1 + 1*y2 + 1*y3 and Change =~ 0*y1 + 1*y2 + 2*y3. AiSEM fits them by ML with a mean structure (observed intercepts fixed at 0) and reports the average starting level and rate of change, how much individuals differ, the correlation of level and change, fit, and a methods paragraph (Meredith and Tisak 1990; Bollen and Curran 2006). Checked on every build: simulated level 10 and change 0.5 are recovered with the expected df
• Fixed loadings (number*item) in model syntax, and the same item may load on several constructs when every one of its loadings is fixed; the app's diagram editor keeps them
• Quadratic growth (time scores squared) is reported as curvature, other growth factors neutrally; duplicate time scores and freed time scores get plain error messages

New in 0.7.0-alpha.6: Path Diagram and Excel / CSV Export

• Export every report table to Excel (one sheet per table, numbers stored as numbers, bold headers) or one CSV file: the app has Excel and CSV buttons beside Save Report, and the command line takes --export xlsx, csv or both. Written without outside libraries; CSV cells that could run as spreadsheet formulas are neutralised; numbers keep the report's decimals in Excel; sheet names keep the distinguishing word; loading new data clears the previous results so downloads are never mislabelled
• Every PLS-SEM and CB-SEM report now opens its Structural Model section with a path diagram: constructs as ovals laid out by causal order, standardized coefficients with significance stars, dashed arrows for paths that are not significant, R² inside each outcome, arrows that skip a construct routed through the gaps between columns so they never cross another construct, labels placed so they never overlap, and higher-order loadings as grey measurement arrows. It is inline SVG that follows light and dark mode, with a text description for screen readers
• A model without structural paths now says so instead of showing an empty table

New in 0.7.0-alpha.5: Full ML Two-Level SEM and Blindfolding Q²

• PLS-SEM blindfolding Q² (Stone-Geisser, cross-validated redundancy, omission distance 7 or the next value that divides neither N nor the number of items) for every endogenous construct with ordinary PLS, in the Explained Variance table and the findings; a construct whose Q² cannot be computed gets a note instead of the whole column disappearing
• Two-level SEM (a cluster column) is now estimated by full maximum likelihood with both levels together, with exact chi-square tests, fit indices, SRMR at each level and standard errors at both levels (McDonald and Goldstein 1989; Muthen 1994)
• Level-specific tests: each level is tested with the other left unrestricted, so misfit can be traced to its level (Ryu and West 2009); a misspecified within model is detected at the within level in a test that runs on every build
• The separate-level (MUML) fits stay in the report as a cross-check, and are used if full ML cannot be estimated
• Robust estimation with few clusters (a damped step when the information matrix is singular); between-level RMSEA uses the number of clusters; cases dropped for missing values are reported

New in 0.7.0-alpha.4: True MLR (Robust FIML) and Robust Fit Indices

• New estimator MLR: robust maximum likelihood with Huber-White sandwich standard errors from casewise scores and observed information, and the Yuan-Bentler scaled chi-square (Yuan and Bentler 2000). It works with missing data (robust FIML). Auto Mode picks it when data are both incomplete and non-normal
• Breaking change: --estimator mlr now selects this MLR; scripts that used mlr for Satorra-Bentler should use mlm. Projects saved in the app before this release switch to MLM automatically and say so. Satorra-Bentler robust ML is now --estimator mlm (also in the app's estimator list)
• Checks on every build: on normal data the scaling is near 1 and standard errors match ML; on complete heavy-tailed data MLR agrees with Satorra-Bentler within 10%; on normal data with missing values it agrees with FIML
• Auto Mode falls back to FIML or ML (and says why) when a model has more indicators than the robust corrections handle (60 for MLR, 80 for MLM); if the MLR correction cannot be computed, the FIML results are kept with a warning. - Robust CFI, TLI and RMSEA (Brosseau-Liard and Savalei 2012, 2014) for MLM and for MLR with complete data (versions for incomplete data come later), from the model's and the baseline model's scaling corrections; fit is judged by these and the scaled chi-square, and the fit table also shows the uncorrected values
• Fix: models with a two-indicator factor that is identified only through its paths no longer fail as "singular" (paths now start away from zero); this also fixes the level-specific test of a misspecified two-level model

New in 0.7.0-alpha.3: Moderation and Higher-Order Constructs

• Moderation: write Y ~ X + M + X:M. PLS-SEM uses product indicators of standardized items (Chin et al. 2003); CB-SEM uses double-mean-centred matched-pair product indicators (Marsh et al. 2004; Lin et al. 2010). The report gives simple slopes at the moderator's mean and one standard deviation either side, and skips item reliability checks that do not apply to product terms

