The grid rules

§ 14a EnWG, the EEBUS limitation state machine, § 9 EEG, Modul 3, MiSpeL and § 42c — as code, with the citation for every number.

Every rule in hems-grid names the document it comes from. [A1 4.5.2] is Anlage 1 to BNetzA decision BK6-22-300; [LPC-031] is the EEBUS Limitation of Power Consumption use-case specification. A rule without a citation is a bug, and cargo xtask check-citations fails the build if a cited document is not one the reader can obtain.

Several of them are not implemented here at all. The § 14a minimum power, the netzwirksamer Leistungsbezug, the Modul 3 conformance rules, the VDE-AR-N 4100 Unsymmetrieleistung and the allocation identity § 42b and § 42c settle on belong to metering, which owns quantities where hems owns control. Two enumerations of one regulation are two things that can disagree about it.

§ 14a EnWG — which devices, and how far down

Three questions, three functions.

Is this a steuerbare Verbrauchseinrichtung?

[A1 2.4.1] names four Fallgruppen: a non-public charge point, a heat pump heater including its auxiliary and emergency heaters, space cooling, and electricity storage with respect to charging. Each above 4,2 kW.

The trap is [A1 2.4.2]: heat pumps and cooling are summed per Fallgruppe behind one connection before the threshold is applied, and if the sum passes it the whole group counts as a single controllable device. Two 3 kW heat pumps are one 6 kW device, not two devices below the threshold — a distinction worth 4,2 kW of minimum power. Charge points and storage are counted individually.

Under which regime?

[A1 10] is four transitional regimes at once. An old device on the old reduced network fee keeps it until 31.12.2028 and moves to the new rules on 01.01.2029. A night-storage heater keeps the old rule with no end date at all. An old device that never had a reduction is simply outside the Festlegung. And any of them may switch voluntarily at any time — the network operator cannot refuse, and there is no way back.

How far down may it be turned?

P_min = base + (n − 1) × GZF(n) × 4,2 kW

base = max(0,4 × ΣP_Wärmepumpe ; 0,4 × ΣP_Raumkühlung)  if a heat-pump or
                                                          cooling group above
                                                          11 kW is included
     = 4,2 kW                                             otherwise

with the Gleichzeitigkeitsfaktor 0,80 / 0,75 / 0,70 / 0,65 / 0,60 / 0,55 / 0,50 for two to eight devices and 0,45 from nine on.

Note the max, not a sum: heating and cooling do not run at full power at the same time, and adding them would hand out a minimum the network never has to grant.

The netzwirksamer Leistungsbezug — why a roof lifts the ceiling

[A1 2.3] defines the limit as applying not to what the controllable devices draw, but to the part of the grid draw they cause. While the photovoltaic system produces more than the rest of the house needs, its surplus covers the wallbox, and the wallbox can keep charging above the operator’s limit without a single watt of it being netzwirksam.

That is the economic case for owning an energy manager under § 14a, and hems computes it in both places it matters: in the planner against the forecast, and in the guard once a second against measurements.

budget = ceiling + max(0, production − other load)

flowchart LR
  PV["roof
4,0 kW"] --> N{"surplus?"}
  HH["household load
1,2 kW"] --> N
  N -- "2,8 kW spare" --> B["budget for the
controllable devices"]
  BAT["battery discharging
2,6 kW, sustainable"] --> B
  C["§ 14a ceiling
4,2 kW"] --> B
  B --> R["9,6 kW the wallbox and the
heat pump may share, lawfully"]

Electrons carry no labels, so the split between controllable and other load has to be chosen — and the Festlegung does not choose. It defines the quantity as “derjenige Anteil … der zeitgleich durch eine oder mehrere steuerbare Verbrauchseinrichtungen verursacht wird” and stops there, so which part of the remaining grid draw the controllable devices caused is an apportionment nobody performed for us. hems therefore takes a Verursachungsregel rather than picking one silently:

Both are defensible. The default can never overstate the headroom, so a guard built on it errs early rather than late; a household whose network operator has agreed the pro-rata reading in its Technische Mindestanforderungen can say so, and nobody else should.

