Intraspinal anesthetic needle puncture virtuality and reality combination teaching system

Abstract

The invention discloses an intraspinal anesthetic needle puncture virtuality and reality combined teaching system, which relates to the technical field of medical teaching evaluation, and comprises the following steps of: 1, compiling an index definition library, a Chinese word matrix and a report template into a rule net version through unit and dimensional reasoning and equation and interval checking, and carrying out property test; 2, aligning and normalizing the multi-source sequence and the event according to a conversion chain, and packaging into a fact package containing a source, a formula and a unit; 3, generating a rendering plan based on the fact package, preposing word use and display precision, and pre-scoring; 4, constrained generation and reverse checking are carried out, minimum correction is carried out under the condition that a fact numerical value is not changed, and a consistency text and a final consistency score are output; 5, signing and issuing a gating and embedding a pedigree and a signature; in case of data, threshold or template change, calculating a change influence domain, re-rendering the influenced sentences, and outputting differences and auditing; consistency, traceability and verifiability of facts and characters are realized.

Description

Technical Field

[0001] This invention relates to the field of medical teaching assessment technology, specifically to a teaching system that combines virtual and real methods for spinal anesthesia needle puncture. Background Technology

[0002] In training scenarios for spinal anesthesia procedures based on a combination of virtual and real methods, the teaching system typically consists of a physical phantom, an augmented / mixed reality display, ultrasound equipment, and multi-source sensing units. It collects data in real time, including force-displacement-attitude time series, ultrasound frame events, resistance disappearance determination, and start and end times. After training, the platform generates line graphs, box plots, histograms, and tables, providing indicators such as success rate, number of punctures, total duration, angle of incidence, withdrawal, and overshoot. It also outputs textual summaries for teaching records, quality assessment, and examination certification.

[0003] Existing processes generally employ template-driven fixed statements or generalized data-to-text modules, supplemented by manual editing and formatting. However, these output links and indicator calculation chains lack engineering-based consistency constraints: First, units and dimensions are often arbitrarily switched based on display requirements, without compatibility reasoning or conversion memorization, easily leading to distortion of sum-difference inequalities due to rounding or unit substitution; Second, statistical elements such as significance tests, confidence intervals, and effect sizes are not strictly bound to terminology, resulting in the misuse of terms such as "significant" and "better than"; Third, raw numbers are allowed to be written directly in the text without being bound to a unique source of fact; Fourth, when data is updated or thresholds are changed, there is a lack of influence domain identification, often replacing full-link verification with local manual modifications; Fifth, the reporting layer lacks a traceable source-formula-unit-timestamp hierarchy, making on-site verification impossible for external auditors.

[0004] The underlying mechanism is as follows: indicators are derived and aggregated multiple times from the original signal before entering the template. The process lacks pre-verification and publication thresholds, amplifying errors during the calculation, presentation, and expression stages, making timely detection difficult. Once erroneous descriptions are archived or published, they can lead to scoring biases, misallocation of training resources, difficulties in obtaining evidence for disputes, and compliance risks. In situations where multi-site teaching and periodic examinations run concurrently, template versions, pass rates, and statistical standards are frequently adjusted with courses or semesters. Personnel turnover results in incomplete rule transmission, further amplifying the probability of these inconsistencies. This is especially true when charts are displayed in millimeters while text is changed to centimeters, or when success rates are rounded to integers while numerators and denominators are kept to units, leading to apparent consistency but internal contradictions—consistent charts but inconsistent text.

[0005] Therefore, the current technical problem is that, in the above training and evaluation scenarios, there is a lack of an end-to-end consistency constraint and verification mechanism covering units and dimensions, formula relationships, statistical terminology, text generation, and release and issuance, so as to ensure that the facts expressed in the charts / tables are presented uniquely, accurately, and traceably in the textual description. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides a virtual-real integrated teaching system for spinal anesthesia needle puncture. It aligns and normalizes multi-source sequences and events according to a conversion chain, encapsulating them into fact packages containing sources, formulas, and units. Based on these fact packages, a rendering plan is generated, pre-selecting words and displaying accuracy, and providing a pre-scoring. Constrained generation and reverse verification are performed, with minimal corrections without altering factual values, outputting consistent text and a final consistency score. Gating is issued and embedded with genealogy and signatures. Upon data, threshold, or template changes, the impact domain of the change is recalculated, and affected sentences are re-rendered, outputting differences and audit results. This achieves consistency and traceability between facts and text, solving the technical problems described in the background art.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A virtual-real combined teaching system for spinal anesthesia needle puncture includes: performing constraint compilation on an indicator definition library, a Chinese term matrix, and a report template containing data reference markers; completing unit dimension reasoning to determine the conversion chain; solidifying the constraints into a rule network version; and generating a template package through boundary pre-drill.

[0011] Based on the rule network version, the original time series and events are aligned, the units are normalized and the indicators are calculated according to the conversion chain, and the source, formula, unit, rounding and timestamp are written into the genealogy and encapsulated into a fact package with version and verification code.

[0012] Read the fact package and template package, complete the pre-validation on the rule network version, form a word plan based on predicate permission and quantile interval, determine the display unit and reserved number of bits for each data reference mark, and generate a rendering plan with pre-scoring;

[0013] Text is generated using data reference tags according to the rendering plan, and numerical values ​​and units are extracted and backfilled to the rule network for verification. Under the premise of not changing the actual numerical values, word replacement, unit switching and bit adjustment are performed with minimal changes, and consistent text and final consistency score are output.

[0014] Issuance gating is performed based on the final consistency score, and the genealogy and signature are embedded in the metadata through text. When data, thresholds or templates change, the impact domain of the change is calculated according to the rule network, and only the affected sentences are re-rendered to generate difference comparisons and audit logs.

[0015] Furthermore, unit and dimension reasoning is implemented for the indicator definition library to establish a conversion chain; the equation relations, sum-difference consistency, interval restrictions, conditional implications, and statistical terminology gating are compiled into a rule network version and semantically bound to the report template containing data reference tags.

[0016] Furthermore, a boundary set is constructed that includes cases where the number of covered samples is zero or one, confidence intervals cross zero, rounding results in unequal summation, unit mismatch, and significance threshold. The report template is then pre-run for each item, and when full coverage is achieved, a template package is output and the accuracy constraints are solidified and displayed.

[0017] Furthermore, the original time series is aligned with the start and end of puncture as anchor points, and the trajectory is reconstructed using shape-preserving interpolation. The time set of events generated by the disappearance of resistance and interlayer crossing is determined based on the curvature peaks and valleys of the force and displacement curves, and the same name reference is maintained with the rule network version.

[0018] Furthermore, unit normalization is completed according to the conversion chain, and the calculated value and the displayed value are saved separately, and the display precision and rounding rules are recorded; when encapsulating the fact package, the source identifier, formula used, unit and dimension, rounding strategy, timestamp and responsible person are written, and the version and verification code are generated.

[0019] Furthermore, the fact package and template package are read, predicate permissions are calculated on the rule network version, and a word selection table is formed by combining semantic direction; the display unit and the number of bits to be retained are determined for each data reference mark according to the display precision, the sum and difference are checked for consistency, and a display strategy table is generated and written into the rendering plan.

[0020] Furthermore, data reference tagging is applied to the template syntax skeleton to check the integrity of references, consistency of units and dimensions, and consistency of sums and differences. Pre-scoring is calculated and a problem list is generated. At the same time, constraint triggering paths are output to locate the corresponding rule entries and syntax positions.

[0021] Furthermore, constrained generation is performed according to the rendering plan, prohibiting raw numbers from entering the text;

[0022] The generated text is reverse-extracted to obtain triples of values, units, and predicates. These triples are then fed back into the rule network version for judgment and to form a set of violations, which serves as input for subsequent corrections.

[0023] Furthermore, without changing the calculated value of the fact package, the minimum modifications are performed in sequence, including word replacement, unit display switching, fine-tuning of retained digits, and minor syntactic rewriting. After the modifications are made, the final consistency score and correction list are checked and updated repeatedly.

[0024] Furthermore, a tiered issuance gating system is implemented based on the final consistency score and the number of blocking items; for approved texts, the fact package signature, rule network version, template version, issuance timestamp, and issuer identifier are written into the document metadata, and a rendering plan summary is saved for offline verification.

[0025] Furthermore, when data, thresholds, or templates change, the changes are mapped to a set of change sources, the change impact domain is calculated on the rule network dependency graph, and only the affected sentences are re-rendered. The re-rendering order is determined according to reachability depth, syntactic length, and constraint importance.

[0026] Furthermore, while generating a comparison of the differences between the old and new texts, the location of the differences and the corresponding reasons for the changes are recorded and added to the audit log; the audit log, template version, rule network version, data version, correction records, and scoring trajectory are stored in chronological order.