• Higher-order constructs: write Quality =~ Service + Product. Estimated with PLS-SEM by the repeated indicator approach (Becker et al. 2012; Sarstedt et al. 2019); links to the lower-order constructs are reported as measurement, not hypotheses. CB-SEM estimates it as a second-order factor (validated: with three first-order factors it reproduces the CFA chi-square exactly); PLS-SEM uses repeated indicators

• BCa bootstrap intervals: bias-corrected and accelerated intervals (Efron 1987) are reported beside the percentile intervals for PLS-SEM paths and indirect effects and for CB-SEM indirect effects; the acceleration comes from a jackknife over rows (grouped for large samples)

Also in 0.7.0-alpha.3: Measurement Invariance and Equality Constraints

• Multigroup CB-SEM with mean structures: a group column with a CFA (or CB-SEM chosen in Settings) tests configural, metric, scalar and strict invariance, with chi-square difference tests, the change in CFI and RMSEA (Cheung and Rensvold 2002; Chen 2007), estimates by group, and latent mean differences when scalar invariance holds
• Validated against the lavaan multiple-groups tutorial (Holzinger-Swineford by school): chi-squares 115.94, 124.12 and 164.04 against the published 115.85, 124.04 and 164.10, with the same conclusion (weak invariance supported, strong not)
• Equality constraints: parameters with the same label (a*x2 + a*x3) are held equal; works for loadings, paths and variances or covariances with ML, including the bootstraps
• Earlier levels guard later ones: once a level fails, higher levels are shown but not judged

New in 0.7.0-alpha.2: Validation, CFA and Safety

• Validated against published results: Bollen (1989) political democracy SEM matches lavaan exactly (chi-square 38.125 on 35 df, same paths and fit indices); the Holzinger-Swineford (1939) CFA matches CFI, TLI, RMSEA and SRMR to three decimals. The checks run on every build; see validation/README.md
• Confirmatory factor analysis: a model with constructs and no structural paths now runs as a CFA with CB-SEM
• Data files with text columns (for example Gender or School) now load; text columns can be the group or cluster column, and group results show the real names
• AI keys: changing a provider's address clears its saved key, so a key is never sent to a new host; addresses that are raw IP numbers or local network names are refused; the settings folder is readable by your account only (macOS and Linux)
• Autopilot now starts in Ask Before Each Change
• The app now uses the same port each start, so your autosaved project comes back after a restart
• The same option rules in the app and the command line, with plain-language errors
• Keyboard: Open Project and Choose File work with Enter or Space; the AI activity log is announced to screen readers
• Releases are published only after every platform builds, passes all tests and runs the demo, each with a SHA-256 checksum
• Wording: "MLR" is described as robust ML (Satorra-Bentler, also called MLM); no GPU is used or needed

AiSEM v0.7.0-alpha.6

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7.0-alpha.6: Path Diagram and Excel / CSV Export

• Export every report table to Excel (one sheet per table, numbers stored as numbers, bold headers) or one CSV file: the app has Excel and CSV buttons beside Save Report, and the command line takes --export xlsx, csv or both. Written without outside libraries; CSV cells that could run as spreadsheet formulas are neutralised; numbers keep the report's decimals in Excel; sheet names keep the distinguishing word; loading new data clears the previous results so downloads are never mislabelled
• Every PLS-SEM and CB-SEM report now opens its Structural Model section with a path diagram: constructs as ovals laid out by causal order, standardized coefficients with significance stars, dashed arrows for paths that are not significant, R² inside each outcome, arrows that skip a construct routed through the gaps between columns so they never cross another construct, labels placed so they never overlap, and higher-order loadings as grey measurement arrows. It is inline SVG that follows light and dark mode, with a text description for screen readers
• A model without structural paths now says so instead of showing an empty table

New in 0.7.0-alpha.5: Full ML Two-Level SEM and Blindfolding Q²

• PLS-SEM blindfolding Q² (Stone-Geisser, cross-validated redundancy, omission distance 7 or the next value that divides neither N nor the number of items) for every endogenous construct with ordinary PLS, in the Explained Variance table and the findings; a construct whose Q² cannot be computed gets a note instead of the whole column disappearing
• Two-level SEM (a cluster column) is now estimated by full maximum likelihood with both levels together, with exact chi-square tests, fit indices, SRMR at each level and standard errors at both levels (McDonald and Goldstein 1989; Muthen 1994)
• Level-specific tests: each level is tested with the other left unrestricted, so misfit can be traced to its level (Ryu and West 2009); a misspecified within model is detected at the within level in a test that runs on every build
• The separate-level (MUML) fits stay in the report as a cross-check, and are used if full ML cannot be estimated
• Robust estimation with few clusters (a damped step when the information matrix is singular); between-level RMSEA uses the number of clusters; cases dropped for missing values are reported