A discharging battery is production too

[A1 2.3] measures what the controllable devices draw from the grid, so energy that came out of a store never crossed the connection point and never counted. A household with a full battery may therefore charge its car under a teatime reduction, and reading the ceiling as a limit on consumption instead leaves the car short on the one evening a battery exists for.

A discharge is not a roof, though: the same tick that spends it may be about to reverse it. So the guard lends only what survives three tests — it is measured, this tick is about to ask for it too, and the store can sustain it for a whole control period — and pins the lender’s ceiling at zero in exchange. The battery may go on discharging or it may stop; what it may not do is turn into a load while the others spend the headroom it created.

On the reference January evening that is 5,4 kWh of headroom, and the difference between a car that leaves full and one that does not.

The same three tests run the other way for § 9 EEG, whose ceiling is on what leaves rather than on what is drawn. The household’s own consumption is headroom there — § 9 Abs. 2 measures at the Verknüpfungspunkt — and the guard lends it to the inverter, minus the part it could not call back before its next tick: a draw it is commanding away this second, a draw it only guessed from a nameplate, and the share of a thermostat’s draw that lend_window says is longer than one control period away. A heat pump answering its own safety limit drops a kilowatt nobody asked it to; without the deduction that kilowatt is already spent on the roof.

Two further readings go the conservative way, and both are places where the tempting default is the unsafe one:

• a controllable device the drivers cannot hear from is assumed to be drawing its nameplate power, never nothing — otherwise the guard hands its share of the budget to the others while it keeps charging;
• a silent inverter, by contrast, is assumed to be producing nothing, because the alternative would raise every budget on the site.

The verdict records which devices had to be assumed, so a compliance figure computed from an assumption never looks like one computed from a meter.

A nameplate assumption is safe in one direction only, and that is worth saying separately. Guessing that a silent device is drawing its rated power keeps the § 14a budget small, which is the right way to be wrong. Reusing the same guess as headroom for feeding in turns it inside out — a house with three quiet devices would be told it may produce nineteen kilowatts above the § 9 EEG cap because it is probably using them — so the export side counts only what a meter actually saw.

And a small device is not a controllable one

A device is a steuerbare Verbrauchseinrichtung only if it passes the 4,2 kW of [A1 2.4.1] — individually for a charge point and a battery, summed per Fallgruppe for heat pumps, so two 3 kW heat pumps behind one connection are one 6 kW controllable device. A 3 kW heat pump on its own, or a 0,5 kW battery, is ordinary load: a network operator may not reduce it, and it never appears against the ceiling. What it does instead is spend the surplus, like the hot-water tank and the dishwasher.

That is a fact about a site’s nameplates and its commissioning dates, and getting it wrong is expensive in both directions. Counting a device in when it is out leaves the household colder than the Festlegung asks for on the winter evenings a reduction actually happens; counting it out when it is in overstates the surplus and therefore understates the netzwirksamer Leistungsbezug, which is the one direction a compliance record may never be wrong in.

So there is exactly one function that answers it — hems_grid::classify_at, from the site’s own assets and the day — and everything asks it: the guard every tick, the planner through PlanningLimits::steuve_devices, and the evidence record when it writes down what crossed the connection point.

Where the § 14a limit is not the only shared one

A § 14a ceiling is spent only by the controllable devices. The main fuse in both directions, each sub-distribution board, the 4,6 kVA of unbalanced load VDE-AR-N 4100 allows, and the § 9 EEG feed-in cap are spent by everything behind them, uncontrollable load included.

Each goes through one weighted max-min allocator and the results intersect — the mechanism is in Architecture. What matters here is that a limit spent by a set of devices has to be shared between them: bounding each asset by it individually lets a roof allowed 5,88 kW and a battery allowed 5,88 kW put 11,76 kW through a limit of 5,88.

Which devices the symmetry rule reaches

VDE-AR-N 4100 Abschnitt 5.5.2 caps the Unsymmetrieleistung of a customer installation at 4,6 kVA — the same limit as 20 A per Außenleiter, and derived from EN 50160’s 2 % cap on voltage unbalance.

