P67, Gen6's precision rests on a seal friction nobody has measured

[aaaaaaaaaaaavm]

HIGH. The product claim is commanded per-satellite velocity, and at Gen6 that claim was unmodelled until 2026-08-16. A44 modelled it and failed two bands.

Gen5 backed the claim with a designed control loop: 0.0274 m/s at 3σ about a 15.8 m/s setpoint, on a gain designed against phase margin after A28 found the previous one crossing over above both track modes.

Gen6 has no loop. The shot is a 133 ms open-loop adiabatic expansion, so its spread is whatever the hardware's repeatability is.

The error budget

Term	3σ alone	Share of variance
Seal friction, ±20 % of A41's allowance	0.3115 m/s	93.4 %
Payload mass, ±0.5 %	0.0725 m/s	5.1 %
Charge pressure, ±0.25 % FS	0.0399 m/s	1.5 %

3σ dispersion is 1.113 % against a declared 0.5 %. It is worse where the setpoint is lower — 2.290 % at 20 m/s against 1.041 % at 30 — because friction is a fixed force, so its share of the commanded work grows as the command falls. That is exactly where a customer asking for a small trim would use it.

There is no instrumentation route

Transducer class	3σ
0.25 % FS	1.109 %
0.05 % FS	1.101 %

A fivefold better instrument moves the answer by 0.008 %. You cannot buy your way to this claim.

What is actually missing

A41 band 8 computed an allowance — the machine tolerates up to 83.4 N — and no run since has said what the friction is, or how much it varies shot to shot. Band 4's failure is therefore conditional, and the condition is the finding.

What would close it: a measured or supplier-bounded seal friction, with a shot-to-shot spread, for the actual piston seal at 50 bar in vacuum. That is a bench test on a component, not a calculation — the second item after #15 that changes the category of evidence rather than its degree.

Also recorded here

30.535 m/s is a zero-friction number. At A41's own full tolerable friction the same charge gives 29.009 m/s, 5.00 % lower. Both are now in cad/parameters.json with A44 named as the source. The value was not silently changed, because which of the two is the design point is a decision and not a correction.

Records: validation/A44_gen6_dispersion.md, analysis/gen6_dispersion.py.

[aaaaaaaaaaaavm]

Update 2026-08-19 — this got substantially worse, and it is now the highest-leverage action in the project.

ADR-034 took Gen6's stroke from 2.18 m to 8.0 m, the host stage's whole usable acceleration length. Friction scales linearly with contact length while shot work saturates, so the unmeasured coefficient's share of the shot rose sharply:

	Gen6 as adopted	Gen6 at ADR-034
Friction work per shot	181.8 J	667.2 J
As a fraction of shot work	9.75 %	28.39 %
Gap between the two velocity numerators	5.00 %	15.38 %

More than a quarter of the gas work now goes into a seal coefficient that has never been measured on hardware. A49 band 6 failed on exactly this and it is recorded as P78. The design point was adopted anyway, with the failure named as a falsifier rather than absorbed.

This entry now governs five open decisions, up from four:

1. ADR-034's design point itself — no longer a correction to the velocity, but the term the point rests on
2. ADR-033's trim stage — its entire justification
3. P83 — the trim authority was sized at the 9.75 % share; neither A44 nor A48 has been re-run at 28.39 %
4. P77 — the pulse store that feeds the trim stage, still unweighed, and it grows if the authority has to
5. P81 — the per-cell backup ejector, which fails by 40× at 2.18 m and 148× at 8.0 m because clearing the tube is a friction bill

It can still delete work rather than add it. If the measured friction is genuinely small: the dispersion the trim stage exists to correct is not there, ADR-033 reverses, P77 closes for free, and the light ejector in P81 becomes viable. One bench test on a piston seal at 22.73 bar in vacuum, with a shot-to-shot spread.

Nothing here changes the action. It changes how much rests on it.

---

Generated by Claude Code

[aaaaaaaaaaaavm]

This entry now chooses between two architectures — 2026-08-20, ADR-036.

A61 inverted the question. Rather than asking what a seal would give, it asked what the loosest seal is that the design can survive — and the binding requirement turned out to be thermal, not control. A 2 g seal must stay within 50 K of its own friction heating, which needs 17.8 N, 4.00 % of the piston's pressure force. The trim stage stops earning its mass at 22.3 N, 5.00 %.

Since 17.8 < 22.3, any seal that survives its own heat also makes the trim stage unnecessary. The two requirements are not in tension — one strictly implies the other.

So this measurement now has a threshold, in the unit seal data is quoted in:

Measured friction	Consequence
≤ 17.8 N (4.00 %)	P88 closes, and the trim stage is deleted. ADR-033 is amended out
17.8 – 22.3 N	Stage still unnecessary, but the seal does not survive its own heat — P88 stays open
> 22.3 N (5.00 %)	The trim stage is needed. ADR-036 falsifier 1 fires, ADR-033 stands, and P92 becomes the next run

The thing nobody had noticed, in A61's words: "The project has been sized, since A41, against the worst common seal class — and nobody ever chose it." 83.4 N is a ceiling, not a prediction. Every headline that descends from it is an upper bound presented as a design point — including the 3.9798 % dispersion now published as Gen6's, and the 144.01 mm section A55 sized to correct it.

A design sized against its own worst case is not conservative if it then buys hardware to fix the worst case. It is paying twice for a number it has never measured.

ADR-036 therefore suspends the trim stage rather than building or deleting it. Deleting on a specification would repeat what ADR-033 did in the other direction — adopting before the falsifier was answered. Work stops until this issue returns a number.

The honest limit, which ADR-036 states and does not blur: at 22.3 N the open-loop dispersion is 0.9051 % against a declared band of 0.5 %. Unnecessary and sufficient are different words. A specified seal makes Gen6 cheaper; it does not make it accurate.

---

Generated by Claude Code