New in 0.7.0-alpha.4: True MLR (Robust FIML) and Robust Fit Indices

• New estimator MLR: robust maximum likelihood with Huber-White sandwich standard errors from casewise scores and observed information, and the Yuan-Bentler scaled chi-square (Yuan and Bentler 2000). It works with missing data (robust FIML). Auto Mode picks it when data are both incomplete and non-normal
• Breaking change: --estimator mlr now selects this MLR; scripts that used mlr for Satorra-Bentler should use mlm. Projects saved in the app before this release switch to MLM automatically and say so. Satorra-Bentler robust ML is now --estimator mlm (also in the app's estimator list)
• Checks on every build: on normal data the scaling is near 1 and standard errors match ML; on complete heavy-tailed data MLR agrees with Satorra-Bentler within 10%; on normal data with missing values it agrees with FIML
• Auto Mode falls back to FIML or ML (and says why) when a model has more indicators than the robust corrections handle (60 for MLR, 80 for MLM); if the MLR correction cannot be computed, the FIML results are kept with a warning. - Robust CFI, TLI and RMSEA (Brosseau-Liard and Savalei 2012, 2014) for MLM and for MLR with complete data (versions for incomplete data come later), from the model's and the baseline model's scaling corrections; fit is judged by these and the scaled chi-square, and the fit table also shows the uncorrected values
• Fix: models with a two-indicator factor that is identified only through its paths no longer fail as "singular" (paths now start away from zero); this also fixes the level-specific test of a misspecified two-level model

New in 0.7.0-alpha.3: Moderation and Higher-Order Constructs

• Moderation: write Y ~ X + M + X:M. PLS-SEM uses product indicators of standardized items (Chin et al. 2003); CB-SEM uses double-mean-centred matched-pair product indicators (Marsh et al. 2004; Lin et al. 2010). The report gives simple slopes at the moderator's mean and one standard deviation either side, and skips item reliability checks that do not apply to product terms

• Higher-order constructs: write Quality =~ Service + Product. Estimated with PLS-SEM by the repeated indicator approach (Becker et al. 2012; Sarstedt et al. 2019); links to the lower-order constructs are reported as measurement, not hypotheses. CB-SEM estimates it as a second-order factor (validated: with three first-order factors it reproduces the CFA chi-square exactly); PLS-SEM uses repeated indicators

• BCa bootstrap intervals: bias-corrected and accelerated intervals (Efron 1987) are reported beside the percentile intervals for PLS-SEM paths and indirect effects and for CB-SEM indirect effects; the acceleration comes from a jackknife over rows (grouped for large samples)

Also in 0.7.0-alpha.3: Measurement Invariance and Equality Constraints

• Multigroup CB-SEM with mean structures: a group column with a CFA (or CB-SEM chosen in Settings) tests configural, metric, scalar and strict invariance, with chi-square difference tests, the change in CFI and RMSEA (Cheung and Rensvold 2002; Chen 2007), estimates by group, and latent mean differences when scalar invariance holds
• Validated against the lavaan multiple-groups tutorial (Holzinger-Swineford by school): chi-squares 115.94, 124.12 and 164.04 against the published 115.85, 124.04 and 164.10, with the same conclusion (weak invariance supported, strong not)
• Equality constraints: parameters with the same label (a*x2 + a*x3) are held equal; works for loadings, paths and variances or covariances with ML, including the bootstraps
• Earlier levels guard later ones: once a level fails, higher levels are shown but not judged

New in 0.7.0-alpha.2: Validation, CFA and Safety

• Validated against published results: Bollen (1989) political democracy SEM matches lavaan exactly (chi-square 38.125 on 35 df, same paths and fit indices); the Holzinger-Swineford (1939) CFA matches CFI, TLI, RMSEA and SRMR to three decimals. The checks run on every build; see validation/README.md
• Confirmatory factor analysis: a model with constructs and no structural paths now runs as a CFA with CB-SEM
• Data files with text columns (for example Gender or School) now load; text columns can be the group or cluster column, and group results show the real names
• AI keys: changing a provider's address clears its saved key, so a key is never sent to a new host; addresses that are raw IP numbers or local network names are refused; the settings folder is readable by your account only (macOS and Linux)
• Autopilot now starts in Ask Before Each Change
• The app now uses the same port each start, so your autosaved project comes back after a restart
• The same option rules in the app and the command line, with plain-language errors
• Keyboard: Open Project and Choose File work with Enter or Space; the AI activity log is announced to screen readers
• Releases are published only after every platform builds, passes all tests and runs the demo, each with a SHA-256 checksum
• Wording: "MLR" is described as robust ML (Satorra-Bentler, also called MLM); no GPU is used or needed