It does not apply to everything behind the connection, which is the reading that looks obvious. The VDE FNN Hinweis is explicit: “Die Anforderungen zum symmetrischen Betrieb gelten nur für Geräte die elektrische Energie einspeisen oder speichern können, also Erzeugungsanlagen, Speicher, Ladeeinrichtungen für Elektrofahrzeuge.” Generation, storage and charging equipment; not a heat pump, not a hot-water tank, not the household’s own single-phase load.

Summing everything would put a kettle on L1 against the budget of the one device the manager can actually move. A limit applied more widely than the rule is still applied wrongly, and here the household pays for it in charging power.

One caveat. The limit is stated in kVA and hems computes it from active power, because that is what a household driver reports — so it understates the unbalance exactly when the grid has asked an inverter for reactive support under VDE-AR-N 4105. A meter that reports apparent power per conductor is used as it stands.

The EEBUS limitation machine

A § 14a limit reaches the energy manager over EEBUS, from the FNN Steuerbox. Five states, three interacting timers, and one rule that is routinely implemented backwards.

stateDiagram-v2
  [*] --> init : restart
  init --> unlimited_controlled : heartbeat resumes,
no limit in force
  init --> limited : heartbeat resumes,
a limit is in force
  unlimited_controlled --> limited : Energy Guard writes a limit
  limited --> unlimited_controlled : limit released, or its
duration expires [LPC-909]
  unlimited_controlled --> failsafe : heartbeat missed for 120 s
  limited --> failsafe : heartbeat missed for 120 s
  failsafe --> limited : contact returns,
a limit is in force
  failsafe --> unlimited_controlled : contact returns
  failsafe --> unlimited_autonomous : FailsafeDurationMinimum
expires [LPC-922]
  unlimited_autonomous --> unlimited_controlled : contact returns
  note right of init : a restart comes back limited, not free
  note right of unlimited_autonomous : a broken control box must not
block a household for ever

• The heartbeat goes both ways at least every 60 s ([LPC-005], [LPC-006]).
• Missing it for 120 s means the Energy Guard is gone.
• A restart comes back limited, not free ([LPC-901/1]) — a heat pump that reboots during a grid emergency must not return at full power.
• The Failsafe Duration Minimum (2–24 h, [LPC-022]) is not a safety timer that keeps the device down. It is a release valve: once it expires with the heartbeat still missing, the device goes unlimited ([LPC-922]), because a broken control box must not block a household for ever.

That last one is what an implementation written from intuition gets wrong, and it has a test named after it. The machine itself lives in the eebus crate and hems derives its state from it rather than tracking one alongside — see devices and drivers.

The machine is also explored exhaustively, not only sampled: a breadth-first search over a finite event alphabet visits every reachable state (deduplicated on the timing differences the machine actually compares), checks four invariants in each, and quantifies the release-on-silence bound over all of them. Its first run found what two thousand random steps never had: one heartbeat in unlimited/autonomous used to open the write gate for ever, so a control box that heartbeated once and died could limit the household days later with a single delayed write. A write now counts only within sixty seconds of a heartbeat — the implementation guide’s own window, and the same gate the eebus crate’s certifiable machine runs.

It explores the machine a box actually runs, and so do the reference days: two implementations of a certifiable state machine disagree, and the one that is wrong is whichever the laboratory is not looking at. The time grains land on each constant the specification names as well as either side of it, because a boundary the alphabet steps over is one no test can see — “no heartbeat for 120 s” means the release happens on the stroke of 120, and that second is where a certification laboratory measures.

init and failsafe are also not § 14a events. The limit in force there is the device’s own preconfigured value, applied because nothing is talking to it — the manager restraining itself, not an operator reducing the house. The evidence record counts the two apart, because reporting the second as the first tells a household the operator intervened on a day when nobody did.

§ 9 EEG — two caps with different references

Two limits apply to feed-in and they are fractions of different things:

• the 60 % cap of the Solarspitzengesetz (in force 25.02.2025) is a fraction of installed capacity, and applies until an intelligent metering system with a control device is in operation;
• the EEBUS feed-in factor ([MGCP-011]) is a fraction of the cumulated nominal AC power of the inverters: P_feed-in ≤ factor × Σ P_PV,AC,nom.

Those differ on almost every real installation, because inverters are routinely undersized against the modules — the reference household is 9,8 kWp of module behind 8 kW of inverter, so 60 % of one is 5,88 kW and 50 % of the other is 4 kW. Taking the smaller of the two results is the only reading that satisfies both, and there is a third arm beside them: an operator’s live LPP session.