[0027] Furthermore, when generating the final report, the source identifier, formula used, unit and dimension, rounding rules, and conversion chain summary corresponding to each data reference mark are written into the document metadata and endnote area, and stored together with the document signature.

[0028] (III) Beneficial Effects

[0029] This invention provides a teaching system combining virtual and real methods for spinal anesthesia needle puncture, which has the following beneficial effects:

[0030] The constraint compilation and template pre-approval process unifies the compilation of unit and dimension reasoning, equations and intervals, sum and difference consistency, and statistical terminology gating into a rule network version. It also covers boundary scenarios with property testing, so that the template completes constraint solidification and semantic binding before it is put on the shelf, avoiding the inheritance of hidden errors in subsequent stages, and forming a stable entry point and clear interpretation.

[0031] Data normalization and genealogy are packaged on a unified timeline to align the original time series and events. Unit normalization is completed according to the conversion chain to distinguish between calculated values ​​and displayed values. The source, formula used, unit and dimension, rounding strategy, timestamp and responsible person are written into the fact package to ensure the uniqueness and traceability of the facts, so that subsequent rendering and issuance are based on the same fact source.

[0032] Rendering planning and pre-verification are based on the fact package and rule network versions. First, unified decisions are made on word planning, display units and reserved digits. The integrity and sum-difference consistency of data reference tags are checked. A rendering plan and problem list containing word selection table and display strategy table are output to reduce structural modifications and cross-page rework after generation from the source.

[0033] Constrained generation and minimal modification correction only allow data reference tags to enter the text. By reverse extraction, the generated text is fed back into the rule network to locate violations. Without changing the factual values, word replacement, unit display switching and retention bit fine-tuning are performed in sequence with minimal changes, quickly converging into consistent text, significantly reducing the burden of manual proofreading and maintaining the stability of the factual layer.

[0034] The release governance and incremental updates are implemented in a tiered manner based on the final consistency score. The fact package signature, rule network version and template version are embedded in the document metadata, and the rendering plan summary is recorded. When data, thresholds or templates change, the change impact domain is calculated according to the rule network. Only the affected sentences are re-rendered and the difference and reason explanation are generated to ensure that the issuance process is reliable and the update process is controllable. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the teaching system for combining virtual and real spinal anesthesia needle puncture according to the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figure 1 This invention provides a teaching system combining virtual and real methods for spinal anesthesia needle puncture, comprising:

[0038] Step 1: Before any report is generated, compile the implicit specifications involving units and dimensions, algebraic relationships, interval restrictions, and statistical terms into an executable rule network. By using pre-release property testing, potential errors are intercepted at the entry point, thus laying a deterministic foundation for subsequent fact packages, rendering plans, minimum corrections, and trusted issuance.

[0039] The teaching evaluation of spinal anesthesia needle puncture involves multiple indicators such as success rate, number of punctures, total duration, incident angle deviation, resistance loss event, and cerebrospinal fluid pressure, with units spanning seconds, millimeters, degrees, and centimeters / hydraulic columns. The wording also involves terms such as "significant improvement" and "no significant difference." If the units are not aligned, the algebras are not consistent, and the statistical terminology is not gated at the entry stage, any subsequent fact-text conversion will introduce hidden errors.

[0040] Using the constrained intermediate representation as a hub, unit and dimensional reasoning is performed first, and then equations and interval relationships are incorporated into the same rule network. Subsequently, property tests covering extremes and boundaries drive template pre-running. If any condition is not met, an actionable repair list is generated, and the template is not uploaded. The entry point is also the gatekeeper, ensuring that subsequent steps do not have to clean up messes caused by unit incompatibility, sum and difference imbalance, or misuse of statistical terminology. Furthermore, the versioned rule network ensures consistent standards across sites and time periods.

[0041] Specifically, the unit and dimension verification and algebraic-interval-implication relationship are merged and compiled into an executable rule network, ensuring that any subsequent data and templates are interpreted solely by this rule network.

[0042] In spinal anesthesia training, the raw data acquisition often comes from multiple sources: time from a timer, pose from optical tracking, pressure from a pressure sensor, and angle from an angle encoder.

[0043] First, a unit mapping is established for each indicator to be used, and its compatibility is verified using a dimensional mapping matrix-unit exponent vector method. For compatible indicators, the conversion chain from the source unit to the standard unit is recorded as the sole usage of the subsequent fact package; for incompatible indicators, they are marked as fatal errors. To avoid abstract expression, during compilation, indicators-units-dimensions are implemented one by one in the order of indicator-unit-dimension, without skipping steps or merging. Among them, the dimensional compatibility determination is as follows:

[0044]

[0045] Where: Dimensional verification function : Values or , indicating whether they are compatible, serves to provide a compatibility determination; unit exponent vector : Represents the index combination of the source unit and the standard unit of a certain indicator in the base units such as mass, length, time, angle, and pressure; Dimensional mapping matrix A matrix that maps unit exponent vectors to dimensional vectors, where elements are integers, rows represent base dimensions, and columns represent the positions of unit exponents; zero vector. : Dimensional mapping matrix The rows with the same dimension represent completely equal units;

[0046] Furthermore, a unit-dimension conversion table is first constructed to express commonly used units such as centimeters of water column, kilopascals, millimeters, degrees, and seconds using rational exponents. For each indicator, a unit exponent vector of the source unit and a unit exponent vector of the standard unit are generated, and these are substituted into the dimension verification function. Generate conversion chains for compatible components, such as a stepwise mapping from centimeters of water column to Pascals to standard centimeters of water column; immediately locate and revert to revisions for incompatible components in the template (e.g., misusing degrees for time fields).

[0047] By using explicit dimensional verification functions and conversion chains, all metrics are forced to align to the same dimensional semantics during compilation, eliminating potential issues such as mixing angles and lengths later on. Once the conversion chains are explicitly defined, they can be directly referenced in the next step of creating the fact package, avoiding redundant judgments.

[0048] After passing the dimensional verification, the constraints directly related to teaching evaluation are uniformly compiled into a rule network, including consistency of equations and sums (e.g., the total duration equals the sum of the durations of each stage), interval restrictions (e.g., passing through the upper and lower bounds of the line), and conditional implications (e.g., when the significance test probability is greater than the threshold, the term "significantly relevant" cannot be used).

[0049] Furthermore, to prevent rounding from compromising the consistency of sums and differences, a rounding operator—the residual measure—is introduced for checking. Specifically, the sum-difference consistency residual is:

[0050]

[0051] Where: Consistency of sum and difference residuals : It measures the deviation between the rounded sum and the sum of the components, and takes the value of a non-negative real number. Its function is to trigger template revision or precision adjustment.

[0052] Integer operator By reserved number of digits Operators for rounding or banker's rounding.

[0053]

[0054] Integer operator For real numbers reserve Even rounding of decimal places; even rounding function : just in time Round the nearest even number to the nearest even number; retain the number of digits. : Non-negative integer. ; .

[0055] Number of items : The number of sub-items, with values ​​taking positive integers; index components. : , No. Each component has its original value, which is a real number; absolute value sign. Take the non-negative difference value.

[0056] For each sum-difference agreement, first retain the number of digits recommended by the rules network. Execute the rounding operator Calculate the consistency residuals of sum and difference. If the residual is greater than zero, the display precision will be adjusted first. If the result is still not zero, the corresponding sentence in the template is rejected, requiring the use of interval-based display or explicit annotation of rounding rules. Interval constraints and conditional implications are expressed using the same rule network, resulting in a unified output of an executable check sequence. This moves arithmetic consistency to the template entry point, avoiding post-rendering adjustments at the sentence level; it also links display precision with equation consistency, reducing the need for manually setting the number of decimal places, allowing subsequent rendering plans to directly reuse the same rules. .

[0057] As an example: In a certain teaching center, the subject leader imported the indicator definition library and report templates into the system: for example, success rate, number of punctures, total duration, incident angle deviation, resistance disappearance event, cerebrospinal fluid pressure, etc. are used as indicators, and data reference markers are used as placeholders in the templates.

[0058] The system first lists the source and standard units for all fields on the left side of the screen. When the person in charge writes the unit for cerebrospinal fluid pressure as millimeters of mercury (mmHg) while the standard unit is centimeters of water (cmH), the system displays a visible prompt indicating that the two units are dimensionally compatible and provides the conversion chain from mmHg to Pascal (Pa) to cmH. When the person in charge mistakenly enters millimeters as the unit for the incident angle deviation, the system immediately highlights the field in red, displays that the angle and length units are incompatible, and prohibits proceeding to the next step. After the person in charge increases the number of decimal places for the total duration minus the duration of each stage on the right side, the consistency residuals of the sum and difference disappear, and the system allows the user to proceed to the next step.