AiSEM v0.7.0-alpha.5

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7.0-alpha.5: Full ML Two-Level SEM and Blindfolding Q²

• PLS-SEM blindfolding Q² (Stone-Geisser, cross-validated redundancy, omission distance 7 or the next value that divides neither N nor the number of items) for every endogenous construct with ordinary PLS, in the Explained Variance table and the findings; a construct whose Q² cannot be computed gets a note instead of the whole column disappearing
• Two-level SEM (a cluster column) is now estimated by full maximum likelihood with both levels together, with exact chi-square tests, fit indices, SRMR at each level and standard errors at both levels (McDonald and Goldstein 1989; Muthen 1994)
• Level-specific tests: each level is tested with the other left unrestricted, so misfit can be traced to its level (Ryu and West 2009); a misspecified within model is detected at the within level in a test that runs on every build
• The separate-level (MUML) fits stay in the report as a cross-check, and are used if full ML cannot be estimated
• Robust estimation with few clusters (a damped step when the information matrix is singular); between-level RMSEA uses the number of clusters; cases dropped for missing values are reported

New in 0.7.0-alpha.4: True MLR (Robust FIML) and Robust Fit Indices

• New estimator MLR: robust maximum likelihood with Huber-White sandwich standard errors from casewise scores and observed information, and the Yuan-Bentler scaled chi-square (Yuan and Bentler 2000). It works with missing data (robust FIML). Auto Mode picks it when data are both incomplete and non-normal
• Breaking change: --estimator mlr now selects this MLR; scripts that used mlr for Satorra-Bentler should use mlm. Projects saved in the app before this release switch to MLM automatically and say so. Satorra-Bentler robust ML is now --estimator mlm (also in the app's estimator list)
• Checks on every build: on normal data the scaling is near 1 and standard errors match ML; on complete heavy-tailed data MLR agrees with Satorra-Bentler within 10%; on normal data with missing values it agrees with FIML
• Auto Mode falls back to FIML or ML (and says why) when a model has more indicators than the robust corrections handle (60 for MLR, 80 for MLM); if the MLR correction cannot be computed, the FIML results are kept with a warning. - Robust CFI, TLI and RMSEA (Brosseau-Liard and Savalei 2012, 2014) for MLM and for MLR with complete data (versions for incomplete data come later), from the model's and the baseline model's scaling corrections; fit is judged by these and the scaled chi-square, and the fit table also shows the uncorrected values
• Fix: models with a two-indicator factor that is identified only through its paths no longer fail as "singular" (paths now start away from zero); this also fixes the level-specific test of a misspecified two-level model

New in 0.7.0-alpha.3: Moderation and Higher-Order Constructs

• Moderation: write Y ~ X + M + X:M. PLS-SEM uses product indicators of standardized items (Chin et al. 2003); CB-SEM uses double-mean-centred matched-pair product indicators (Marsh et al. 2004; Lin et al. 2010). The report gives simple slopes at the moderator's mean and one standard deviation either side, and skips item reliability checks that do not apply to product terms

• Higher-order constructs: write Quality =~ Service + Product. Estimated with PLS-SEM by the repeated indicator approach (Becker et al. 2012; Sarstedt et al. 2019); links to the lower-order constructs are reported as measurement, not hypotheses. CB-SEM estimates it as a second-order factor (validated: with three first-order factors it reproduces the CFA chi-square exactly); PLS-SEM uses repeated indicators

• BCa bootstrap intervals: bias-corrected and accelerated intervals (Efron 1987) are reported beside the percentile intervals for PLS-SEM paths and indirect effects and for CB-SEM indirect effects; the acceleration comes from a jackknife over rows (grouped for large samples)