The factor arrives over EEBUS MGCP scenario 1, where hems is the Monitoring Appliance — it binds and subscribes to the grid connection point’s DeviceConfiguration and reads the factor; it never writes one, because the factor is set by whoever configures the connection. What crosses the wire is a percentage, and the sum it applies to is a property of the building that no EEBUS message carries: only the two together are a number of watts. So the driver reports the percentage and the site turns it into a ceiling, because the driver does not know the roof.

The cap comes from the site, not from a wire. § 9 Abs. 1 applies to the plant by force of law: no operator sends it, no session announces it, and a box that waited to be told would leave an uncontrolled roof feeding in without limit. So the guard derives it from the household’s own Para9Status on every tick, and the planner derives the same ceiling from the same call — a plan made without it runs the roof at its rating and then watches the guard curtail it every sunny midday. An operator asking for less through an LPP session still wins, and the reason chain says which of the two it was, because a household told “§ 9 EEG” for an operator’s intervention has been told the wrong thing about its own bill. A plant that has passed its Ansteuerbarkeit test is not capped at all.

Both are measured at the connection point, which is where the statute measures them, so what the house is using is headroom: a household drawing 4 kW may produce 4 kW above the cap and still feed in exactly the cap. Applied per inverter instead, a house consuming 3 kW would have its roof curtailed as though it consumed nothing and throw away exactly those 3 kW every sunny hour.

Both limits are enforced on power, on every tick. § 14a is a control instruction with a five-minute response presumption [A1 4.2], so the guard may never be over it even for a tick; § 9 Abs. 2 says Leistung, so the same is true of the 60 % cap. Reading that one as a quarter-hour average — which the register in the day report might suggest — would let a roof sit above a statutory line for minutes at a time on the strength of a forecast about the rest of the slot. The register is a diagnostic; the statute says Leistung.

What differs between them is not the resolution but what each measures. § 14a bounds the netzwirksamer Leistungsbezug of the controllable devices, so a dishwasher’s draw is not headroom for a wallbox. § 9 EEG bounds the Einspeisung at the Verknüpfungspunkt, so everything the house is using is headroom, whoever is using it — and curtailing a roof as though the car were not charging throws away real kilowatt-hours for nothing.

just demo capped runs a clear May day on a 20 kWp roof with a small store — May, not June, because the cap is a fraction of direct-current power and how close a roof gets to that fraction is decided by cell temperature — and prints the quarter-hour feed-in peak against the ceiling. The cap binds, at 12,01 of 12,00 kW around solar noon, and the household loses 2,4 kWh of export — of which 2,3 kWh is the price of the one-minute control period the simulated day runs at rather than the one second a box runs at. The guard lends the roof the household’s own consumption as feed-in headroom, which § 9 Abs. 2 EEG says it may, but only the part it could call back before the next tick; the same day at a five-second period curtails 0,1 kWh and reaches 11,99 kW.

That is far less than “60 %” sounds, and the arithmetic is worth stating: a German roof’s clear-day peak is only about two thirds of its direct-current rating once system losses, soiling and a 50 °C cell are taken off, so the 60 % line clips the top tenth of the peak on the clearest days of the year — and a household with a store, a tank and a heat pump absorbs most of that rather than throwing it away.

What it costs needs a baseline that is capped too, because § 9 EEG does not ask whether there is an energy manager behind the meter: lifting the cap moves the managed household’s own cost by seven cents and the unmanaged one’s by twelve — and most of those seven are the simulated day’s one-minute control loop rather than the cap.

The three things that decide the figure are worth naming, because getting any of them wrong inflates it several-fold: the month (a 50 °C cell in June keeps a roof barely above the 60 % line), whether the planner is being shown the weather in advance, and whether the roof is modelled at its datasheet or at what a roof three years old actually delivers. With the cap lifted there is no curtailment at all.

An LPP session is a third source and a different story: it is a network operator asking for something right now. The three are reported apart, because telling a household that a network operator intervened when in fact the statute simply applied is a different claim about the world.

