[frahlg]

The house lives in this repo, and FTW's fast closed-loop simulation (go/internal/sitesim in srcfl/ftw) should judge its fuse by the same model. This moves the fuse model into an importable package, fuse/ at the module root, with no dependencies. It also replaces the model itself (Fredrik, 22 October): the simulator is the judge of Core's fuse protection, so its fuse must follow physics and IEC 60269, and must not copy Core. The model is a central estimate, one maker's mean curve within the standard's tube, not a worst case.

Why the old model could not judge

The old fuse was a damage counter: each second above the rating spent 1/t of the fuse's life, and the damage recovered linearly over a fixed cooldown.

• No preload. A fuse that had carried 0.9–1.0 × In for an hour blew at an overload no sooner than a cold one.
• Wrong long-time limit. 1.13 × blew within the hour, where IEC 60269-1 lets a gG carry 1.25 × In for it.
• Too slow at high multiples. 5 × took 20 s at every rating, against the 5 s gate.
• Recovery did not depend on the current.

The model

Each phase has three thermal bodies in a ladder:

heat I²  ->  element  ->  fuse body and sand  ->  holder and cabinet  ->  ambient

• Units. Heat into the element is u² (u = I/In). The three resistance shares Re, Rb and Rh add up to 1, so at a steady current the element sits u² above ambient.
• Melting. The element melts at UMin², UMin being the steady multiple that just melts it at 20 °C. That rise maps to tin's melting point, 232 °C: the M-effect solder spot. A warmer cabinet leaves less margin.
• Time constants. Each body's τ is its heat capacity times its own outgoing resistance. The system's modes come out of the coupled ladder: at 25 A they are 0.12 s, 9.4 s and 1,308 s.
• Exact steps. Step solves the ladder exactly for a current held over the step, from an eigendecomposition done once per band. Steps of 0.5, 1 and 2 s reach the same state to 1e-9. dt = 0 changes nothing and returns the time left at that current, which makes it a "time to melt now" probe.
• Export heats as import (I²). Cooling follows the modes; after an hour it is the slowest mode's exponential.
• API:
  • Lookup, All, TripSeconds(multiple) at the reference band, and Step(*Phase, amps, ratingA, dt) Status;
  • MeltSeconds(multiple, ratingA, ambientC), Band, Ratings, Supports;
  • Preheat(*Phase, amps, ratingA), which sets the steady state of a held current, so a scenario that starts mid-evening starts warm;
  • Phase.Theta (the state) and Phase.AmbientC *float64 (nil means 25 °C; 0 is 0 °C).
• Status.
  • ratio is current over rating.
  • trip_s is the time to melt from cold at this ratio.
  • remaining_s is the time to melt from now.
  • damage is the element's way from ambient to melting, 0 to 1.
  • active means it will melt if the current stays.
  • invalid says why a step was refused.
• Refused input. A NaN current, a negative, NaN or infinite step, an ambient that is not finite or not below 232 °C, and a rating the profile does not model are refused. The state stays as it was and invalid says why. An infinite current, or one whose square overflows, blows the fuse; the reported ratio stays finite for JSON.
• Preheat on a blown fuse is refused: a judge does not replace a fuse by itself.
• Independent of Core. Nothing is shared with Core's heat filter (τ 60 s, 1.25 × In).

Calibration

What the standard fixes. IEC 60269-1 fixes, per rating:

• the conventional currents: Inf = 1.25 In must not melt and If = 1.6 In must melt within 1 h, for 16–63 A;
• four gates at 10 s, 5 s and 0.1 s;
• the pre-arcing I²t window at 0.01 s.

These come from the standard's tables as given in:

• Schneider Electric, Complementary technical information, CA908036E v3.2 (2018);
• Enel GSCL004 rev. 0 (2016), Tables 6, 7 and 14, for 25–63 A;
• Mersen EduPack TM-104 (2012), slide 6, for 25 A.

35 A is not in those tables; its gates and I²t are interpolated on log axes between 32 A and 40 A. The reference ambient of 20 °C, with the curves approximately valid at 30 °C, is from Enel GSCL004 §6.1.