Also in 0.7.0-alpha.3: Measurement Invariance and Equality Constraints

• Multigroup CB-SEM with mean structures: a group column with a CFA (or CB-SEM chosen in Settings) tests configural, metric, scalar and strict invariance, with chi-square difference tests, the change in CFI and RMSEA (Cheung and Rensvold 2002; Chen 2007), estimates by group, and latent mean differences when scalar invariance holds
• Validated against the lavaan multiple-groups tutorial (Holzinger-Swineford by school): chi-squares 115.94, 124.12 and 164.04 against the published 115.85, 124.04 and 164.10, with the same conclusion (weak invariance supported, strong not)
• Equality constraints: parameters with the same label (a*x2 + a*x3) are held equal; works for loadings, paths and variances or covariances with ML, including the bootstraps
• Earlier levels guard later ones: once a level fails, higher levels are shown but not judged

New in 0.7.0-alpha.2: Validation, CFA and Safety

• Validated against published results: Bollen (1989) political democracy SEM matches lavaan exactly (chi-square 38.125 on 35 df, same paths and fit indices); the Holzinger-Swineford (1939) CFA matches CFI, TLI, RMSEA and SRMR to three decimals. The checks run on every build; see validation/README.md
• Confirmatory factor analysis: a model with constructs and no structural paths now runs as a CFA with CB-SEM
• Data files with text columns (for example Gender or School) now load; text columns can be the group or cluster column, and group results show the real names
• AI keys: changing a provider's address clears its saved key, so a key is never sent to a new host; addresses that are raw IP numbers or local network names are refused; the settings folder is readable by your account only (macOS and Linux)
• Autopilot now starts in Ask Before Each Change
• The app now uses the same port each start, so your autosaved project comes back after a restart
• The same option rules in the app and the command line, with plain-language errors
• Keyboard: Open Project and Choose File work with Enter or Space; the AI activity log is announced to screen readers
• Releases are published only after every platform builds, passes all tests and runs the demo, each with a SHA-256 checksum
• Wording: "MLR" is described as robust ML (Satorra-Bentler, also called MLM); no GPU is used or needed

AiSEM v0.7.0-alpha.4

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7.0-alpha.4: True MLR (Robust FIML) and Robust Fit Indices

• New estimator MLR: robust maximum likelihood with Huber-White sandwich standard errors from casewise scores and observed information, and the Yuan-Bentler scaled chi-square (Yuan and Bentler 2000). It works with missing data (robust FIML). Auto Mode picks it when data are both incomplete and non-normal
• Breaking change: --estimator mlr now selects this MLR; scripts that used mlr for Satorra-Bentler should use mlm. Projects saved in the app before this release switch to MLM automatically and say so. Satorra-Bentler robust ML is now --estimator mlm (also in the app's estimator list)
• Checks on every build: on normal data the scaling is near 1 and standard errors match ML; on complete heavy-tailed data MLR agrees with Satorra-Bentler within 10%; on normal data with missing values it agrees with FIML
• Auto Mode falls back to FIML or ML (and says why) when a model has more indicators than the robust corrections handle (60 for MLR, 80 for MLM); if the MLR correction cannot be computed, the FIML results are kept with a warning. - Robust CFI, TLI and RMSEA (Brosseau-Liard and Savalei 2012, 2014) for MLM and for MLR with complete data (versions for incomplete data come later), from the model's and the baseline model's scaling corrections; fit is judged by these and the scaled chi-square, and the fit table also shows the uncorrected values
• Fix: models with a two-indicator factor that is identified only through its paths no longer fail as "singular" (paths now start away from zero); this also fixes the level-specific test of a misspecified two-level model

New in 0.7.0-alpha.3: Moderation and Higher-Order Constructs

• Moderation: write Y ~ X + M + X:M. PLS-SEM uses product indicators of standardized items (Chin et al. 2003); CB-SEM uses double-mean-centred matched-pair product indicators (Marsh et al. 2004; Lin et al. 2010). The report gives simple slopes at the moderator's mean and one standard deviation either side, and skips item reliability checks that do not apply to product terms

• Higher-order constructs: write Quality =~ Service + Product. Estimated with PLS-SEM by the repeated indicator approach (Becker et al. 2012; Sarstedt et al. 2019); links to the lower-order constructs are reported as measurement, not hypotheses. CB-SEM estimates it as a second-order factor (validated: with three first-order factors it reproduces the CFA chi-square exactly); PLS-SEM uses repeated indicators

• BCa bootstrap intervals: bias-corrected and accelerated intervals (Efron 1987) are reported beside the percentile intervals for PLS-SEM paths and indirect effects and for CB-SEM indirect effects; the acceleration comes from a jackknife over rows (grouped for large samples)