What lifts the 60 % cap is a technical fact, not a commercial one: an intelligent metering system with a control device in operation and the network operator’s first successful Ansteuerbarkeit test — § 9 Abs. 2 waits for both — or the output sold on the market with the Fernsteuerbarkeit § 10b EEG demands. “The contract is signed” and “the control path works” are different days, and on the days between them the cap still applies. The default is that nothing has lifted it: the cap staying on costs a household some feed-in, lifting it wrongly means feeding in above a statutory limit, and only one of those is the operator’s problem.

The scope of the cap is three conditions and two of them are not sizes

Reading § 9 Abs. 2 as “2 to 100 kWp, from 25.02.2025” is the obvious summary and every part of it is slightly wrong — in both directions.

There is no lower bound. § 9 Abs. 2 S. 1 Nr. 3 reaches every system below 25 kW that draws the Einspeisevergütung; the 2 kW everybody quotes is S. 4, which exempts Steckersolargeräte — and only those, and only with at most 800 VA of inverter, behind a Letztverbraucher’s Entnahmestelle. Read as a size class it exempts every small roof array from a cap the statute puts on it, so it is a declared fact about the installation rather than an inference from a nameplate.

The upper bound is exclusive: from 100 kW up, Nr. 1 demands remote reducibility and no percentage at all.

And the date is a window. § 100 Abs. 3b disapplies the cap to systems commissioned between 01.01.2023 and 24.02.2025, so “before the Solarspitzengesetz” does not mean “uncapped”: a system from 2019 still carries the obligation of the EEG version applicable to it, which it met either with equipment the operator can reduce it with — no static ceiling — or by limiting itself to 70 % (§ 100 Abs. 3 S. 2 Nr. 2). Which of the two is, again, a declaration.

§ 51 — one meter, two rules, two clocks

§ 51 Abs. 1 reduces the anzulegender Wert to zero in a quarter hour with a negative spot price. Both remuneration schemes are computed from the anzulegender Wert — § 53 Abs. 1 for the Einspeisevergütung, Anlage 1 zu § 23a Nr. 1 for the Marktprämie — so the rule zeroes the tariff and the premium. A household in Direktvermarktung priced at spot + Prämie in a negative hour is being told it still earns the premium in exactly the hours the statute is paying it to stop.

Whether the rule reaches a plant at all is a date. § 51 Abs. 2 Nr. 1 exempts anything below 100 kW for every period before the end of the calendar year in which it is fitted with an intelligent metering system — so a meter fitted in March keeps the negative hours of that whole year — and Nr. 2 exempts anything below 2 kW until a Bundesnetzagentur Festlegung that has not been made.

That date is deliberately not the same fact as the one that lifts the 60 % cap. § 9 waits for the Ansteuerbarkeit test; § 51 asks only when the meter went in. The ordinary German § 14a household of 2026 is exactly the gap between them: an intelligent metering system in for years, so the negative quarter hours earn nothing, and the 60 % cap still on because nobody has run the test.

The evidence record — and what “acted on it” means

[A1 7.2] asks the operator of a controllable device to show, for an individual case and in a way the network operator can follow, that a commanded reduction was carried out; [A1 7.3] keeps the records for two years. hems writes one append-only record per control event at the moment the guard acts: every ceiling commanded and when, the minimum the customer was owed against each, a minute-resolution trace of the netzwirksamer Leistungsbezug, and how long it took to take effect.

That last one has a trap in it. [A1 4.2 S. 5] requires the reduction to be acted on without delay, and it is tempting to record “acted on” as the first setpoint issued. But a reduction to 4,2 kW on a house drawing 2,1 kW requires nothing to be sent to anything: the manager is inside the limit from the instant it arrives, and there is no setpoint because there is no change to make. Waiting for the next thing that happens to move and calling that latency reports eight minutes on a household that was compliant from the first second — which a network operator reads as a breach.

So a reduction takes effect when the household is inside it, and the record says which of the two happened:

  slowest reaction                   0 s, commanded
  slowest reaction                   0 s, already below

Both satisfy [A1 4.2]. They are different facts, and an operator asking “what did you do?” is owed “nothing, we were at 2,1 kW” rather than a blank.