What the standard leaves open. The gates leave a wide tube. No gate constrains 1.4–2 ×, where home overloads land, and 2 × from cold is constrained by nothing. The review's sensitivity run on the a6baf29 structure shows how wide the tube is for 25 A: varying the holder's share (0.2–0.6), the holder's τ (600–2,400 s) and UMin (1.30–1.55), with every gate kept, puts a cold 1.6 × anywhere from 82 s to 2,293 s, and 2 × from 24 s to 351 s.

Fitted to a manufacturer. To pin that region, each band follows ABB's pre-arcing curves for OFAF gG fuse-links, size 000, 500 V (ABB Fusegear DIN-type HRC-fuse links, 1SCC317001C0201, page 14). The curves were digitized from the PDF's vector paths, so the read-off error is under 2 % in time (48 vertices a curve, linear on log axes between them).

All six parameters (UMin, Re, τe, Rh, τb, τh) are fitted per band, by least squares on log time from cold at 20 °C, to eleven ABB points:

• the curve's start at 7,990 s (the exact vertex, so UMin stays below it);
• 1.5, 1.6, 2, 2.5, 3, 4, 5, 8 and 10 × In;
• the current that melts in 0.01 s.

The fit keeps every gate.

R(T). An element resistance rising with temperature (silver, 0.0038 per K) was tried and made the fit worse: rms 0.11–0.14 decades instead of 0.08–0.095, UMin at its bound, and the 32 A I²t outside its window. It is left out.

Ferraz Shawmut. Their NH gG chart (gGNH, F 24) was also tried. It is a raster image with label leader lines across the curves, and it could not be traced reliably, so it is not used. Read by eye it looks slower than ABB at 2–3 ×, which shows how far makers sit apart inside the tube.

Validation against ABB (from cold, 20 °C)

Each cell is ABB / model, with the model's error in decades in brackets. The last two columns give the error at the eleven fitted points and at all 48 vertices of the ABB curve.

In	1.5 ×	1.6 ×	2 ×	3 ×	5 ×	rms, max at 11 points	rms, max at 48 vertices
16	3,547 s / 2,760 s (−0.11)	1,165 s / 1,483 s (+0.10)	53.1 s / 53.2 s (+0.00)	3.29 s / 3.48 s (+0.02)	0.333 s / 0.308 s (−0.03)	0.092, 0.14	0.199, 0.46
20	2,593 s / 1,941 s (−0.13)	820 s / 1,094 s (+0.13)	57.8 s / 57.9 s (+0.00)	4.58 s / 4.73 s (+0.01)	0.450 s / 0.425 s (−0.02)	0.095, 0.14	0.170, 0.39
25	2,341 s / 1,816 s (−0.11)	822 s / 1,058 s (+0.11)	58.7 s / 58.9 s (+0.00)	4.22 s / 4.48 s (+0.03)	0.445 s / 0.404 s (−0.04)	0.088, 0.13	0.159, 0.36
32	2,581 s / 2,060 s (−0.10)	970 s / 1,211 s (+0.10)	72.9 s / 73.2 s (+0.00)	5.06 s / 5.38 s (+0.03)	0.527 s / 0.470 s (−0.05)	0.080, 0.12	0.147, 0.33
35	2,680 s / 2,180 s (−0.09)	1,050 s / 1,284 s (+0.09)	76.5 s / 76.9 s (+0.00)	5.21 s / 5.53 s (+0.03)	0.539 s / 0.480 s (−0.05)	0.080, 0.13	0.145, 0.33
40	2,778 s / 2,246 s (−0.09)	1,067 s / 1,312 s (+0.09)	80.2 s / 80.6 s (+0.00)	5.71 s / 6.04 s (+0.02)	0.598 s / 0.544 s (−0.04)	0.081, 0.13	0.143, 0.32
50	3,566 s / 3,005 s (−0.07)	1,471 s / 1,721 s (+0.07)	122 s / 123 s (+0.00)	8.24 s / 8.65 s (+0.02)	0.829 s / 0.776 s (−0.03)	0.087, 0.15	0.139, 0.28
63	1,798 s / 1,562 s (−0.06)	910 s / 1,026 s (+0.05)	135 s / 139 s (+0.01)	9.99 s / 10.1 s (+0.01)	0.767 s / 0.710 s (−0.03)	0.092, 0.17	0.129, 0.22

