Calibration & Unit Normalization in Biosignal Processing, biosignal_, creates a derived waveform using explicit affine coefficients. Start with the values and units actually loaded. A successful inverse check does not establish that those coefficients describe the measuring device.
1. Establish what the loaded values represent
Load the recording and select the intended channels before choosing calibration parameters. Retain the original file, its metadata and the source of your coefficients. Determine whether the reader has already applied header scaling: a parameter named “gain” is not a reason to apply a second conversion to values already expressed in physical units.
| Inspect before Run | Keep with the result |
|---|---|
| Loaded value domain | Document whether the samples are digital counts, physical values or another declared representation. The app applies its equation to the loaded values; it does not infer a device conversion from waveform magnitude. |
| Channel identity and order | Use exact selected channel names for channel-specific parameters. Unknown names are refused. The resolved coefficient table follows the selected recording’s channel order, even when JSON keys appear in another order. |
| Recorded and asserted units | Keep recorded input, asserted input and output labels separate. The default conflict policy is reject; an explicit override is a recorded assertion, not a proof that the new label is correct. |
| Timing and source identity | Retain sampling rate, time origin, source fingerprint and any events. Calibrating sample values does not establish event-detection accuracy or provide a new timing reference. |
2. Resolve coefficients for every selected channel
The app’s equation is y = gain × x + offset. Gain and offset must be finite; each resolved gain must be nonzero. A negative gain reverses numerical polarity. A zero gain is refused because the original values could not be recovered by the inverse equation.
- Choose the calibration mode. In
manualmode, the visible gain/offset controls provide defaults;channel_in advanced JSON can supply named per-channel values.calibration - For a channel using endpoints, supply all four:
digital_,minimum digital_,maximum physical_andminimum physical_. The digital endpoints must differ. Do not mix endpoint fields with gain/offset fields for that channel.maximum - Read the resolved coefficient table. Endpoint gain is the physical span divided by the digital span; offset aligns the lower named digital endpoint with its corresponding physical endpoint. Explicit channel coefficients take precedence over the global defaults.
- If using a source-bound calibration contract, select
source_mode and supply the JSON through Source-bound calibration contract. The schema binds its channels and evidence digest to the loaded recording fingerprint. Do not also supply a separate channel-calibration mapping.bound_ contract
A matching evidence digest records which document was used. It does not authenticate a calibration certificate or establish metrological traceability. Choose workflow sections for the outputs you need; their required dependencies may add further result tables.
3. Keep numerical scaling separate from unit claims
Changing an output label alone does not rescale values. With gain 1 and offset 0, the app records a numerical identity transform even if the labels differ. The unit table explicitly retains unit_.
The app can screen an asserted UCUM label for basic character/whitespace conditions. It does not implement a dimensional UCUM parser or establish equivalence between differently written units. The UCUM specification defines both unit-expression syntax and semantics; a basic lexical screen is a narrower check.
An inverse round-trip checks numerical recovery using the supplied coefficients. Its scale-normalized error is the largest absolute reconstruction error divided by the greater of 1 and the largest absolute finite input value. This is not a per-sample relative-error percentage or a measurement of device accuracy. With no finite values, the check is unavailable.
Uncertainty routes use declared assumptions. A default zero uncertainty is not evidence of a perfect instrument. The retained examples below do not validate uncertainty propagation, certificate authenticity, physical unit semantics or downstream clinical use.
4. Read two retained refusal/recovery examples
Both checks reuse one synthetic fixture: 385 samples at 128 Hz, with channels named maternal and abdominal, both labeled mV. These are test labels and impulses, not an acquired maternal/fetal study or two independent datasets. Each retained sequence loaded once, attempted a refused Run, corrected one setting, ran again and exported once.
| Retained sequence | Refusal and saved recovery |
|---|---|
| Global gain | Gain 0 was refused with Export disabled and Run available again. Changing only gain to −2, with offset 1, produced y = −2x + 1 for both channels. The saved round-trip table has two PASS rows with zero error; both polarity flags are true. This bundle contains 23 files, including its manifest. |
| Mixed channel coefficients | Equal digital endpoints for maternal were refused. Changing its digital maximum from 0 to 1 resolved endpoints 0/1 → 3/5 to gain 2, offset 3. abdominal used explicit gain 4, offset −2; global controls remained 7 and 9. This bundle contains 22 files, including its manifest. |
The second endpoint table contains two channel rows, but only one endpoint-derived channel. Its endpoint metric is 1. The four endpoint cells for the gain/offset channel remain empty in CSV; those empty cells are not zeros or missing calibration failures. The selected sections require nine routes in the first sequence and eight in the second. These dependency counts are not independent scientific cases.