[0059] Furthermore, after compilation, the template must be pre-run line by line using test samples covering extremes and boundaries to ensure that the template can satisfy the rule network under any combination of conditions; if there are any non-compliance, an actionable list of fixes should be output and the template should not be uploaded.

[0060] The key to property testing lies in sample coverage. First, a boundary set is constructed, covering scenarios such as zero or one sample size, confidence intervals crossing zero, significance thresholds, unit mismatch triggers, and sum / difference thresholds. Then, a pre-test sample set is generated, and finally, the coverage rate is calculated. Only after sufficient coverage is achieved does the subsequent licensing decision proceed. The coverage rate is constructed as follows:

[0061]

[0062] Where: Coverage rate : Value in Between these, the function is to measure whether the property test covers all boundary cases, if and only if Time indicates complete coverage; pre-simulation sample set : A sample set automatically generated by the system according to rules, with elements being parameterized combinations of indicators; boundary set : A predefined set of extreme and boundary cases; size symbol of the set : Returns the number of elements contained in the collection;

[0063] Furthermore, based on the interval endpoints, critical statements, and unit compatibility boundaries appearing in the rule network, the boundary set is automatically enumerated; then, combined with the data reference tags used in the template, a pre-run sample set is generated, and each rule is run sequentially to count the number of intersections; if the coverage rate... Less than The system will automatically prompt for missing boundary scenarios and guide you to create corresponding sentence structures or supplementary words.

[0064] Thus, the template completes full boundary crossing before being uploaded, reducing the probability of boundary issues being exposed after going live; coverage rate, as a single-value indicator, facilitates unified control across multiple sites.

[0065] Textual conclusions in teaching evaluations often contain statistical terminology. To avoid misuse of terms such as "significantly improved," the combination of the significance test probability and whether the confidence interval crosses zero is used as a threshold to calculate the predicate permission function, which is then used as one of the necessary conditions for template permission. The predicate permission function is as follows:

[0066]

[0067] In the formula: predicate permission function : Values or Its function is to determine whether terms such as "significantly improve" or "superior" can be used, requiring significance; the term is used only if the probability of the significance test is less than the significance level and the confidence interval does not cross zero. .

[0068] Significance test probability : The probability value from a defined testing procedure, with a range of values. Significance level Thresholds set by users in the indicator definition library; commonly used value ranges. ; Lower limit of confidence interval Upper limit of confidence interval : The endpoints of the interval estimate of the effect, taking real values; Indicator function The value is taken as follows: when the condition inside the curly braces is true. Otherwise .

[0069] For each sentence in the property test pre-run, calculate the predicate permission function using words. When templates are used, words that significantly improve or outperform similar words are used. If the error occurs, immediately output a repair list, suggesting replacement with sentences that have insignificant or comparable differences, and add the sentence to the prohibited list; once all sentences meet the four thresholds of dimensional compatibility, sum-difference consistency, sufficient coverage, and permitted word usage, the system will mark the template with an upload permission flag.

[0070] Therefore, using explicit threshold functions to bind statistical terminology with data constraints avoids arbitrary word choice; the existence of license marks provides a verifiable basis for subsequent site deployment and version management.

[0071] Step 2: Unify the original time series and event-based rule network versions to the same scale and display precision, and generate a fact package with complete source, formula, unit, rounding, timestamp, and responsible person. This provides a unique and reproducible factual basis for subsequent rendering plans, minimal modification corrections, and credible issuance.

[0072] The raw data for teaching spinal anesthesia needle puncture comes from multiple sources. Force, displacement, posture, pressure, ultrasound frames, and manually labeled events are located on different timelines and in different unit systems. Direct aggregation can easily lead to dimensional conflicts, temporal misalignments, and unbalanced equations. Even if the sentence structure is correct in subsequent rendering, errors will be perpetuated. Therefore, it is necessary to first standardize all collected data to a standardized unit under the constraints of the rule-based system, and to map each indicator to its source, formula, and responsible person, thereby forming an auditable fact package.

[0073] First, the multi-source timeline is aligned and reconstructed to stably extract key events such as the start and end of puncture and the disappearance of resistance. Then, various physical quantities are standardized to uniform units according to the conversion chain. Subsequently, a read-only encapsulation of indicator vector, metadata, and data spectrum is generated under the same display precision. Finally, the fact package version and checksum are output. Before entering text rendering, the data is cleaned and solidified once, ensuring that subsequent steps are based on the same fact package, eliminating hidden errors caused by inconsistent units and time misalignment, and significantly reducing manual rework costs.

[0074] Furthermore, force, displacement, attitude, pressure, and ultrasonic frames are mapped onto a unified time axis to reconstruct a continuous trajectory at the same density. On the reconstructed trajectory, the time points and time periods of key events are automatically determined according to physical characteristics, generating an event time set as the time reference for the calculation of all subsequent indicators.

[0075] On-site, the operator imported force sensor data, displacement encoder data, 3D attitude data, pressure sensor data, and ultrasonic frame timestamps all at once. The system first reads the rule network version and conversion chain, using the manual button press at the start of the puncture as the anchor point to align the start times of each channel. Then, monotonic shape-preserving interpolation is used to reconstruct the discrete sampling points, avoiding overshoot and false peaks. To ensure the reproducibility of the process, the interpolation kernel is fixed during import, all interpolation nodes are derived from the original timestamps, and the end point of the interpolation segment strictly falls before the end of the puncture, without crossing the termination mark.

[0076] After aligning the starting point, interpolation reconstruction is performed channel by channel. The force, displacement, and attitude channels use piecewise cubic shape-preserving interpolation, while the pressure channel uses piecewise quadratic shape-preserving interpolation. Ultrasonic frames are reconstructed into an equally spaced sequence using nearest-neighbor shape-preserving interpolation. After reconstruction, the aligned time axis-multi-channel trajectory is output at the same time resolution.

[0077] All physical quantities are mapped to the same time axis, ensuring continuous trajectories and preserving original monotonicity, thus avoiding spurious events caused by interpolation overshoot. Subsequent event extraction and unit normalization have a unified carrier, guaranteeing that every indicator in the fact package can be traced back to the same time reference. Among these, the disappearance of resistance is a key event in epidural training, and relying solely on thresholds is easily affected by individual differences.

[0078] Furthermore, to improve robustness, events are identified at the curvature peaks and troughs of the force-displacement trajectory: the neighborhood where curvature rises to a peak and then rapidly falls back typically corresponds to the instant of ligamentum flavum rupture; curvature troughs correspond to the stable segment of the push. A curvature function is used to geometrically characterize the force-displacement curve to determine the event time set. The curvature is constructed as follows:

[0079]

[0080] Where: curvature : Measuring force-displacement over time The degree of curvature of the trajectory, taking a non-negative real number, serves as a geometric characterization factor for event determination. ,time Belongs to a closed interval; force sequence Forces aligned on the time axis, with values ​​as real numbers; displacement sequence Align the displacement on the time axis, and take values ​​as real numbers; first derivative. : These are the first derivatives of force and displacement with respect to time, respectively, which are used to characterize the rate of change;

[0081] Second derivative : These are the second derivatives of force and displacement with respect to time, respectively, used to characterize acceleration and bending tendency; time variable The event falls within the closed interval from the start to the end of the puncture; the derivative is given by the central difference (internal points) and the one-sided difference (boundary points); the interpolation uses piecewise cubic shape-preserving Hermite interpolation (PCHIP, no abbreviation needed) to avoid overshoot; event extraction is performed using a peak-valley neighborhood pattern combined with start and end anchor point filtering.

[0082] The first and second derivatives of the reconstructed trajectory are calculated at fixed time steps to obtain a curvature sequence. Peak-valley-peak or valley-peak-valley structural patterns are then searched on this curvature sequence. The search range is limited by aligning the start and end of puncture events, outputting event time sets such as the moment resistance disappears, the interlayer crossing period, and the needle tip stillness period. Therefore, the event time sets are derived from a unified curvature determination, reducing manual threshold adjustments; events correspond one-to-one with the time axis, facilitating subsequent index calculations using the same time period for integration, counting, and segmented aggregation; thus, the time dimension of the fact package has a stable source.

[0083] Furthermore, based on the existing aligned timeline and event time set, the units from different sources are unified into standardized units according to the conversion chain. Then, statistical quantities such as index vectors and quantile intervals are generated, and the source-formula-unit-rounding-timestamp-responsible person are written into the genealogy record, and finally encapsulated into a read-only fact package.