Also in 0.7.0-alpha.3: Measurement Invariance and Equality Constraints

• Multigroup CB-SEM with mean structures: a group column with a CFA (or CB-SEM chosen in Settings) tests configural, metric, scalar and strict invariance, with chi-square difference tests, the change in CFI and RMSEA (Cheung and Rensvold 2002; Chen 2007), estimates by group, and latent mean differences when scalar invariance holds
• Validated against the lavaan multiple-groups tutorial (Holzinger-Swineford by school): chi-squares 115.94, 124.12 and 164.04 against the published 115.85, 124.04 and 164.10, with the same conclusion (weak invariance supported, strong not)
• Equality constraints: parameters with the same label (a*x2 + a*x3) are held equal; works for loadings, paths and variances or covariances with ML, including the bootstraps
• Earlier levels guard later ones: once a level fails, higher levels are shown but not judged

New in 0.7.0-alpha.2: Validation, CFA and Safety

• Validated against published results: Bollen (1989) political democracy SEM matches lavaan exactly (chi-square 38.125 on 35 df, same paths and fit indices); the Holzinger-Swineford (1939) CFA matches CFI, TLI, RMSEA and SRMR to three decimals. The checks run on every build; see validation/README.md
• Confirmatory factor analysis: a model with constructs and no structural paths now runs as a CFA with CB-SEM
• Data files with text columns (for example Gender or School) now load; text columns can be the group or cluster column, and group results show the real names
• AI keys: changing a provider's address clears its saved key, so a key is never sent to a new host; addresses that are raw IP numbers or local network names are refused; the settings folder is readable by your account only (macOS and Linux)
• Autopilot now starts in Ask Before Each Change
• The app now uses the same port each start, so your autosaved project comes back after a restart
• The same option rules in the app and the command line, with plain-language errors
• Keyboard: Open Project and Choose File work with Enter or Space; the AI activity log is announced to screen readers
• Releases are published only after every platform builds, passes all tests and runs the demo, each with a SHA-256 checksum
• Wording: "MLR" is described as robust ML (Satorra-Bentler, also called MLM); no GPU is used or needed

AiSEM v0.7.0-alpha.3

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7.0-alpha.3: Moderation and Higher-Order Constructs

• Moderation: write Y ~ X + M + X:M. PLS-SEM uses product indicators of standardized items (Chin et al. 2003); CB-SEM uses double-mean-centred matched-pair product indicators (Marsh et al. 2004; Lin et al. 2010). The report gives simple slopes at the moderator's mean and one standard deviation either side, and skips item reliability checks that do not apply to product terms

• Higher-order constructs: write Quality =~ Service + Product. Estimated with PLS-SEM by the repeated indicator approach (Becker et al. 2012; Sarstedt et al. 2019); links to the lower-order constructs are reported as measurement, not hypotheses. CB-SEM estimates it as a second-order factor (validated: with three first-order factors it reproduces the CFA chi-square exactly); PLS-SEM uses repeated indicators

• BCa bootstrap intervals: bias-corrected and accelerated intervals (Efron 1987) are reported beside the percentile intervals for PLS-SEM paths and indirect effects and for CB-SEM indirect effects; the acceleration comes from a jackknife over rows (grouped for large samples)

Also in 0.7.0-alpha.3: Measurement Invariance and Equality Constraints

• Multigroup CB-SEM with mean structures: a group column with a CFA (or CB-SEM chosen in Settings) tests configural, metric, scalar and strict invariance, with chi-square difference tests, the change in CFI and RMSEA (Cheung and Rensvold 2002; Chen 2007), estimates by group, and latent mean differences when scalar invariance holds
• Validated against the lavaan multiple-groups tutorial (Holzinger-Swineford by school): chi-squares 115.94, 124.12 and 164.04 against the published 115.85, 124.04 and 164.10, with the same conclusion (weak invariance supported, strong not)
• Equality constraints: parameters with the same label (a*x2 + a*x3) are held equal; works for loadings, paths and variances or covariances with ML, including the bootstraps
• Earlier levels guard later ones: once a level fails, higher levels are shown but not judged

New in 0.7.0-alpha.2: Validation, CFA and Safety

• Validated against published results: Bollen (1989) political democracy SEM matches lavaan exactly (chi-square 38.125 on 35 df, same paths and fit indices); the Holzinger-Swineford (1939) CFA matches CFI, TLI, RMSEA and SRMR to three decimals. The checks run on every build; see validation/README.md
• Confirmatory factor analysis: a model with constructs and no structural paths now runs as a CFA with CB-SEM
• Data files with text columns (for example Gender or School) now load; text columns can be the group or cluster column, and group results show the real names
• AI keys: changing a provider's address clears its saved key, so a key is never sent to a new host; addresses that are raw IP numbers or local network names are refused; the settings folder is readable by your account only (macOS and Linux)
• Autopilot now starts in Ask Before Each Change
• The app now uses the same port each start, so your autosaved project comes back after a restart
• The same option rules in the app and the command line, with plain-language errors
• Keyboard: Open Project and Choose File work with Enter or Space; the AI activity log is announced to screen readers
• Releases are published only after every platform builds, passes all tests and runs the demo, each with a SHA-256 checksum
• Wording: "MLR" is described as robust ML (Satorra-Bentler, also called MLM); no GPU is used or needed