The retained reviews compared all 770 calibrated values and their inverses, plus all 385 times, in each sequence. This lesson reopens the saved coefficients, dispositions and complete table structures; it does not repeat those numerical analyses. Neither retained sequence supplies a visual-layout review of the product GUI.
5. Save the complete bundle and inspect its scope
Use Export once the intended result is available. The source-defined exporter writes a separate run directory, verifies its staged members and refuses to overwrite an existing bundle. File counts depend on the selected routes and available results; 23 and 22 are the counts of these two saved examples.
| Saved material | What it establishes |
|---|---|
| Calibrated waveform and channel map | The full saved waveform table retains 385 rows and both named channels in each example. Keep its time column and the channel map containing output units, gain, offset and value-column names. |
| Coefficients, units and route tables | Read coefficient source, recorded/asserted/output units, route status, required dependencies and unavailable outputs. A compact metric count and a published table’s row count can describe different things, as the endpoint example shows. |
| Input and parameter documents | Retain the input receipt and parameter manifest, including channel-specific values and workflow choices. These documents identify the result configuration; they do not replace the original recording and its acquisition metadata. |
| Bundle manifest and review documents | Keep every listed member and its hash. The saved validation-scope table keeps preparation disposition proposal; the exact-result review document keeps decision unreviewed. A successful export does not promote either status. |
6. Keep preparation review and result review distinct
A calibrated layer can become accepted preparation only through its matching source- and derived-layer-bound accepting review. The separate Result disposition control requires an exact result-review JSON with an accepting decision before accepting that result. Neither control supplies external scientific validation. Both saved examples retain proposal/unreviewed states.
The viewer may show only a bounded selection of channels or points. Inspect the full numerical export and its channel map for completeness. Keep the original data separately from the derived waveform, and carry forward any unresolved unit, evidence or review limitation.
7. Read three later contract and uncertainty examples
Three additional native refusal/recovery sequences were reviewed locally on 10 November 2026. They reuse the same 385-sample, 128 Hz synthetic fixture as the two earlier examples. The channel names maternal and abdominal are test labels; these are not acquired patient recordings or three independent datasets. The tables below transpose each saved two-channel CSV so that every field remains visible. Their exact CSV columns and strings remain in the evidence download.
Keep selected routes and required dependencies distinct
| Sequence | Selected routes | Required routes | Files including manifest |
|---|---|---|---|
| Source-bound contract | 3 | 7 | 21 |
| Declared uncertainty | 2 | 8 | 21 |
| Conflicting unit label | 2 | 7 | 21 |
Each sequence loaded the waveform once, attempted one refused Run, corrected its specified input, ran again and exported once. The source-contract sequence also loaded two auxiliary contract documents. These route counts and overlapping assertions do not represent independent scientific cases. All three exports retain preparation disposition proposal and result-review decision unreviewed.
Bind the contract to the loaded source, not a selected-view fingerprint
The first contract incorrectly used the selected-recording fingerprint. The app refused it because the contract fingerprint did not match the loaded source recording. Replacing only that fingerprint in the next loaded contract allowed the retained recovery. The visible gain 7 and offset 9 remained unchanged; source-bound coefficients resolved separately to maternal gain 3, offset 1 and abdominal gain −2, offset 4.
| Saved field | maternal | abdominal |
|---|---|---|
channel_index | 1 | 2 |
| channel | maternal | abdominal |
| gain | 3.0 | -2.0 |
| offset | 1.0 | 4.0 |
coefficient_source | source_bound_contract:channel_gain_offset | source_bound_contract:channel_gain_offset |
calibration_reason_code | synthetic_maternal_pair | synthetic_abdominal_pair |
The coefficient-source strings identify the source-bound channel mappings. The following complete source table retains the evidence digest and both authenticity flags. The label laboratory_calibration_record was part of this synthetic test contract; it does not establish a genuine laboratory certificate.
| Saved field | maternal | abdominal |
|---|---|---|
channel_index | 1 | 2 |
| channel | maternal | abdominal |
coefficient_source | source_bound_contract:channel_gain_offset | source_bound_contract:channel_gain_offset |
calibration_reason_code | synthetic_maternal_pair | synthetic_abdominal_pair |
calibration_mode | source_bound_contract | source_bound_contract |
evidence_basis | laboratory_calibration_record | laboratory_calibration_record |
evidence_sha256 | 91221daf | 91221daf |
evidence_authenticity_validated | False | False |
The retained review compared the full 770 calibrated values and inverses and all 385 times. This addition checks saved identities and fields without repeating those calculations. A matching source fingerprint or evidence digest establishes document binding; neither authenticates the device, certificate or physical unit interpretation.