[0084] Since multiple devices use different units, without prior standardization, consistency of sums and differences, as well as interval constraints, cannot be guaranteed. Linear conversion is used to map source unit values ​​to standardized units; the constraints on accuracy and consistency of sums and differences are derived from the rounding operator in step one, and are thus solidified into the display rules of the fact package. Among these, unit conversion is included.

[0085]

[0086] Where: Normalized value : A physical quantity in standardized units, serving as the sole input for subsequent index calculations; source value Physical quantity in source unit; scale factor : The scaling factor from the source unit to the standard unit, with a positive real number as the value; Offset term : The translation amount used when there is a zero-point difference between units; it is usually zero and takes the value of a real number; value range: , , .

[0087] The scale factor and offset term are determined item by item according to the conversion chain, unifying force, displacement, pressure, angle, etc., to standardized units. Then, under the constraint of display accuracy, the rounding operator is called to save the values ​​used for display and the values ​​used for calculation separately, and the mapping relationship between calculated values ​​and display values ​​is explicitly recorded to avoid imbalances in the equations introduced by rounding during subsequent rendering. Thus, all physical quantities are unit-normalized before entering index aggregation, and the display accuracy is controlled and traceable. The coexistence of calculated values ​​and display values ​​makes the boundary between the fact layer and the display layer clear, reducing the frequency of triggering subsequent minimum modification corrections.

[0088] Furthermore, after standardizing units and display precision, various indicators are calculated around the event time set: time-period integral indicators, frequency indicators, and angle deviation indicators; simultaneously, quantiles are used instead of variance statistics to express heterogeneity, reducing sensitivity to outliers; finally, the rule network version, indicator vector, conversion chain, and genealogy record are encapsulated with an irreversible signature to generate the fact package version and checksum. Among these:

[0089]

[0090] In the formula, the quantile interval : Used to describe the range of a set of indicators, serving as a reference for interval display and terminology gating during subsequent rendering; quantile function In the dataset The quantile function returns the value that makes the distribution function reach the quantile level. The numerical value, in a specific form (Hyndman–Fan type 7): for ascending samples ,

[0091]

[0092] Parameters and their range: ; .

[0093] Dataset : A set of similar metrics derived from the aligned timeline and event time set; quantile level :satisfy Two numbers: For a closed interval, ,and .

[0094] For each type of indicator, it is accumulated into the dataset, and a specified quantile interval is calculated to form an indicator vector-quantile interval pair. Then, a signature input string is constructed, and the rule network version, indicator vector, conversion chain, and genealogy record are concatenated in a fixed order. The signature is calculated and written into the fact package so that any tampering can be detected by the verification code.

[0095] Furthermore, to avoid symbol confusion, the signature process does not change any index values, but only performs an irreversible mapping on the byte sequence, where:

[0096]

[0097] In this formula: Signature The checksum of the fact packet is used for integrity verification. For a fixed-length byte sequence, a hash function : Irreversible mapping function, takes a fixed-length byte sequence as its value; version identifier The immutable representation of the rule-based network version is the input byte sequence; index vector. : All calculated values ​​of indicators arranged in a fixed order, which is the input byte sequence; conversion chain The ordered set of scale factors and offset terms used in unit normalization; genealogical record. The system is an ordered set of source, formula, unit, rounding, timestamp, and responsible party. In application, the quantile interval ensures that the interval expression of indicators and the gating of terminology are based on evidence; the signature binds the rule network version, indicator vector, conversion chain, and genealogical record into an indivisible whole, ensuring the integrity of the fact package and facilitating subsequent offline verification and audit playback.

[0098] As an example, in the Skills Center, the engineer loads the rule network version and template package into the system interface and inserts the original folder of the day's training. The interface first displays an alignment timeline, and the operator uses the mouse to align the start and end marks of the puncture to a unified zero point and end point. Then, the operator clicks "Reconstruct Trajectory," and a force-displacement curve appears on the right side of the screen. Sharp turns in the curve are automatically marked as the disappearance of resistance, and the corresponding timestamps appear in the event list below.

[0099] The engineer clicked on "Unit Normalization" to switch the pressure source unit from millimeters of mercury to the standardized unit centimeters of water column, and the interface displayed a conversion chain. Next, they clicked "Display Precision" to round the time used for display to the nearest whole second, and the system showed a comparison between the calculated and displayed values. Finally, they clicked "Generate Fact Package," and the system generated a read-only file with a version number and checksum in the folder. The endnote preview area listed the source, formula, unit, rounding strategy, and responsible person's signature for each indicator.

[0100] Step 3: Before generating any text, a rendering plan is formed based on the unified semantics of the fact package, rule network version, display precision, conversion chain, boundary set, and predicate permission function. At the planning level, it is determined once and for all what to say, how to say it, what units and precision to use, and which words are allowed. A pre-scoring score and a list of issues are also given so that there will be no fundamental deviations when entering the next step of constrained generation.

[0101] Readers of teaching reports often prioritize whether the standards are met, the relationship to the pass / fail line, the significance of differences, and how key events are described. If these crucial deliverables are left to be passively patched after generation, repeated rework will occur. Therefore, predicate selection, units and precision, referential integrity, and sum / difference consistency need to be moved to the rendering plan before generation, supported by computable threshold and cost functions.

[0102] First, based on the indicator vectors, quantile intervals, boundary sets, and predicate permission functions in the fact package, the word choice and semantic direction of each sentence are planned. Then, under the constraints of display accuracy and consistency of sum and difference, the unit and number of digits to be retained for each data reference mark are determined. Subsequently, a pre-scoring is calculated, and items that cannot be automatically repaired are marked as a list that must be manually confirmed. Major items that may lead to rework are resolved before generation, and subsequent corrections only require minimal changes within a small scope. The rendering plan becomes the hard constraint input for the next step, ensuring that the same fact is consistently represented across different templates and sites.

[0103] Among them, the predicate permission function-semantic direction-word cost is the main thread. First, the sentence statement is determined, and then the number is written under the constraints of display precision-sum consistency. The two together form the word selection table and the display strategy table.

[0104] Within the framework of the fact package, the semantic core of a sentence consists of two parts: whether the use of words requiring salience is permitted and the direction of the numerical value. The former depends on the predicate permission function, and the latter depends on the semantic direction; both jointly determine the word selection. To form an executable decision criterion, a threshold function is first used to determine permission, then a cost function is used to measure the consistency between candidate words and directions, and finally, a word selection table is output. The predicate permission function is as follows:

[0105]

[0106] In the formula: predicate permission function : Values or Its function is to determine whether words that significantly improve or outperform the class requirement of significance can be used;

[0107] Significance test probability : Test results from the fact package, value range This is used to measure the confidence level of the difference relative to the line of passage; The test function outputs: If comparing proportions with passing lines, use an exact binomial test or a permutation test (number of permutations). With random seed Write metadata); if comparing continuous quantities with thresholds, use a single-sample permutation test; confidence intervals are given by Clopper–Pearson (proportional) or quantile intervals (continuous).

[0108] Significance level The threshold value range given in the indicator definition library. Used to trigger a word choice threshold; lower bound of the confidence interval Upper limit of confidence interval : Derived from the endpoints of the fact packet quantile interval and the test result, taking real values, used to determine whether it crosses zero; Indicator function If the condition is true, take Otherwise take .

[0109] For each sentence that requires a condition, first calculate the predicate permission function. If the value is If the candidate words are significantly improved or better than, then only words with insignificant differences, equivalent to, or not meeting the standard are retained. Then the semantic direction is calculated, which is determined by the sign of the difference between the index and the passing line, and words with higher directional consistency are selected from the allowed word set accordingly.

[0110] Furthermore, to quantify consistency, a word cost function is defined, with lower costs taking precedence. The word cost function is as follows:

[0111]

[0112]

[0113] Where: word cost function : Non-negative real numbers, Its function is to measure the availability and directional relevance of a candidate word; a smaller value is better. (Weight coefficient) : The value is a positive real number, and its functions are to characterize the penalty strength for permission violation and the penalty strength for directional deviation, respectively; semantic direction : Values The sign of the difference between the indicator and the passing line is determined by the sign of the line. Indicates below, It means quite, Indicates "above"; direction of candidate words : Values Values ​​are pre-assigned according to word meaning; for example, "not meeting the standard" is mapped to... , which is equivalent to Better than mapping to ; Directional deviation Defined as Non-negative real numbers are used to measure directional consistency.

[0114] Among them, the word cost function is calculated sentence by sentence. The candidate word with the lowest cost is selected as the word for the sentence; if the costs are equal, the more robust word is selected according to a pre-defined rule of prioritizing robustness. This process generates a word selection table, providing explicit predicates for subsequent syntactic skeleton binding.

