Methodology changelog

This page records every change to the methodology: to the calculation, to the certificate and to the documentation itself. Older entries are never overwritten — that is what keeps a certificate issued today traceable years from now. Which version is currently in force is shown at the top of the methodology page; every certificate names the version it was calculated with.

• 2 September 2026 — The certificate states a range — this vehicle's own. Alongside the percentage there is now a WLTP-referenced range equivalent in kilometres: this vehicle's official WLTP range when new, scaled by the measured ratio of today's usable energy to the as-new figure. The as-new figure comes from the vehicle's certificate of conformity or from open EU vehicle registers for precisely this specification. On the card for a classified ad the kilometre figure may lead; on the certificate, energy retention remains the certified quantity.

  What made way for it: an older range figure that divided the measured energy by one consumption value used for every vehicle. It looked like a measurement and was a division. WLTP consumption no longer appears in the calculation at all: it sits inside the as-new range and cancels out.

  What the number is not: more precise than the percentage (it is the same measurement in a different unit), a WLTP measurement (no cycle was driven), or a forecast for everyday driving (WLTP is a standardised test cycle). Where the vehicle's specification cannot be determined unambiguously, the figure is omitted with nothing put in its place — under one variant code the registers hold up to four official ranges as much as 19 percent apart, and an average of those would describe no vehicle. The methodology version stays at v0.5; the certificate method version rises to cert-2.5.0 and cert-3.3.0 respectively, and the change is recorded as an amendment of 2 September 2026 in the preregistered study protocol.

• 1 September 2026 — The vehicle variant now also comes from the telemetry stream. Which battery pack a car was built with depends on its variant — "Long Range", "Standard" and so on. Until now we read that detail from Tesla's REST interface alone, and it answers only while the car is awake; we never wake it. A car that sleeps a lot kept us waiting accordingly. The same detail travels in the telemetry stream of newer vehicles, and from now on we take it from there while the REST answer is still outstanding.

  What this changes: a certificate can be issued sooner. What it does not change: the calculation. It is the same detail about the same car on a different route; where both are present, the REST answer still applies. Of the 30 vehicles in our fleet that deliver both, all 30 agree. Where neither is available, no certificate is issued, as before. The methodology version remains v0.5; the change is recorded in the pre-registered study protocol as an addendum dated 1 September 2026.

• 27 August 2026 — The headline figure comes from the characteristic curve rather than a projected slope (cert-3.0.0), and in the curve's edge zones the discharge side counts (cert-3.1.0) — this page explains both. Two steps on the same day:

  Step 1 — from the straight line to the curve (cert-3.0.0). Up to cert-2.3.0 the numerator was the slope of the middle measurement region, times a hundred. That assumes a battery yields the same energy per percentage point across the whole charge range. The full-swing telemetry of the test vehicle shows the opposite: at both edges the curve is flatter than in the middle, roughly 0.44 against 0.52 to 0.58 kWh per percentage point. The projection therefore overstates the energy across the full displayed span, on the test vehicle by 1.3 to 2.7 kWh.

  What is calculated now: from every interval that charging sessions and measurement runs provide, we estimate a piecewise linear, monotone curve of remaining energy over the displayed state of charge. The numerator is its integral between a displayed 0 and 100 percent. Zones without a measurement of their own carry a versioned shape template per data-channel manufacturer and cell chemistry: stored relatively rather than as a kilowatt-hour figure, with its own uncertainty in the range, and replaceable by the vehicle's own measurement. There is no correction factor on the result anywhere.

  Both quantities appear in the certificate. The curve figure up front, with a note on which part of the charge range was measured, and below it the previous calculation as a comparison line. Where the data does not support the curve calculation, the document stays at cert-2.3.0; the second quantity never prevents an issue. The issuing gate and the confidence threshold still work on the slope figure.