What a reduction is worth to a household

The § 14a ceiling is a constraint in the planner, so it has a shadow price: what one kilowatt-hour of relief from it would be worth to this household, computed from its own plan rather than assumed.

  relief from § 14a was worth            3.93 €/kWh

That is a house with no store whose car will otherwise leave short. The same reduction on the same household with a 10 kWh battery is worth nothing at all, because the store lends the controllable devices the headroom [A1 2.3] allows and the ceiling stops binding — and a limit that costs a household nothing is a limit nobody should be compensated for. It is [A1 2.3] measured in money, and it is what a § 41e Aggregatorvertrag offer or an OpenADR bid should be priced from — aggregators currently price both households at “30 % of nominal”.

How exposed is this connection?

[A1 8.4] makes the network operator publish how much it has been reducing — per Netzbereich, per month, as an Eingriffsdauer in hours and an Eingriffsintensität in percent. It has no timestamps in it: it is a monthly aggregate over a postcode area, so it cannot tell a household when it will be reduced, and it says nothing at all about one connection point. hems does not model it, and the absence is deliberate — nothing fetches it, and a type mirroring a document no service reads is a schema rather than a feature.

The question a household actually wants answered — how often does my operator reduce me, in which quarter hour, and to what? — can therefore only come from its own [A1 7.2] record, which the box is keeping anyway. Ninety days of it, bucketed by day type, is /v1/status:

"exposure": {
  "days_of_record": 90,
  "hours_reduced": 11.5,
  "busiest_quarter_hour": "17:45",
  "busiest_frequency": 0.8,
  "typical_ceiling_kw": 4.2
}

Your operator reduces you at teatime, four times in five, to 4,2 kW. The days of record are on the page next to the figures they are drawn from, because a share whose denominator is invisible cannot be checked — and a household that has never been reduced gets null rather than a tidy sheet of zeroes, since nought hours out of nought days is not the same claim as nought out of ninety.

Only operator reductions count. A box holding itself below its own fuse is not the grid reducing it, and counting those would report every household with a small connection as heavily controlled.

It is reported and not planned against, and that is a measurement rather than a reservation. Feeding anticipated windows into the planner was built and then removed: on a dynamic tariff a § 14a reduction lands in the evening peak, which is exactly where the price has already told the plan not to be — both signals track residual load — so anticipating it moved the reference days by between €0,000 and €0,011. On a flat tariff, where the price cannot stand in for it, it was worth −€0,018, because a charging deadline priced at €5/kWh already front-loads the session. Four measurements, none of them positive — so the wiring went, and so did the mechanism behind it. What stayed is the number a household reads.

Modul 3 — time-variable network charges

Available only together with Modul 1, only at a location without registrierende Leistungsmessung, and only with an intelligent metering system. Windows and price levels are fixed per calendar year for the whole network area, must be billed in at least two quarters, and must appear in the preliminary price sheet by 15 October of the preceding year.

There is no machine-readable national format for any of it — a PDF and an Excel sheet per network operator — so whoever commissions the box transcribes the operator’s calendar into its configuration, records where it came from, and the box refuses to start unless it conforms:

$ hemsd run --check --config /etc/hemsd/hemsd.toml
📅 Modul 3 `NB-14A-3-2026` for 2026 conforms — HT 17:00–20:00, NT 22:00–06:00, billed in Q1, Q4
   transcribed from https://www.example-netz.de/preisblatt-2026.pdf#modul3

Transcribing is not inventing, and the refusal is what keeps the difference. The seven rules are checked before a byte moves — three levels that are all reachable, a day with no gaps, at least two hours of Hochtarif on every day class, windows identical all year, at least two billed quarters, one calendar year of validity, and the three preconditions of § 1 — and a source is required, because when a household queries a bill the first question is which document said so.

The reachability rule is the one worth stating on its own, because its failure is silent. A Niedertarif band written as a single wrapping window leaves that level declared and unreachable: the register appears, the day is still fully covered by the fallback, every other rule passes, and the calendar reads as a conforming three-level tariff while the household is billed for a module whose cheap hours it can never be in.

Window membership is decided on local wall-clock time, because that is how the price sheet is written. On the long October day the repeated hour is inside the same window twice, and on the short March day the skipped hour is inside none. Both are what the document means.