[0115] Thus, through the joint constraints of permission, direction, and cost, the core words of a sentence are determined before generation; when the word cost function... In this process, any salient words are excluded, thus preventing incorrect wording from entering subsequent stages; the measurement of directional consistency provides an objective standard for selection, reducing subjective decision-making.

[0116] Furthermore, the numbers in the sentence must not only be correct, but also aligned. If the number of digits retained is not set properly, the displayed value after summing the items may not be equal to the displayed value of the total items, which will confuse the reader.

[0117] Therefore, during the planning phase, a one-time check must be performed based on the display accuracy and conversion chain in the fact package, and suggestions for interval-based display or accuracy adjustment should be given if necessary. The method for handling the consistency residuals of sums and differences is as follows:

[0118]

[0119] Where: Consistency of sum and difference residuals : Non-negative real numbers, used to measure the number of decimal places to retain. The difference between the total and sub-item values; the number of sub-items : A positive integer representing the number of terms in the summation; the original value of each term. : Real number, calculated value from the fact pack; absolute value symbol : Returns a non-negative difference.

[0120] Integer operator Perform bitwise operations on the input. Rounding or banker rounding, It is a non-negative integer, and the specific rules are fixed in the rule network version, where:

[0121]

[0122]

[0123] And define the parity function:

[0124]

[0125] Parameters and range of values: number of decimal places ; ; .

[0126] Iterate through all total-to-part relationships in sequence and calculate the consistency residuals of sum and difference. When the sum and difference are consistent, the residuals In this case, prioritize increasing the number of bits to retain. If adjusting to the upper limit allowed by the rule network version still fails... In the presentation strategy table, mark this location as a priority for range-based presentation, and indicate in the issue list that range-based expression should be prioritized over single-value expression in the next step. The presentation unit directly references the conversion chain to ensure consistency with the fact package. Therefore, during the planning phase, it's crucial to identify elements that guarantee consistency between sums and differences. Alternatively, a clear path can be established for transitioning to a range-based display, preventing subsequent generation from triggering rework due to display precision issues; the display unit should be consistent with the conversion chain to avoid unit confusion.

[0127] As an example, on the workstation in the Skills Center, after loading the fact package and template package, the instructor clicks on word planning. A word selection table appears in the center of the screen, with each row corresponding to a sentence, and candidate words and their costs displayed on the right. The instructor hovers the cursor over the word "better than," and the system explains below that the cost of that word in this sentence is a constant, and it is automatically hidden because the predicate permission function is in a prohibited state. The instructor then switches to the display strategy, and the left side of the interface displays preview values ​​of the sub-items and the total item, while the upper right corner displays the current number of bits to retain. The tutor moves the cursor to a summation sentence, and the system highlights the sum-difference consistency residual value below it, providing a suggestion for interval display on the button. The tutor clicks "Apply," and the sentence is marked for interval display in the plan. Throughout the entire process, no raw numbers or statistics are displayed; only visible clicks, toggles, and markings.

[0128] Furthermore, after completing the word selection and display strategy, it is necessary to bind the data reference tags, units, precision, and predicates to the specific syntactic skeleton, calculate the pre-score, and include items that cannot be automatically repaired in a list that must be manually confirmed, thus forming a rendering plan that can be directly used for the next step.

[0129] The syntactic skeleton is the structural part of the template that does not change with the data, while the data reference markers are the variable parts of the skeleton. To ensure that the bound sentences can be directly generated in the next step, citation integrity checks and skeleton binding need to be completed during the planning phase. Simultaneously, to provide quantitative basis for subsequent scoring and gating, a pre-score is calculated for the bound sentences. This score considers citation integrity, consistency of units and dimensions, consistency of sums and differences, and consistency of word permission and direction. Specifically:

[0130]

[0131] In the formula: pre-scoring : Value in Its function is to measure the feasibility of a rendering plan; a higher value is better. (Weight coefficient) : Positive real numbers, representing the penalty intensity for missing citations, dimensional conflicts, sum / difference residuals, and word choice costs, respectively; number of missing citations. : A non-negative integer representing the number of data reference tags in the skeleton that are not mapped by the fact package;

[0132] Dimensional conflict Quantity Dimensional conflict : A non-negative integer representing the number of entries whose units and dimensions are inconsistent after binding; average cost of terminology. The arithmetic mean of the word cost functions of all sentences to be judged, with values ​​being non-negative real numbers;

[0133] Step function When the sum and difference are consistent, the residuals A positive number is applied as a penalty if the time condition is met; otherwise, a negative number is applied. The specific value is fixed in the version of the rules network;

[0134]

[0135] This is the step penalty function; The input is the scalar value to be evaluated; This is the penalty amplitude constant. Its range and function are as follows: ; Output a fixed penalty value if and only if the condition is true; otherwise, output zero.

[0136] Check each data citation marker to see if there is a corresponding item in the fact package; if a missing one is found, record it as a missing citation. The displayed units are compared with the conversion chain item by item, and any discrepancies are recorded as dimensional conflicts. For sentences involving summation, read the result obtained in the previous step. Determine the step function Deductions; for all sentences requiring gating, calculate and average word count. Finally, a preliminary estimate was given. And generate a list of issues based on the source of the deduction.

[0137] When using it, the scoring criteria are placed in advance, so that the risks and priorities are clear before the next generation stage begins; after binding, the skeleton and data reference tags, units and precision are matched one by one, providing a one-click rendering plan for the next constrained generation.

[0138] However, numerical scoring alone is insufficient; operators need a clear understanding of where changes are needed, why they are necessary, and how to implement them. Therefore, during the planning phase, constraint triggering paths are constructed based on the dependencies between rule network versions, mapping each penalty to a specific rule, marker, and syntactic location. Based on this, pre-repair suggestions are generated, with their priority correlated with penalty intensity, repair cost, and the scope of impact on other sentences.

[0139] Using the dependency graph of the rule-based network version as a foundation, a path from the root constraint to the sentence fragment is constructed for each deduction, and presented sequentially on the interface in a highlighted manner, allowing the operator to view each point along the path; at the same time, pre-remediation suggestions are provided on the right, such as increasing the number of bits to retain. One suggestion is to replace words with those that have higher directional consistency, or switch to a range-based display. All suggestions use already calculated values ​​for display accuracy, word cost, and sum / residual, without redundant calculations.

[0140] When using it, the operator can locate the source of the problem and the solution without having to repeatedly jump between multiple interfaces; since the suggestions are based on existing calculation results, when entering the next stage of constrained generation, the operator can directly make minimal changes along the suggested path, shortening the distance from planning to finalization.

[0141] As an example, the tutor clicks on the pre-score on the rendering plan page. A panel pops up, displaying the pre-score value at the top and listing four categories of items categorized by source: missing citations, unit conflicts, sum / difference residuals, and word choice costs. The tutor clicks on a unit conflict item, and a constraint trigger path appears on the right side of the screen, pointing from the unit and unit rules to a specific data citation marker in a sentence, with each node in the path highlighted. Pre-repair suggestions appear in the lower right corner of the panel, including switching to standardized units along the conversion chain and marking the sentence as an interval display. The tutor checks the suggestion and clicks "Apply to Plan," and the item disappears from the issue list, while the pre-score automatically updates.

[0142] Step 4: Based on the unified semantics of the rendering plan, fact package, and rule network versions, first generate a draft of the constrained text, then reverse extract the draft back to the rule network for secondary verification. If a violation of the constraints is found, the text and format changes are made with minimal changes without altering the fact values. Finally, the consistent text, correction list, and final consistency score are output.

[0143] Among them, the teaching report on spinal anesthesia needle puncture must be consistent with the facts and also meet the hard constraints of consistency of sum and difference, consistency of units and dimensions, and gating of statistical terminology.

[0144] If the code is directly allowed to proceed to the manual editing stage, issues such as raw data, misuse of units, and overstepping wording are highly likely to occur, leading to interception during the release governance phase and repeated rework. Therefore, it is necessary to first perform a constrained generation, binding the word choices, display strategies, and skeleton determined in the rendering plan into text. Then, the text is reverse-extracted and rules are fed back to form a set of violations. Minimal modifications are then made to the set of violations, ensuring that the facts remain unchanged while the expression is finely adjusted, thus ensuring a stable path to trusted issuance.