The fitted points hide the shape between them. Against all 48 vertices the model is rms 0.13–0.20 decades off, at most 0.22–0.46:

• too slow at 1.6–1.95 ×: 16 A at 1.82 × melts after 447 s in the model and 150 s in ABB's curve;
• too fast at 2.0–2.5 ×: down to −0.27 decades.

The cause is a step in the model's curve. Where the element's quasi-steady rise over the body reaches the melting rise, at u = UMin/√(Re + Rb) = 2.02–2.09 × In (16–63 A), the body stage takes over from the holder. Between 1.9 and 2.3 × the model's log-log slope reaches −18 to −52; ABB's reaches −7 to −11. The review refitted all 48 vertices and reached only rms 0.12–0.15, so this is the three-body ladder's shape, not a poor fit. A test holds every vertex within 0.5 decades, as a regression check.

For sitesim and the site test: overloads at 1.9–2.1 × In sit on that step, so a small change of current there moves the melting time a lot. A finding there deserves a second look against the fitted points before it is called a defect.

Calibration against IEC 60269-1 (from cold, 20 °C)

In	UMin	Imin(10 s), ≥ 10 s	Imax(5 s), ≤ 5 s	Imin(0.1 s), ≥ 0.1 s	Imax(0.1 s), ≤ 0.1 s	I²t at 0.01 s (window)	1.25 ×	1.6 × (< 1 h)	2 ×	2 × after 1 h at In	step at
16	1.430	33 A: 22.3 s	65 A: 1.09 s	85 A: 0.197 s	150 A: 0.028 s	566 A²s (300–1,000)	never	1,483 s	53.2 s	9.0 s	2.02 ×
20	1.413	42 A: 24.7 s	85 A: 1.20 s	110 A: 0.210 s	200 A: 0.031 s	1,097 A²s (500–1,800)	never	1,094 s	57.9 s	11.0 s	2.03 ×
25	1.396	52 A: 26.3 s	110 A: 0.919 s	150 A: 0.164 s	260 A: 0.036 s	2,205 A²s (1,000–3,000)	never	1,058 s	58.9 s	10.0 s	2.03 ×
32	1.392	75 A: 14.3 s	150 A: 0.716 s	200 A: 0.174 s	350 A: 0.040 s	4,532 A²s (1,800–5,000)	never	1,211 s	73.2 s	12.2 s	2.04 ×
35 *	1.390	82 A: 14.4 s	165 A: 0.709 s	222 A: 0.168 s	387 A: 0.040 s	5,499 A²s (2,210–6,331)	never	1,284 s	76.9 s	13.0 s	2.04 ×
40	1.396	95 A: 15.1 s	190 A: 0.767 s	260 A: 0.170 s	450 A: 0.041 s	7,682 A²s (3,000–9,000)	never	1,312 s	80.6 s	14.2 s	2.04 ×
50	1.410	125 A: 17.2 s	250 A: 0.776 s	350 A: 0.163 s	610 A: 0.041 s	14,012 A²s (5,000–16,000)	never	1,721 s	123 s	22.5 s	2.05 ×
63	1.314	160 A: 19.4 s	320 A: 0.612 s	450 A: 0.125 s	820 A: 0.029 s	18,388 A²s (9,000–27,000)	never	1,026 s	139 s	19.0 s	2.09 ×

* Interpolated gates.

Preload, 1 h at In, then:

In	Damage after 1 h	1.5 × after 1 h	1.5 × cold
16	0.46	1,945 s	2,760 s
25	0.50	1,112 s	1,816 s
63	0.54	754 s	1,562 s

At 2 × the preloaded fuse takes 9–22.5 s, against 53–139 s cold.