MiSpeL — which kilowatt-hour was green

From 1 October 2026 a household battery that has ever been charged from the grid, or a bidirectional charge point, keeps its levy privileges and its EEG support only if the energy through it is separated: which quantity was green, which was grey, quarter hour by quarter hour, summed over a calendar month. Get it wrong and the household loses the privilege on its grid draw, the market premium on its feed-in, or both.

hems-grid::mispel implements the two options that need arithmetic, formula by numbered formula, in exact decimals.

Abgrenzungsoption (Anlage 1, formulas (1)–(33)) follows the electrons. The whole thing turns on one MIN applied per quarter hour:

(1)¼ = MIN[ Z1NB¼ ; Z2V¼ ]     grid electricity that went into the store
(2)¼ = MIN[ Z1NE¼ ; Z2E¼ ]     feed-in that came out of it

That is the Speichervorrang — the legally willed priority that attributes grid draw to the store before anything else. Summing the month first and taking the minimum afterwards is the mistake that turns a perfectly ordinary day into kilowatt-hours of arbitrage, and it is the one a test pins.

From there: the round-trip efficiency is measured where a battery is alone on the meter and presumed at 85 % wherever a charge point makes measurement impossible; energy a car brought home from somebody else’s charge point (Fremdtankstrom) is taken back out before anything is settled or supported; and storage losses are privilegeable only where the storage system’s own meter can see them.

That last clause is a Basisfall, and it is worth its own paragraph because it decides who pays for the conversion loss. Z2V¼/Z2E¼ are the store and the charge point together, which is Basisfall A3 — and a pair of meters that cannot be told apart cannot say which of them lost the energy, so A3 charges it to the household. A4 is the installation where the store has a meter of its own, reported as Z3V¼/Z3E¼, and (17)A4 then charges the round-trip loss to the store where it belongs. A box that reads its battery writes both pairs; one that does not writes Z3 as absent, which is not zero — a zero would be a settlement claiming the battery stood still — and a household that declared A4 is then refused a Nachweis rather than handed a wrong one.

Pauschaloption (Anlage 2, formulas (P1)–(P15)) is open to solar up to 30 kWp and draws two flat lines through the year instead: 500 kWh per kilowatt installed is supportable, an indifference band above it is neither, and anything above that — in quarter hours where the spot price was not negative — is settleable. The band widens as the store shrinks against the roof, because a small store leaves less room for the arbitrage the band exists to keep out.

What grid charging is actually worth, and why the planner still does not price it. Differentiate that chain at one more kilowatt-hour of (1)¼, with η the round-trip efficiency. The plant-charged quantity falls by one, the renewable storage output by η — so the grey share of what the store fed back rises by η, the levy-reducing quantity rises by η, and the levied grid draw therefore rises by only 1 − η. Grid charging costs the levies on the round-trip loss, about 0,8 ct/kWh, not the levies on the whole kilowatt-hour: the Saldierung is exactly what stops a battery being levied twice on one electron. The support side is the larger half — the anzulegender Wert is lost on η of it, about 6,7 ct/kWh — so together roughly 7,5 ct/kWh, which is the size of the whole day-ahead spread an arbitrage is chasing.

It applies only to the share of the store’s output that leaves the house, and that share is a decision the planner makes, which makes the exact term bilinear. So the optimiser prices it at zero and says so, rather than at a plausible constant: a plan that spends real money on an invented number is worse than one that admits it cannot see this.

The third option, Ausschließlichkeit, is for a store that is never charged from the grid, and it has nothing to settle. It does have something to plan, and that is where this got interesting: the optimiser expressed “no grid charging” as b_ch ≤ pv, which reads like the same sentence and is not. With a roof making 3 kW, a house drawing 1 kW and a battery taking all 3, that constraint is satisfied, the meter is running, and every kilowatt-hour in the battery is grey — the household’s whole claim spent by a plan that thought it was honouring it. What the option means is the same MIN as above set to zero, min(import, charge) = 0, which is a disjunction that no inequality implies. It is a binary now, declared only for the households that chose the option.

And the Nachweis is checked rather than declared: histd sums (1)¼ over the period, says whether the claim held, and names the quarter hours that broke it. A household should hear that from its own box rather than from its network operator.