[0145] Therefore, the processing involves the following steps: First, constrained generation is performed, filling the syntactic skeleton with data reference markers, units, reserved digits, and predicate selections one by one to form the initial text draft; Second, the initial text draft is reverse-extracted to obtain triples of numerical values, units, and predicates, and post-validation is performed on the rule network to form a violation set; Third, the violation set is corrected with minimal modifications in a fixed order, strictly adhering to the boundary of not changing the factual numerical values; Fourth, if there are still irresolvable dimensional incompatibilities or missing references, a failure handling repair script is generated and export is blocked. This completes the generation-validation-correction closed loop within the same stage of text generation, relieving subsequent release governance of the pressure of content-level rework and objectively reducing the probability of common errors such as misused words, unit mismatches, and sum / difference imbalances.

[0146] The process involves rendering each sentence based on the word selection table, display strategy table, and skeleton binding table of the rendering plan, prioritizing data reference tags, adjusting display precision accordingly, and using units from the conversion chain to form a draft text. Subsequently, a second verification is performed on the same rule network to map the draft back to the algebra-dimension-interval-implied criteria and locate potential violations.

[0147] Furthermore, the process revolves around the coupled execution of the syntactic skeleton, data reference tags, and presentation strategies. First, the skeleton binding table is read, matching each data reference tag with its corresponding item in the fact package. Then, units and reserved digits are assigned to each tag according to the presentation strategy. Finally, predicates are placed on the skeleton based on the word selection table. To make this process have a verifiable functional form, a text generation mapping is introduced, jointly mapping the skeleton, value table, and strategy to a draft text.

[0148] During execution, the system scans each data reference marker position according to the skeleton's order, retrieves the corresponding calculated value-display value comparison from the fact package, and writes the display value into the sentence according to the display strategy. If the position is marked as a range-based display in the rendering plan, the quantile range is read for range expression; if it is a single-value display, the display value is strictly written. Predicate positions are assigned words according to the word selection table, and all unapproved candidate words are blocked. Any raw numbers are rejected from entering the text; that is, if a marker has no corresponding fact package item, execution stops and a missing reference is marked.

[0149]

[0150] In the formula: the first draft of the text : An ordered set of string sequences, used as input for subsequent validation and correction; skeleton : The ordered structure of the syntactic skeleton, taking values ​​as a sequence of alternating spaces and fixed phrases; value table : Calculated values ​​from the fact package - a set of values ​​compared to quantile intervals; strategy The rendering plan includes a word selection table, a display strategy table, and unit and bit depth settings.

[0151] Value range: A finite-length string sequence; A placeholder sequence with a fixed structure; It is a bounded set; Given a finite set of rules. Generating operators. A deterministic operator that maps skeletons, value tables, and strategies to text;

[0152]

[0153]

[0154] In the formula: syntactic skeleton :A collection of literal fragments Composed of placeholder fragments; placeholders are divided into numerical placeholder sets. With predicate placeholder set .

[0155] Value table :Tag for each data reference Give the calculated value Quantile dataset Exhibition Units Display digits These all come from the fact package and rendering plan.

[0156] Strategy Includes display methods for "single values ​​or ranges", unit writing rules, number localization rules, and mapping of predicate selection results and placeholders to predicates.

[0157] Connection symbol This indicates that the strings are concatenated in order to form a string sequence.

[0158] Numerical evaluation functions Specific format: If the display method is a single value, then take... And use the rounding operator by digit. If the display method is range, then in the dataset Take the two ends of the quantile Combining into interval values ​​( (Given by the rendering plan).

[0159] Formatting functions : Input numerical value or range Number of bits With unit string Generate target text (including fixed rules for decimal places, thousands place, spaces before and after units, and interval conjunctions).

[0160] Predicate selection function Retrieve placeholders from the strategy The corresponding decided terms (determined by the previous term gating and cost-minimizing selection).

[0161] Placeholder Map placeholder fragments to value table keys or predicate key .

[0162] By using a functional representation of text generation mapping, constrained generation becomes inspectable and reproducible; rejecting raw numbers ensures that every value in the text can be traced back to a corresponding item in the fact package; units and reserved bits are implemented exactly according to the rendering plan, avoiding logical inconsistencies caused by subsequent adjustments.

[0163] At the text level, the generated initial text draft is parsed, extracting the displayed values, units, and predicates one by one, and combining them into a set of triples of value-unit-predicate; this set is then fed back into the rule network for judgment, forming a set of violations. Reverse extraction requires lexical recognition of number format, interval format, and unit name, and mapping predicates to direction labels.

[0164] The process involves parsing each sentence according to the skeleton binding table of the rendering plan to locate the actual text fragments that are marked with data references; performing lexical segmentation on numbers using fixed decimal point and thousandths rules; using a process of left and right parentheses recognition and endpoint parsing for intervals; matching units with the standardized names in the conversion chain; and labeling predicates with a candidate word-direction mapping table.

[0165] The set of triples is then evaluated according to the equations, dimensions, intervals, and implication relations of the rule network. Any triples that do not satisfy the conditions are added to the violation set, where the violation set is defined as follows:

[0166]

[0167] Where: set of violations : Contains all constraint entries that are not satisfied in the initial draft of the text; constraint set : All constraint entries from the RulesNet version for subsets; constraint determination The Boolean result obtained by substituting the initial draft of the text into a certain constraint has a value of [value]. or , .

[0168] When used, the text-to-rule loop is clarified by defining the violation set in a set-like manner, and any item that does not meet the requirements can be located in the set; the size of the set and the source of the item become the direct inputs for the sorting of the correction strategy in the next sub-step; during the extraction process, the name is matched with the conversion chain to ensure that the dimensional verification is still valid at the text level.

[0169] As an example, the tutor clicks "Generate Draft" on the workstation. Segmented text appears on the left side of the screen, each with a citation marker. The tutor hovers the mouse over a number in a segment, and the system displays the source item, calculated value, and displayed value from the fact package below. Clicking "Validate" brings up a list on the right, with items arranged in algebraic format: unit name - predicate. Some entries are marked as failing. The tutor clicks on one entry, and the system highlights the corresponding sentence in the text, displaying the dimensional consistency entry from the rule network at the bottom. The tutor doesn't input any data, simply pressing "Enter Correction" on the interface, and then proceeds to the next sub-step.

[0170] Furthermore, based on the set of violations, minimal modifications are performed in the following order: no change to factual values, priority replacement of high-risk predicates, adjustment of display units, and final fine-tuning of the number of bits to retain. Each modification is measured by a cost function, achieving a balance between penalty intensity, repair cost, and scope of impact. If unrepairable situations such as dimensional incompatibility or missing references occur, a repair script is generated and export is blocked.

[0171] The cost function measures the cost of different correction actions, with lower costs indicating higher priorities. Correction actions include predicate substitution, unit display switching, fine-tuning of retained digits, and minor rewriting of sentence structure, while requiring that the factual values ​​remain unchanged.

[0172] The cost function takes into account the action type, number of actions, and scope of influence. The search sequence follows a progressive amplification method from words to numbers and from local to whole sentences.

[0173] First, the violation set is categorized by source. For violations involving out-of-bounds statistical terminology, the candidate set for predicate replacement is directly entered. For violations involving inconsistent unit names but compatible dimensions, the candidate set for unit display switching is entered. For violations involving inconsistent sums and differences, the candidate set for fine-tuning the retained digits or changing to interval display is entered. For violations involving syntactic incompleteness, the candidate set for minor structural rewriting is entered. Then, the cost of each candidate solution is calculated, and applications are attempted in ascending order of cost until the violation is resolved or no solution is available. The minimum modification cost is:

[0174]

[0175] Where: minimum modification cost : A non-negative real number used to measure the total cost of a set of correction actions; the number of predicate substitutions. : Non-negative integer; number of display toggles per unit : Non-negative integer; number of decimal places to retain; number of adjustments : Non-negative integer; Number of minor rewrites of sentence structure : Non-negative integer; scope of influence quantification : A non-negative real number representing the weight of the number of sentences or the span of paragraphs affected by this correction; weighting coefficient : Positive real number, pre-set according to the strategy of penalty intensity - repair cost - scope of influence.

[0176] When in use, the minimum changes are concretized into a set of comparable and selectable actions with the cost function as the core, avoiding subjective judgment; the search sequence is from light to heavy, and the correction is completed first at the word and presentation layers, and the syntactic structure is only touched when necessary, thereby reducing the disturbance to the entire text; the fact values ​​are kept unchanged so that the correction process does not change the fact layer.

[0177] Furthermore, once the correction action is applied, the text needs to be updated immediately, the post-validation run needs to be rerun, and the final consistency score needs to be calculated accordingly. If there are still violations that cannot be resolved, a failure handling script is generated. The script clearly lists the missing fact items, the source of units with incompatible dimensions, the syntactic segments that need to be added, and blocks the export.