  What changes on this page: section 2 explains the curve, the shape template and the dual reporting; both diagrams have been redrawn. The history chart now shows what the calculation does: the published range narrows with the measurement runs and settles on a floor, instead of staying equally wide across 30 months. The methodology version stays at v0.5 — the change to the calculation itself is registered and justified as version 1.7 of the pre-registered study protocol; this page catches up with the explanation. In the same pass, an editorial change: the recurring opener "honesty requires …" has been cut from the methodology page, as has the comparative formula "X beats Y" (the section 3 heading, the shape-template rule, the literature line). Finally, the antithesis "not X but Y" has been cut from 13 occurrences to 6: the explanatory ones remain (such as "no extrapolation fills the gap. A shape template does"), the defensive ones are gone ("not as a marketing gesture, but …", "not a slogan here but …"). The statements behind each are unchanged.

  Step 2 — in the edge zones the discharge side counts (cert-3.1.0). The certificate states extractable energy — and at the edges of the charge range in particular, charging and discharging read measurably differently. An edge section therefore counts as measured only once discharging runs cover it; charging alone leaves the shape template, with its uncertainty, in place. Where both directions cover the same section, their measured difference enters the uncertainty range as a term of its own and replaces the previous blanket direction surcharge. In the middle of the charge range nothing changes; without discharge data the calculation stays as it was.

  This changes the meaning of the stated curve figure — hence a new method version rather than a silent correction. The step is registered and reasoned in the preregistered study protocol as revision 1.9, together with predictions set down in advance and their evaluation across ten vehicles. Section 2 of the methodology page adds a paragraph to the rule; the methodology version remains v0.5.

  Step 3 — the two uncertainties are labelled separately on the certificate. The document states the comparison with the as-new state; it additionally carries the uncertainty of the variant's as-new reference and is therefore wider than the vehicle's measurement precision. Until now both figures stood side by side unlabelled and read like a contradiction. The certificate now names both parts, and the usable energy carries its own measurement band in kilowatt-hours — that figure contains no reference uncertainty. Every band figure appears exactly once. The signed values and the calculation are unchanged; issuing criterion 4 (confidence) in section 2 explains what the threshold named there refers to.

• 26 August 2026 — Two additional pieces of evidence for the as-new scale: the drive variant from the vehicle identification number, and cell chemistry measured from the pack voltage. The entry of 18 August sets out when no certificate is issued: when the as-new scale is open or ambiguous. Two of those cases can be sourced rather than left open.

  The variant. Whether a car has one motor or two used to sit in the manufacturer interface's badging field alone, and that field only arrives while the car is awake. A car that streams continuously is rarely caught awake and therefore never got a variant — even though the same information sat at position 8 of the vehicle identification number on every call. The mapping is sourced against Tesla's own decoder table filed with the US traffic safety authority and holds for cars built at the US plants only; for Shanghai and Berlin that position carries something else, and there the old route still applies.

  The cell chemistry. A 2021 Model 3 base variant carries either a Panasonic pack with 96 cells in series or a CATL pack with 106 to 108. A measured pack voltage yields a cell voltage under either assumption, and at a high state of charge only one of them is physically possible. What this decides is the build, not the capacity: the number of cells in series does not change with ageing, with temperature or with the state of charge. Inferring the scale from the measured capacity would be the circular argument the 18 August entry rules out; it stays ruled out, and the calculation never sees the remaining energy in the first place. It needs at least five data points with no appreciable charge or drive current, in the window between 65% and 97% state of charge. Below that the two curves converge too closely, above it the LFP curve stands too steep — outside the window the scale stays open.

  What this changes. The 2021 Model 3 base variant (roughly 49 kWh NCA against roughly 57 kWh LFP) is no longer locked; it is decided as soon as the data points are there. Where the table already resolves a variant, or carries it as a band reference, the additional evidence leaves it alone — it applies only where nothing was produced before. The Model 3 Long Range (2019–2023) therefore keeps its band reference: both candidates share the same chemistry, and voltage does not tell them apart. Nothing changes in the calculation, and no new field appears in the signed wording of the certificate; how the scale came about is recorded alongside it.