Retain covariance and input uncertainty explicitly
The uncertainty sequence used manual gain 2 and offset 3, gain and offset standard uncertainties of 1, input standard uncertainty 0.5 and coverage factor 2. Covariance −2 was refused by the positive-semidefinite uncertainty contract. Changing only the visible covariance to −1 produced the saved recovery. This is a boundary case of the declared first-order model, not evidence that the assumptions fit a real instrument.
| Saved field | maternal | abdominal |
|---|---|---|
channel_index | 1 | 2 |
| channel | maternal | abdominal |
gain_standard_uncertainty | 1.0 | 1.0 |
offset_standard_uncertainty | 1.0 | 1.0 |
gain_offset_covariance | -1.0 | -1.0 |
input_standard_uncertainty | 0.5 | 0.5 |
output_unit | mV | mV |
coefficient_covariance_assumed_zero | False | False |
input_value_uncertainty_included | True | True |
input_coefficient_covariance_assumed_zero | True | True |
coverage_factor | 2.0 | 2.0 |
method_id | jcgm-gum-first-order-affine-propagation-v1 | jcgm-gum-first-order-affine-propagation-v1 |
The two false/true covariance fields have different scopes: coefficient covariance was supplied, while input-to-coefficient covariance was assumed zero. Input-value uncertainty was included. Preserve these distinctions when reading the full summary below.
| Saved field | maternal | abdominal |
|---|---|---|
channel_index | 1 | 2 |
| channel | maternal | abdominal |
gain_standard_uncertainty | 1.0 | 1.0 |
offset_standard_uncertainty | 1.0 | 1.0 |
gain_offset_covariance | -1.0 | -1.0 |
input_standard_uncertainty | 0.5 | 0.5 |
assumed_independent_coefficients | False | False |
input_value_uncertainty_included | True | True |
input_coefficient_covariance_included | False | False |
coverage_factor | 2.0 | 2.0 |
minimum_propagated_standard_uncertainty | 1.0 | 1.118033988749895 |
median_propagated_standard_uncertainty | 1.4142135623730951 | 1.4142135623730951 |
maximum_propagated_standard_uncertainty | 1.4142135623730951 | 1.4142135623730951 |
minimum_expanded_uncertainty | 2.0 | 2.23606797749979 |
median_expanded_uncertainty | 2.8284271247461903 | 2.8284271247461903 |
maximum_expanded_uncertainty | 2.8284271247461903 | 2.8284271247461903 |
output_unit | mV | mV |
method_id | jcgm-gum-first-order-affine-propagation-v1 | jcgm-gum-first-order-affine-propagation-v1 |
The retained independent review checked twelve minimum/median/maximum summary values and 48 separate scalar, flag and string slots. Those observations overlap the same two-channel sequence. The first-order reference excluded the higher-order product of coefficient error and input error; this example does not establish the physical adequacy of that approximation.
This export contains uncertainty contracts and summaries, but no calibrated-waveform table or per-sample uncertainty table. A summary’s minimum, median and maximum do not constitute a full uncertainty trace. The expanded values use the declared factor 2; this guide does not assign them a confidence or coverage probability. The saved mV label is retained without certifying physical unit semantics.
Correct a conflicting label without claiming unit conversion
The third sequence asserted input unit V against the fixture’s recorded mV label under the reject policy. The app refused it. Only the visible input-unit label changed to mV for recovery; gain 1, offset 0, output mV and the reject policy stayed unchanged. The retained output was the numerical identity transform, not a volts-to-millivolts conversion.
| Saved field | maternal | abdominal |
|---|---|---|
channel_index | 1 | 2 |
| channel | maternal | abdominal |
recorded_input_unit | mV | mV |
asserted_input_unit | mV | mV |
output_unit | mV | mV |
unit_override | False | False |
unit_code_system | source_declared | source_declared |
ucum_lexical_check_passed | False | False |
unit_dimensional_semantics_validated | False | False |
unit_conflict_policy | reject | reject |
The false UCUM lexical flag reflects the saved source_declared code system; it is not a dimensional unit validation result. No override was exercised, and dimensional semantics remained false. The earlier independent review checked the complete 770 identity waveform values, 385 times and twenty saved unit fields. This addition does not repeat that numerical reference.
Keep the source and export limits attached
In this addition’s dated inspection, the calculation app, catalog, exporter and three supporting files match all three recorded payload bindings. The app retains the 7fecec73… identity. The current original shared UI module differs from the recorded module, so these retained native results do not accept the complete current dependency tree.
The saved-file review covers 63 export files and 51 complete CSV schemas across these three sequences. It adds no scientific run, product GUI action, export, recording or graphical-layout review. Certificate authenticity, metrological adequacy, clinical use, general cancellation behavior, hosted execution and whole-method acceptance remain outside the evidence.
Download the complete saved tables, route inventories and dated source relationships. The two earlier examples and all prior failed attempts retain their original evidence.