[0178] The process involves applying a set of actions with minimum cost, updating the text, and repeatedly performing reverse extraction and rule backfeeding. When all violations are cleared, a final consistency score is calculated, and a correction list is generated. If unresolvable items exist, the script is output in the order of fact package first, then template, guiding the process back to the data normalization and genealogy packaging or constraint compilation and template pre-review stages for repair. The final consistency score is as follows:

[0179]

[0180] Where: Final consistency score The value is not greater than The real number serves as the direct input for governance; weighting coefficient : Positive real numbers, corresponding to the penalty intensity for missing citations, dimensional conflicts, sum / difference residuals, mean cost of word usage, and correction span, respectively; number of missing citations : A non-negative integer, referring to the number of missing references that still exist after correction; number of dimensional conflicts. : A non-negative integer, referring to the number of dimensional inconsistencies that still exist after correction;

[0181] Average cost of word usage The mean of the corrected word cost function, a non-negative real number; correction span. : A non-negative real number that quantifies the sentence or paragraph span affected by the correction;

[0182] Step penalty A positive penalty is applied when the final sum / residual is greater than zero; otherwise, a negative penalty is applied. ;

[0183]

[0184] Step penalty function : A threshold function that maps the input to 0 or a fixed constant; the final sum-difference consistent residual. : The final difference between the total term and each component term at a predetermined number of decimal places (calculated after correction in step four); penalty amplitude constant. A positive number configured by the rules network version, used to reflect a fixed deduction in the scoring; indicator function. : If the condition is true, take 1; otherwise, take 0.

[0185] This item is invoked in both the preliminary and final consistency scoring: when, after minimal modifications, unequal sums and differences still exist, it is determined by... Output a fixed penalty constant; output zero if the residual has been eliminated. This approach, which uses a fixed threshold upon triggering rather than continuously scoring based on residual size, aligns better with discrete decisions regarding the number of bits retained and the presentation strategy.

[0186] When in use, the scoring update and correction list are output simultaneously, enabling the release governance to make tiered decisions accordingly; the failure handling script specifies the reasons for the blockage and the backtracking path, avoiding aimless rework; the linkage between scoring items and correction spans suppresses the inappropriate practice of making major corrections for minor errors.

[0187] As an example, on the same workstation interface, the mentor clicks "Execute Correction." In the interface, the system first replaces "significant increase" in the sentence with "not significant difference," and the highlight in that location disappears; then, it switches the unit names in two places to the standardized names consistent with the conversion chain; finally, it changes the display method of a summation sentence to a range display. The mentor clicks "Re-verify," and only one missing reference remains in the list. The system automatically generates a repair script, prompting that the sentence needs to select the corresponding item from the fact package or return to the template to add a data reference marker to that location. The mentor clicks "Generate Correction List," and the right side of the screen displays a list of replacement words, unit switching items, and display method adjustments, with the final consistency score displayed at the bottom. The mentor clicks "Submit to Release Governance," moving to the next step.

[0188] Step 5: Under the unified semantics of the final consistency score, consistency text, fact package, and rule network version, complete the trusted issuance with hierarchical gating, and when data or thresholds change, perform minimal re-rendering and difference presentation only for the affected sentences, making the report auditable, verifiable, and reproducible.

[0189] Once a teaching report enters the publication stage, any issues such as inappropriate wording, inconsistent units, imbalances in sums and differences, or unclear sources will directly impact the credibility of the teaching evaluation. Therefore, the publishing side must implement tiered decisions on consistency texts based on calculable thresholds and include the source, formula, unit, timestamp, and issuer in the document metadata for offline verification. Meanwhile, training data rollover or online adjustments are inevitable; rewriting the entire report each time would result in significant rework. The publishing side should modify the impact domain calculation, re-rendering only the affected sentences and outputting the differences and their causes.

[0190] First, based on the final consistency score and statistical calculation of blocking items, a tiered gating system is implemented. Texts that pass the gating are issued, while those that do not meet the criteria are automatically sent for manual review or directly rejected. For texts that pass the gating, a document fingerprint is formed by embedding a fact package signature, rule network version, template version, issuance timestamp, and issuer identity into the document metadata. Then, changes to data, thresholds, and templates are monitored, the impact domain of the changes is calculated, and affected sentences are re-rendered sequentially according to a priority queue, generating a difference comparison and the reason for the change. The document is updated while retaining the audit trail. The release decision has clear and verifiable basis, and the document can be verified offline. Incremental updates only touch on necessary content, avoiding a complete rewrite. The entire release and update process is traceable, meeting the requirements for teaching audits and spot checks.

[0191] Based on the final consistency score and whether there are still blocking items, three levels of decision are made; for the text that is allowed to pass, a document fingerprint is generated and the source-formula-unit-timestamp-signer is written into the document metadata so that the report can be verified outside the system.

[0192] In this process, the release gating is not a single threshold switch, but a segmented decision combining the final consistency score and the number of blocking items. The gating threshold is derived from the rule network version, and the blocking items refer to fatal problems that cannot be resolved even after the fourth step of correction.

[0193] Read the final consistency score, blocking item count, and gating threshold, and substitute them into the piecewise function to obtain the decision result; once the decision is determined, the system will solidify the corresponding release action and logging action into the audit track, and clearly prompt the signatory on the interface.

[0194]

[0195] In the formula: issuing decisions Values ​​can be released, reviewed, or rejected; the function is to drive the release process; final consistency score. The score from step four, with a value no greater than [value missing]. A real number used to represent the degree of text consistency; upper threshold : The threshold for direct approval upon release is a real number, determined by the version of the rules website; the lower threshold The threshold between review and rejection, a real number, is given by the rule network version; blocking item count blocking. : A non-negative integer representing the number of fatal problems that cannot be solved after the fourth step.

[0196] Thus, the scoring, threshold, and blocking items are linked into clear segmented decisions, eliminating verbal judgments; thresholds are uniformly managed by the rule network version to ensure consistent interpretation across different sites; and the audit trail contains replayable decision-making evidence to meet regulatory requirements.

[0197] However, issuing a document is not simply about outputting a single document; it involves binding facts, rules, templates, issuers, and timestamps into a verifiable whole. To achieve this, a document fingerprint needs to be written into the document metadata, and a summary genealogy list needs to be embedded, allowing anyone to verify the document's origin and consistency offline.

[0198] The fingerprint input includes the factual packet signature, rule network version, template version, issuance timestamp, issuer identity, and final consistency score. The document signature is calculated and written into the document metadata. Simultaneously, a digest genealogy list is written, including the source indicator for each data citation, the formula used, units and dimensions, rounding rules, and conversion chain digest. The metadata is written as structured key-value pairs, independent of external services.

[0199]

[0200] Where: document signature A fixed-length byte sequence used for report integrity verification;

[0201] Document hash function : Irreversible hash function, outputs a fixed-length byte sequence; refers to a cryptographic hash function used to generate report fingerprints, which maps a byte string concatenated in a fixed order and with a fixed encoding to a fixed-length digest, used for integrity verification and offline verification. ;

[0202] Final Consistency Score As before, it serves as a quality characterization of the signed text; fact package signature. The signature from step two is used to bind the text to the fact package; RulesNet version. Version identifier from step one; template version : Current template package version identifier; Issuance timestamp : The time marker of the issuance action; the identifier of the issuer. : The issuer's identity hash or certificate identifier.

[0203] Once a document is generated, it carries a verifiable fingerprint and digest lineage; fingerprints can be compared and the source list can be read in any offline opening scenario; the combination of final consistency score - fact package signature - rule network version - template version - issuance timestamp - issuer identity makes the issuance behavior itself accountable and reproducible.

[0204] As an example, at the workstation in the skills center, the instructor clicks "Submit for Issuance." A gating panel appears on the screen, displaying the final consistency score, threshold description, and block count. The instructor presses "Confirm," and the interface automatically generates a footer section in the document footer with the source, formula, unit, timestamp, and issuer information. Simultaneously, a metadata preview pops up in the sidebar, listing the rule network version, template version, fact package signature, and document signature. The instructor copies the file to a teaching reader without network access, opens it, and clicks the "Verify Signature" button. The screen indicates that the signature matches and the source is consistent.

[0205] Furthermore, after publication, data may be rolled over, thresholds may be adjusted, and templates may be revised. Indiscriminately re-rendering the entire document would destroy existing traces and increase workload. Therefore, the impact domain of the change should be calculated first, and only the affected sentences should be re-rendered. Then, a difference comparison between the old and new texts should be performed, and the reason for the change should be provided. If the change causes the scoring or licensing function to cross a threshold, a review or re-issuance should be automatically triggered.

[0206] The core of the influence domain lies in determining which sentences are indeed affected by the changes. The rule network is a directed dependency graph, where sentences can be viewed as terminal nodes, and changes to data and thresholds are the source set. Using graph reachability operators, a directed traversal is performed starting from the source set to obtain the set of affected nodes. The intersection of this set with the sentence set yields the influence domain.