AiSEM v0.7.0-alpha.2

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7.0-alpha.2: Validation, CFA and Safety

• Validated against published results: Bollen (1989) political democracy SEM matches lavaan exactly (chi-square 38.125 on 35 df, same paths and fit indices); the Holzinger-Swineford (1939) CFA matches CFI, TLI, RMSEA and SRMR to three decimals. The checks run on every build; see validation/README.md
• Confirmatory factor analysis: a model with constructs and no structural paths now runs as a CFA with CB-SEM
• Data files with text columns (for example Gender or School) now load; text columns can be the group or cluster column, and group results show the real names
• AI keys: changing a provider's address clears its saved key, so a key is never sent to a new host; addresses that are raw IP numbers or local network names are refused; the settings folder is readable by your account only (macOS and Linux)
• Autopilot now starts in Ask Before Each Change
• The app now uses the same port each start, so your autosaved project comes back after a restart
• The same option rules in the app and the command line, with plain-language errors
• Keyboard: Open Project and Choose File work with Enter or Space; the AI activity log is announced to screen readers
• Releases are published only after every platform builds, passes all tests and runs the demo, each with a SHA-256 checksum
• Wording: "MLR" is described as robust ML (Satorra-Bentler, also called MLM); no GPU is used or needed

AiSEM 0.7.0-alpha.10

New in 0.7.0-alpha.10: CVPAT (cross-validated predictive ability test) in PLSpredict, comparing the model's out-of-sample loss with the indicator average and the linear-model benchmark by one-sided paired t tests. Full notes: https://github.com/yeerock/aisem/blob/main/RELEASE_NOTES.md

AiSEM 0.7.0-alpha.1

AiSEM 0.7: AI Native Structural Equation Modeling Software, free CB-SEM and PLS-SEM written from scratch and backed by published research. Free to download and use, including for business (see LICENSE). AiSEM was called OpenSEM up to version 0.6.

The AiSEM app. Double-click aisem (or run aisem) and the app opens in your web browser. Everything runs on your own computer; only what you choose to share goes to the AI provider you pick.

New in 0.7: AI Copilot

• Autopilot runs the whole study: it reads your column summaries, drafts the model, chooses the methods, runs the analysis, reviews the results against published guidelines, improves the model and writes up the findings, with a full activity log
• Full Auto, or Ask Before Each Change; choose which steps the AI does and how many review rounds
• AiSEM checks every AI suggestion before use: models are parsed and matched to your columns (the AI repairs rejected models), settings are checked against the allowed choices
• Number verification: every number in an AI write-up or answer is matched against the computed report, and any number not found is flagged
• AI buttons on each step: Draft with AI, Recommend Settings with AI, AI Review, AI Write-Up, plus Ask the AI about your data and results
• Providers: OpenAI (ChatGPT), Anthropic (Claude), Google Gemini, DeepSeek, Moonshot (Kimi), Z.ai and Zhipu (GLM), xAI (Grok), Meta (Muse), Alibaba Cloud Model Studio (Qwen), Mistral, Groq, OpenRouter, any OpenAI-compatible service, and free private models on your own computer with Ollama or LM Studio
• Free options: Gemini, GLM Flash, Mistral and Groq free tiers, OpenRouter free models, and local models
• The model list is loaded live from each provider, so new models appear without an update
• Privacy: choose Column Names Only, Column Summaries (default) or Summaries and First 15 Rows; keys stay on this computer, are masked on screen, never appear in reports or project files, and are saved (owner-only file) only if you tick Remember on This Computer