• 20 August 2026 — v0.4 + v0.5: capacity is a slope, the unit of measurement is the measurement run, and the band no longer narrows on its own (cert-2.0.0). Prompted by an observation on our own vehicle: the stated confidence improved every single day — from ±1.1 to ±0.6 percentage points in five days — without any new charging sessions. That was not new knowledge; it was a flaw in the formula. Three changes:

  1. The numerator now measures the right quantity. Previously we computed remaining energy / (state of charge/100) per data point. That assumes remaining energy is proportional to state of charge — it is not. Below the displayed zero sits a reserve, roughly 3 kWh on the test vehicle, measured and cross-checked against a calibrated charge-point bill. The old ratio counted that reserve as usable capacity, and the lower the state of charge, the more heavily: at 30% it yielded 61.1 kWh, at 90% only 56.6 kWh — a 7.5 kWh difference from the state of charge alone. We now determine the slope of remaining energy against state of charge. It does not depend on where in the range you measure, and the reserve is reported separately instead of being counted in.

     This lowers the headline figure. On the test vehicle from roughly 95% to roughly 89% against the same denominator. The battery has not changed — the earlier figure included a reserve nobody can reach. The lower value is the more honest one.

  2. We count measurement runs, not calendar days. A measurement run is one continuous movement of the state of charge in a single direction — a charge or a drive. A thousand readings during one charge are one measurement, not a thousand. A vehicle parked for a week produces no measurement run at all: its band now stays put instead of “improving”.

  3. The band is a budget of four named parts rather than a single formula, and it no longer tends towards zero:

     • Scatter between measurement runs — the only part that falls with more data, and only through genuine measurement runs.
     • Direction — do charging and driving agree? As long as only one direction has been observed, that ignorance is carried as a range.
     • Extrapolation — the capacity figure covers the full charge range, but we only measure where the vehicle actually went. Always driving between 60 and 80% earns a wider band. This part does not shrink by measuring longer, only by using a wider range of the state of charge.
     • Source resolution — some manufacturers report whole kilowatt-hours. Turned into a slope that quickly becomes ±1 kWh, and it now sits in the band instead of claiming a precision the data cannot support.

     The figure is therefore called a model uncertainty range from now on, with a stated coverage factor of k = 1. “95% confidence interval” appears nowhere and will not be claimed until it has been checked against independent reference measurements.

  Consequence for issuing: the threshold for a certificate moves from ±1.5 to ±2.0 percentage points — not a relaxation, but because the band now contains parts that were previously missing. Vehicles with coarsely resolved data, or a very narrow range of state of charge in daily use, may no longer reach it. That is deliberate: better no figure than one that feigns precision.

• 18 August 2026 — A reference value for Model S and Model X, and no more guessed denominator. The health figure is a fraction: measured usable energy divided by what the battery held when new. For Model S and X that second value was never resolved — variant detection did not know their badges ("75D", "100D", "P100D") and fell back on a per-line approximation. Two changes; the calculation itself is unchanged (engine v0.3):

  1. The variant table now covers the pack generations of S and X. Sourced are the 75 pack (2016–2019, around 72.5 kWh usable) and the Palladium generation from model year 2022 including Plaid (around 95 kWh usable). For the 100 packs from 2016 to 2020 and the early 18650 generation (60, 70, 85, 90) the sources remain inconsistent — there the reference value stays open.
  2. Where the reference value is open or ambiguous, no certificate is issued any more. Previously a per-line approximation was substituted and labelled as such on the certificate. That was not enough: too large a denominator makes a battery look worse than it is, and in a signed report that is the wrong error to make. Besides the S/X variants named above this also affects Model 3 Long Range from 2019 to 2023, where two different cell packs shipped under the same badge. Anyone affected now sees this in their battery status; the free history keeps running unchanged. Certificates already issued remain verifiable and keep their provenance note.
  3. New: the band reference. Where a model variant carries two sourced battery packs of the same size class that cannot be told apart under the same name, we compute against their midpoint and add the spread to the stated uncertainty. For the Model 3 Long Range (2019–2023) that is roughly 75 and roughly 79 kWh usable, so 77 as the reference; the absolute band grows by about one percentage point. The certificate spells this case out — on the page and in the PDF — because a reference figure that no single car had exactly deserves to be named. We explicitly do not claim which pack is fitted: inferring the pack from the measurement and then computing against it would be circular reasoning that makes every battery look healthier. We do not average across different pack sizes — the 2021 Model 3 base variant (roughly 49 kWh NCA versus roughly 57 kWh LFP) stays blocked, as there is no meaningful midpoint there.