Short overloads, 25 A. 30 s at 1.5 × every 2 min (0.5 × between) raises the damage after each burst from 0.53 to 0.64 over ten cycles. A held 1.5 × then blows in 1,532 s instead of 1,816 s cold.

Ambient, 1.5 × on 25 A. 3,053 s at 0 °C, 1,816 s at 20 °C, 1,630 s at 25 °C and 1,191 s at 40 °C.

aM, any rating. Gates from Mersen TM-104 slide 7, and all six hold. Fitted with UMin = 4 (an assumption) and the holder's share at 0.2.

Free choices and limits

• The three-body shape. Between the fitted points the model misses ABB's curve by up to 0.46 decades, most around its step at 2.02–2.09 × In (see the validation).
• The fit makes it ABB-like. UMin (1.31–1.43), the holder's share (0.50–0.60) and τ (1,250–1,950 s), the body's τ (7–19 s) and the element (Re 0.063–0.076, τ 0.07–0.20 s) are fitted to one maker's curves, inside the IEC tube. They are not fixed by physics or by the standard. Another maker's fuse-links sit elsewhere in the tube: the review's range above, and Ferraz looking slower at 2–3 ×.
• R(T) is omitted (see above).
• One melting temperature. Tin's 232 °C is used for all times; the ambient acts through it.
• gG covers 16–63 A only. IEC's 4 < In < 16 A band has other conventional currents (1.5 / 1.9 In), and ABB gives no 10 A curve here to fit. 10 A is therefore refused for gG and the profiles scaled from it: by Step (invalid) and by the simulator's API and Wails binding. Above 63 A is refused too. aM and manual model any size and take any.
• Pre-arcing only. No arcing time, current limiting or ageing.
• Not calibrated: gM, gR and gS are gG with scaled times (× 2, × 0.3, × 0.5).
• Phase current. The simulator takes it as P / 230 V.

Also here

• cmd/simulator/fuse.go keeps its names as aliases. The site keeps each phase's thermal state (Site.FuseHeat) between ticks, clears it with the fuse, its size or its profile, and takes the cabinet's air from Site.FuseAmbientC (nil is 25 °C).
• Fuse size. A size is checked together with its profile, in the API and the Wails binding. gG, gM, gR and gS take 16, 20, 25, 32, 35, 40, 50 and 63 A; aM and manual take any size. Changing the profile alone is checked against the saved size.
• Dashboard. A phase the model refuses shows the reason under its bar, for example "iec_gg is modelled for 16-63 A, not 10 A" for a site saved with 10 A before this change. The size list keeps a saved size outside the bands, so it is never blank. Reviewed in a browser: a saved 10 A gG site shows the reason on all three phases and "10 A" in the settings; picking 16 A clears the reasons and the model judges the phases again.
• Tests:
  • every gate and I²t window, ABB within 0.2 decades at the eleven fitted points and within 0.5 at all 48 vertices;
  • 1.25 × holds and 1.6 × blows through Step;
  • preload, Preheat against a 6 h soak, accumulating short overloads, cooling along the slowest mode;
  • export as import, exactness to 1e-9;
  • the status fields, 0 °C as a temperature, every refused input, infinite and overflowing currents (blown, finite ratio, valid JSON), Preheat on a blown fuse;
  • the warm cabinet, the aM gates;
  • the site loop's heat memory, sizes checked with their profiles.
• go test -race ./... passes.

What this changes for FTW

• FTW's site test (test(site): run Core against device-simulator, short set on PRs, full set nightly ftw#1797) runs on the stack minor: close the six gaps FTW's site test patched around #8 → minor: headless runs from a house spec, with a timeline and a record #10 → minor: faults and seeded random timelines for headless runs #11, which still carries the old curve. Once this merges, that stack rebases onto it. The headless run then gets site.fuse_ambient_c and a preheat for scenarios that start warm.
• sitesim's switch to this package (waiting for this) will judge preloaded cases. Expect new findings in its field matrix; they will be listed, not hidden.

Part of the work described in #8. Independent of #8; either can land first.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-22T06:52:45.513781Z	7156a59	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.