New in 0.7: Advanced Models

• Bayesian SEM: Gibbs sampler with 4 chains on all CPU cores; posterior means, medians, credible intervals, one-tailed p, standardized estimates, indirect effects, R-squared, rank-normalized R-hat and effective sample size, posterior predictive p value and DIC (--estimator bayes)
• Two-level SEM for clustered data: intraclass correlations, design effects, and the model fitted within and between clusters (--cluster <column>)
• Latent class and latent profile analysis (categorical, continuous or mixed items, missing values allowed): 1 to K classes with AIC, BIC, sample-size adjusted BIC, CAIC, entropy, bootstrap likelihood ratio test, class sizes, average posterior probabilities and item profiles (--classes <K>)
• FIMIX-PLS hidden segments in every PLS-SEM run: AIC, AIC3, AIC4, BIC, CAIC, HQ, MDL5, EN and NEC for 1 to 4 segments, a suggested segment count, and segment-specific paths and R-squared (--segments <K>)
• Necessary condition analysis in every PLS-SEM run: CE-FDH and CR-FDH ceilings, effect size d, permutation test, accuracy, inefficiencies and bottleneck tables (--nca-perms <N>)
• Rename: the program is now aisem (was opensem); OpenSEM project files and saved work still open

New in 0.6: Ordinal Data

• WLSMV for Likert-type and other ordered-category items: thresholds, polychoric correlations with their asymptotic covariance matrix, diagonally weighted least squares, robust standard errors and the mean-and-variance adjusted chi-square, with CFI, TLI and RMSEA from the adjusted statistics
• Auto Mode picks WLSMV when every item has 2 to 4 categories and no values are missing; choose it yourself with --estimator wlsmv or Ordinal Items (WLSMV) in the app
• Polychoric correlations run on every CPU core

New in 0.5

• More CB-SEM estimators: generalized least squares (GLS), unweighted least squares (ULS) and asymptotically distribution-free (ADF / WLS), in the app and with --estimator gls|uls|adf
• Monte Carlo power analysis for every structural path at your sample size, and the sample size that gives every path 80% power
• Fix: no crash when an estimator has no chi-square test (ULS)

New in 0.4: The AiSEM App

• Four guided steps: Data, Model, Run, Results
• Load a CSV or TSV file by drag and drop, with a preview and missing-value count, or try the demo data
• Visual path diagram editor: add constructs, tick their items, connect paths by clicking, drag to arrange, reflective or formative; loops are refused
• Suggest Model groups items by name (T1, T2, T3 become construct T); Tidy Layout arranges the diagram
• Model text stays in sync with the diagram and can be edited directly
• Run with Auto Mode or choose method, estimator, resamples, confidence level, seed, group comparison and IPMA scale
• The full report opens in the app; Save Report keeps it as an HTML file
• Save Project and Open Project keep data, model and settings in one file
• Security: the app listens only on this computer, needs a secret per-session link, rejects other websites and uses a strict content security policy

New in 0.3: PLS-SEM Add-Ons

• PLSpredict: 10-fold cross-validation repeated 10 times, RMSE and MAE against a linear-model benchmark, Q2_predict, and the predictive-power verdict
• Multigroup analysis with --group <column>: permutation test of path differences (PLS-MGA) for every pair of groups
• MICOM measurement invariance: compositional invariance, equal means and equal variances
• Importance-performance map analysis (IPMA) for each final outcome construct

New in 0.2: CB-SEM

• Maximum likelihood estimation (Fisher scoring) with standard errors and the completely standardized solution
• Robust ML (Satorra-Bentler scaled chi-square and robust standard errors) for non-normal data
• FIML for missing data, compared against the EM saturated model
• Chi-square, CFI, TLI, RMSEA with 90% confidence interval and test of close fit, SRMR, AIC, BIC
• Modification indices with expected parameter change
• Bollen-Stine bootstrap model test and Mardia normality tests
• Residual covariances in the model syntax: x1 ~~ x2

New in 0.2: Mediation

• Specific and total indirect effects for CB-SEM and PLS-SEM, with bootstrap intervals (delta method under FIML)
• Mediation type for each indirect effect (complementary, competitive, indirect-only, direct-only, no effect)

Auto Mode

• Chooses CB-SEM for reflective models and PLS-SEM when a construct is formative
• Chooses ML, robust ML or FIML from the data, and falls back to PLS-SEM with an explanation if CB-SEM cannot be estimated

Still Included

• PLS path modelling and consistent PLS, reliability, AVE, HTMT, VIF, f squared, bootstrap on all CPU cores
• Self-contained HTML report with APA-style tables, a methods paragraph and the full reference list

Quick Start

1. Download the file for your system below and unzip it.
2. Double-click aisem (Windows: aisem.exe). The app opens in your browser; keep the small window open while you work.
3. Command line also works: aisem analyze --data mydata.csv --model mymodel.txt

Run aisem --help for the model syntax and all options. Each download has a .sha256 checksum. The macOS build is not notarised yet; if macOS blocks it, right-click the file and choose Open.