  ---

  20 August 2026 — v0.5: batteries get worse, and we now account for it (cert-2.1.0). Until now, every measurement from a vehicle was combined as though it described the same state. It does not: a measurement from a year ago describes the battery as it was a year ago, and it was better then. The figure we published was therefore not today's value but a blend of past and present — with a year of history, roughly four months stale.

  What changed:

  1. The trend is now calculated as a trend. A straight line is fitted through the measurement runs over time and over distance travelled; the value published is the one at the current end of that line. The decline itself becomes a statement in its own right: how much capacity is lost per year and per 10,000 km.
  2. Scatter now measures only measurement scatter. Previously the capacity decline was itself part of the scatter and inflated the uncertainty range. We now measure deviation from the trend line.
  3. Three new checks inside the uncertainty range. If the time-based and distance-based trends contradict each other, the range widens. If the calculation says the battery has become better, the range widens too — because that cannot happen, and it points to the vehicle system having recalibrated. We deliberately do not discard this case; we disclose it. And if no new measurements arrive for a while, uncertainty about today's state grows.

  What this means for published figures: for vehicles with a long history, the published value drops slightly — it is now today's value rather than an average across months. With a short history, almost nothing changes. The uncertainty range becomes somewhat wider for long histories; that is intended, because the narrower figure did not cover the actual error.

  ---

  Same-day addendum (cert-2.2.0): the certificate now shows the trend as well. New on the document and the status page: the measured decline in kWh per year and per 10,000 km — but only for the axis actually measured, never as a zero. Alongside it, one sentence on what the stated uncertainty depends on and what would improve it; for most vehicles that is the charge range used, not waiting time. A calculation error went with it: for vehicles with plenty of time but few kilometres, the document could previously show "−0.00 kWh per 10,000 km" as a trend — a figure the calculation itself had refused to produce. It no longer appears at all.

  Second same-day addendum: the reference figure now sits on the same basis everywhere. The health figure is a fraction. Since v0.4 the numerator measures only the energy between the displayed 0 and 100 per cent — the reserve below the displayed zero no longer counts. The denominator, however, still counted it on the free surfaces: there it was the gross figure for the model line. Usable divided by gross gives a figure that is too low — on the test vehicle 88.9 instead of 93.6 per cent. Around 4.7 of the roughly 6 lost points were this basis error, not the condition of the battery.

  Previously two errors roughly cancelled each other out: as long as the numerator counted the reserve, the gross denominator matched it well enough. Now that the numerator is clean, the denominator is wrong on its own.

  The certificate was never affected. It has been calculated against the resolved usable reference figure since 27 July. The signed headline figure does not move as a result of this change, and certificates already issued remain verifiable unchanged. What we brought into line are the battery status, the purchase page, the check before issuing, and the study export.

  The reference figure is now the one UN GTR 22 uses for SOCE everywhere: measured usable energy divided by the usable energy when new — both without the hidden reserve.

  Intended side effect: the uncertainty band is now expressed against the same, smaller denominator and is therefore about five per cent wider. The issuing criterion (±2.0 points) was previously easier to meet by exactly that margin.

• 16 August 2026 — v0.2 + v0.3: measurement window instead of discarding, weighted span, one energy source per calculation, confidence from independent measurement days. Shipped the same evening in two steps (v0.2, shortly after v0.3) — recorded here as one change. It is the first version bump since v0.1. Six changes, five of them to the calculation:

  1. A charge that runs past the usable SoC limit (such as a full charge) is no longer discarded entirely; instead its measurable part is used and the capacity is derived from that window. Charging to full is explicitly fine — nobody needs to change their charging habits for the Akkubrief.
  2. The minimum span of a charge drops from 25 to 15 percentage points; in return, every charge now weighs into the weighted median proportionally to its span. Wide charging windows carry more information; short ones are dampened rather than hard-rejected (cert-1.4.0).
  3. Each calculation uses exactly one energy source (directly measured stored energy preferred, otherwise the charge counter) — two different yardsticks in one calculation could otherwise appear as spurious degradation. The source used is stated in the certificate (cert-1.5.0).
  4. The confidence interval counts independent measurement days instead of data rows. A parked car delivers hundreds of readings of the same vehicle estimate — that is one measurement situation, not hundreds of measurements. Previously the band could become arbitrarily narrow (±0.1 percentage points); now it widens to an honest magnitude. Likewise, with very few charging sessions an outlier visibly widens the band instead of vanishing inside the robust statistics. The issuance criterion "confidence ≤ ±1.5 points" is unchanged — it just finally measures something.
  5. The measurement basis is source-pure: points from the range back-calculation and directly measured energy values never enter one calculation together. Once enough directly measured points exist, only those count — including for the minimum period, which may therefore temporarily restart: a period meant to confirm the value must consist of the same measured quantity.
  6. Made visible on the certificate: the provenance marker stored since 9 August is now actually displayed on the certificate page and in the PDF (source class, quality tier, whether measured independently of the vehicle's own estimate). Newly stated: the DC fast-charging share of the recorded charging energy (cert-1.6.0) and the mileage trend — the latter only when the mileage span of the data points supports a reliable statement. Fixed beforehand: data points without an odometer reading entered the trend calculation as "0 km" and could massively overstate the span; such points now inherit the last known reading or stay out.

• 9 August 2026 — Provenance marker for the measured energy (cert-1.3.0). Every certificate now states where its energy figure came from: a measured charging session, a manufacturer's remaining-energy reading, or the vehicle's range display converted back into kWh. The document also records whether the figure is a measurement of our own or merely reproduces the vehicle's own estimate — a distinction the percentage alone cannot show. Where a conversion was involved, the constant used is named, along with whether it has been calibrated on the vehicle. The calculation itself is unchanged.

• 27 July 2026 — Honest denominator + two bands (cert-1.2.0). The reference figure is now called "reference energy when new" instead of "certified energy" — nothing about it is certified. Where the vehicle variant resolves unambiguously, the percentage is computed against the variant's documented usable new value (source: specialist databases/teardowns; the makers publish no kWh figures). Where the variant is ambiguous (the same model shipped different battery packs) or unknown, the coarse model-line value is used and flagged as such — no silent fallback. Two bands are new: a tight one for trend/comparison (measurement precision only) and a wider one for the absolute claim that honestly carries the denominator uncertainty. The capacity estimate itself is unchanged — still v0.1.

• 25 July 2026 — Grade and Euro 7 line removed from the certificate, data section corrected, research evidence added. Three changes, none of them to the calculation — still v0.1:

  1. The A/B/C/D grade is gone. It was arithmetically untenable next to an honest uncertainty band (at ±3 points no vehicle can be unambiguously "B" or "C"), and a fleet-position alternative needs a peer cohort that does not exist yet. The Euro 7 reference went with it: it applies to new type approvals from 29 Nov 2026 and new vehicles from 29 Nov 2027, does not apply to the used vehicles we certify, and in the chart it read as pass/fail. Both are gone from the certificate, the PDF and the listing card; the Euro 7 context remains on this methodology page. The trajectory chart goes with them, because it drew a straight line from the as-new state at 0 km to today — a path we never measured.
  2. Correction to the section "How the data reaches us": it described Fleet Telemetry as the preferred data path and charging-session reconstruction as the fallback. Neither was accurate — in reality only periodic status reads are used, and energy is converted from the vehicle's range estimate. Section corrected and extended to state where the energy figure comes from. New: the energy added during charging is now recorded (not yet used in the calculation).
  3. The section "Scientific basis & standards" gained research evidence (field-SoH accuracy, partial-charge formula, robust statistics, trajectory; with sources).

• 24 July 2026 — Added section "Scientific basis & standards" (context on UN GTR 22 / SoCE and Euro 7, scientific sources). The calculation itself is unchanged — still v0.1.

• 23 July 2026 — Issuance criteria published (data-quality gate: usable readings, charging activity, minimum span of 5 days, confidence ≤ ±1.5 percentage points). The calculation itself is unchanged — still v0.1.

• v0.1 (19 July 2026) — First release: charging-session/telemetry path, robust statistics (weighted median, MAD, Theil–Sen), range output, Ed25519 signature with client-side verification page.