The Festlegung is an Arbeitsstand of 05.08.2026, so every result carries the RuleSet that produced it and the day that rule set starts to apply. A Nachweis that cannot say which rules made it is not evidence of anything.

§ 42c EnWG — sharing electricity with the neighbours

Since 1 June 2026 final customers inside one distribution network’s Bilanzierungsgebiet may use renewable electricity together over the public grid. What is shared is an allocation, not physics: each quarter hour, the community’s generation is divided among its consumers by an Aufteilungsschlüssel agreed in writing, and each member’s share is billed at the community’s price instead of their supplier’s.

A static key and a dynamic one are two different contracts, not one rule and its improvement. Applying the key once and capping each member at what they used leaves generation on whoever happened to be away: a member allocated 3 kWh who consumed 1 kWh cannot take the other two, and those two are simply not shared. Re-offering them to whoever still has unmet consumption shares more — but it is a different allocation, and Abs. 3 Nr. 2 makes the Aufteilungsschlüssel a written agreement between the parties.

So hems-grid::sharing takes an Aufteilung. Statisch applies the key once and reports the remainder as unallocated — generation that went to the public grid and is settled the ordinary way. Dynamisch cascades, and is the default because it is what an energy-sharing community is for. Choosing between them silently would be choosing one of two contracts on the household’s behalf.

Each pass is metering::allocation::allocate, so Σ shared + unallocated equals the generation exactly, whichever key applies: shares are cut to a millionth of a kilowatt-hour before anything is subtracted, and the residual is a difference rather than an accumulation. A settlement that loses a millionth loses it into nobody’s account, and the loss is invisible until somebody reconciles a year.

And the plan acts on it

Settling an allocation after the fact is half of § 42c. The other half is that a household in a community should move its flexible load into the quarter hours the community is generating, which is the whole behavioural reason to join one.

What § 42c changes is which energy price applies to the kilowatt-hours the Aufteilungsschlüssel allocated, and the allocation is capped at what the member actually drew. So a slot costs

shared_price · min(share, g_in)  +  import_price · (g_in − min(share, g_in))

— a cheap block first, which is convex exactly while the community is the cheaper of the two. A convex cheap-block price needs one bounded column and one row rather than a binary: the allocation is at most the share the key offers, at most what the household drew, and it comes off the bill.

Where a community is dearer than the supplier the same function is concave, and a model that treated the discount as optional would let the plan believe it could decline an allocation it cannot — the key applies whatever anybody prefers. So the discount floors at zero and the plan claims no advantage rather than inventing one.

The price lives in the tariff, because a price is: § 42c replaces the supplier’s energy component and nothing else, since the electricity reaches the member over the public grid. On the reference winter day a community selling at 12 ct/kWh net still delivers a 32,5 ct kilowatt-hour against the supplier’s 47,9 — a third off, and not the ninety per cent that “free solar from the neighbours” suggests. Network charges, the Stromsteuer, the Konzessionsabgabe and 19 % value added tax do not care where the electron came from.

hemsd simulate --day winter --sharing runs the comparison, and the baseline is in the same community: a household joins one and then does nothing about it, and the key allocates it anyway. Leaving the baseline outside would report the value of the membership as though the planner had produced it. The day settles 12,9 kWh through the community against the 3,9 kWh the same key allocates a member who does nothing — from its own quarter-hour registers, through the same allocation a Nachweis would use, and €2,00 of credit against €0,60. It is measured on the allocation rather than on a difference of two savings: on a January day that difference is three cents, which is noise with a sign.

It runs a year on from the winter day’s own date, because § 42c Abs. 4 Nr. 1 obliges a network operator to make sharing possible only from 1 June 2026 and every reference day is dated before that. Both sides move together, by whole weeks, so the weekday and the load profile’s day type survive and the household inside a community is measured against the same Thursday as the one outside it. A household in a community on a day the rule does not reach is refused rather than settled: an allocation before the duty is one no operator would perform.

Whether a delivery point may take part at all is a metering question (Zählerstandsgangmessung or quarter-hourly registrierende Leistungsmessung, § 42c Abs. 1), and it lives in metering::sharing. hems consumes that decision rather than restating it.