[0207] Processing method: Map data change items, threshold change items, and template change items to a unified source set, and perform a directed traversal on the rule network dependency graph to record reachable nodes; find the intersection of reachable nodes and sentence node sets to obtain the position index of sentences that need to be re-rendered; highlight these sentences on the interface for easy verification and confirmation by operators, where:

[0208]

[0209] In the formula: the influence domain The set of sentences that need to be re-rendered; graph reachability operator Operators for computing the set of nodes reachable from a source set in a directed graph; changing the source set... A unified set of data change items, threshold change items, and template change items; dependency graph. : Directed dependency structure of rule-based networks; Sentence node set The set of nodes corresponding to all sentences in the current consistent text.

[0210] When used, the influence domain calculation based on graph reachability can strictly and reliably determine the sentences that must be re-rendered, avoiding the need to rewrite the entire document; the sentence position index is visualized on the interface, making the operation intuitive and leaving clear traces.

[0211] The graph reachability operator `Reach(·,·)` is used to determine the set of all nodes reachable along directed edges from a given set of change sources in a directed dependency graph of the "rule network version," forming the basis for calculating the "change influence domain." Let the dependency graph G=(V,E) be a directed graph, and Δsource⊆V be the set of change sources (nodes corresponding to data, thresholds, or templates). Then: ;

[0212] Equivalent iterative closure form: Let R0 = Δ source, define the successor operator. Iteration:

[0213]

[0214] Until it stops growing, the limit set That is .

[0215] Input: Dependency graph G (nodes and directed edges have been determined by the rule network version in step one), change source set Δsource (derived from the actual changes to "data, threshold, and template" in step five).

[0216] Output: Set of reachable nodes In step five, the influence domain is obtained by intersecting the sentence node set S.

[0217] However, more than one sentence is often affected, and the order of re-rendering affects the efficiency of review and reader comprehension. A re-rendering priority weight is constructed using factors such as reachability depth, syntactic length, and constraint importance, prioritizing sentences with broad impact, complex structures, and important constraints. After re-rendering, the differences between the old and new texts and the reasons for the changes are output together.

[0218] The process involves calculating the priority weight for each sentence in the impact domain and placing them into a queue in descending order. Each sentence in the queue is then re-rendered using the rendering plan fragment generated in step three. After rendering, the old and new texts are highlighted with character-level differences, and the reasons for the changes are listed in a sidebar, distinguishing between data changes, threshold changes, and template changes. If the predicate permission function or the final consistency score crosses the threshold after re-rendering, a review process or re-issuance prompt is automatically triggered.

[0219]

[0220] Where: priority weight : Real number, the larger the value, the higher the priority; weighting coefficient : Positive real numbers, controlling the relative impact of reachability depth, syntactic length, and constraint importance respectively; Reachability depth : The shortest path length from the source of the change to the sentence node, a non-negative integer; syntactic length The number of characters or lexical units in the sentence, a non-negative real number; constraint importance. The importance score of the constraint item involved in this sentence is a non-negative real number, given by the rule network version.

[0221] When used, the priority weight combines the breadth of impact, structural complexity, and constraint importance into a single-value indicator, making the re-rendering order clear; differences and reasons are output side by side, making it easy for tutors to quickly determine whether a second review or re-issuance is needed; automatic flow triggered by threshold crossing avoids missing key changes.

[0222] As an example, on the same workstation, after the administrator increases the site's settings via a line in the configuration page, a notification bar appears at the top of the page indicating that a threshold change has been detected.

[0223] When the administrator clicks "Update," three sentences in the text are highlighted, and the sidebar displays one data change and two threshold changes. When the administrator clicks the first highlighted sentence, the system first displays the old sentence, then the new sentence, with the differences highlighted as characters. The sidebar notes the reason for the change: threshold change, and provides a checkbox for whether to trigger a review. If the administrator checks this option, the system temporarily saves the file as pending review, and an entry triggering a review due to a threshold change is automatically added to the audit track.

[0224] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0225] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0226] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0227] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0228] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A teaching system combining virtual and real methods for spinal anesthesia needle puncture, characterized by: include, Constraints are compiled on the indicator definition library, Chinese term matrix and report template containing data reference marks, and the conversion chain is determined by unit dimension reasoning. The constraints are solidified into a rule network version and generated into a template package after boundary pre-drilling. Based on the rule network version, the original time series and events are aligned, the units are normalized and the indicators are calculated according to the conversion chain, and the source, formula, unit, rounding and timestamp are written into the genealogy and encapsulated into a fact package with version and verification code. Read the fact package and template package, complete the pre-validation on the rule network version, form a word plan based on predicate permission and quantile interval, determine the display unit and reserved number of bits for each data reference mark, and generate a rendering plan with pre-scoring; Text is generated using data reference tags according to the rendering plan, and numerical values ​​and units are extracted and verified against the backflow rules. Without altering the factual values, word replacement, unit switching, and digit adjustment are performed with minimal changes to output consistent text and final consistency score. Issuance gating is performed based on the final consistency score, and the genealogy and signature are embedded in the metadata through text. When data, thresholds or templates change, the impact domain of the change is calculated according to the rule network, and only the affected sentences are re-rendered to generate difference comparisons and audit logs.

2. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 1, characterized in that: Unit and dimension reasoning is performed on the indicator definition library to establish a conversion chain; equality relations, sum-difference consistency, interval restrictions, conditional implications, and statistical terminology gating are compiled into a rule network version and semantically bound to a report template containing data reference tags.

3. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 2, characterized in that: Construct a boundary set that includes samples with zero or one coverage, confidence intervals crossing zero, rounding leading to unequal summation, unit mismatch, and significance thresholds. Perform a pre-run of each report template. When full coverage is achieved, output a template package and solidify the display of accuracy constraints.

4. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 3, characterized in that: The original time series is aligned with the start and end of puncture as anchor points. The trajectory is reconstructed using shape-preserving interpolation. The event time set of resistance disappearance and interlayer crossing is determined based on the curvature peaks and valleys of the force and displacement curves, and the same name reference is maintained with the rule network version.

5. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 4, characterized in that: Units are normalized according to the conversion chain, and the calculated and displayed values ​​are saved separately, along with the display precision and rounding rules. When encapsulating the fact package, the source identifier, formula used, unit and dimension, rounding strategy, timestamp and responsible person are written, and the version and verification code are generated.

6. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 5, characterized in that: Read the fact package and template package, calculate the predicate permission on the rule network version, and combine it with the semantic direction to form a word selection table; The display unit and the number of bits to be retained are determined according to the display precision of each data reference mark, the sum and difference are checked to be consistent, and a display strategy table is generated and written into the rendering plan.

7. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 6, characterized in that: Data reference tagging and binding are applied to the template syntax skeleton, and the integrity of references, consistency of units and dimensions, and consistency of sum and difference are checked. Pre-scoring is calculated and a list of issues is generated. At the same time, constraint triggering paths are output to locate the corresponding rule entries and syntax positions.

8. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 7, characterized in that: Constrained generation is performed according to the rendering plan, prohibiting raw numbers from being included in the text; The generated text is reverse-extracted to obtain triples of values, units, and predicates. These triples are then fed back into the rule network version for judgment and to form a set of violations, which serves as input for subsequent corrections.

9. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 8, characterized in that: Without changing the calculated value of the fact package, perform the minimum changes in sequence, including word replacement, unit display switching, fine-tuning of retained digits, and minor syntactic rewriting. After the changes are made, repeat the verification and update the final consistency score and the correction list.

10. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 9, characterized in that: A tiered issuance gating system is implemented based on the final consistency score and the number of blocking items. For approved texts, the fact package signature, rule network version, template version, issuance timestamp, and issuer identifier are written into the document metadata, and a rendering plan summary is saved for offline verification.

11. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 10, characterized in that: When data, thresholds, or templates change, the changes are mapped to a set of change sources. The change impact domain is calculated on the rule network dependency graph, and only the affected sentences are re-rendered. The re-rendering order is determined according to reachability depth, syntactic length, and constraint importance.

12. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 11, characterized in that: While generating a comparison of the old and new texts, the location of the differences and the corresponding reasons for the changes are recorded and added to the audit log; the audit log, template version, rule network version, data version, correction records and scoring trajectory are stored in chronological order.

13. The teaching system combining virtual and real methods for spinal anesthesia needle puncture according to claim 12, characterized in that: When generating the final report, the source identifier, formula used, unit and dimension, rounding rules and conversion chain summary corresponding to each data reference mark are written into the document metadata and endnote area and stored together with the document signature.

Images