# CURVE-DEPTH — FINAL CALIBRATION: 85/85 vs 90/90

> ## ⚖⚖⚖ RULED 2026-08-25 (`#24-15` item 1) — **90/90. THE CONSTANT IS CLOSED FOR THE ERA.**
>
> **`PLATEAU_P` 85 → 90 · `CEILING` 85 → 90 · permanent reserve floor 10% of book.** Per `#24-14`
> item 3, **any further change requires a new campaign under JOURNEY-R1 discipline** — hypothesis
> frozen before data, both tails, the lot. A successor editing either constant without one is
> breaking a ruling, not tuning a value.
>
> **⛔ THE ACCEPTED PRICE, RECORDED BESIDE THE SIGNATURE RATHER THAN IN A FOOTNOTE:** −59.03% max
> drawdown on the trough path (against −57.25% at 85/85 and −50.49% under the old rule), −38.89%
> on the recovery path, and −21.33% of book in the worst single week.
>
> **▣ THE OPERATOR'S GROUND, FILED WITH IT: the PA book IS the BTC allocation itself.** Life
> reserves and other assets sit outside it — so the internal reserve is **pullback ammunition and
> discipline, not a savings cushion.** *A 10% reserve is a deployment decision here and would be a
> solvency decision somewhere else; the drawdown figures above mean something different without
> that sentence.*
>
> **⛔ NOT ONE MEASURED FIGURE IN THIS FILE CHANGED.** The tables are byte-identical to the run
> that priced the decision — and the runner is constants-independent by construction (it takes the
> plateau and the ceiling as explicit arguments and never reads the live module), so **re-running
> it after the ruling reproduces every number exactly, which was verified rather than assumed.**
> *Original stamp `e01a6a1b32d72654ddef1493e75b804be325dba27a9810d6fa324a8901258473`, kept as
> `FINAL_CALIBRATION.v1-preruling.md.ots`; the annotated file is re-stamped in
> `FINAL_CALIBRATION.sha256`.*

**Ordered `#24-14`, run 2026-08-25.** Runner: `run_final_calibration.py`. Raw: `final_calibration.json`.

**⛔ NO CONSTANT MOVES WITH THIS BLOCK.** `engine/pa_curve.py` keeps `PLATEAU_P = 85` and
`CEILING = 85` exactly as ruled. **A 90/90 configuration does not exist in the live module and this
run did not create one there.** The plateau and the ratchet are mirrored locally, and the mirror is
**proven equal to the shipped module at the ruled 85/85 — 1001 grid points and a 14-week ratchet
path, to 1e-12 — before any 90/90 number is computed.** The runner exits rather than run if they
disagree anywhere.

**Two self-checks, both green:** the control reproduces `RESULTS.md` to 1e-9, and the 85/85 rows
match `CEILING_COMPARE.md` to 1e-12. *Two instruments, same answer, or the table is not evidence.*

**Price-list rules carry (`#24-08`): both tails on every row, no coins figure without its drawdown,
no ranking.**

---

## THE TABLE

*Full signed configuration, ratchet `EXIT(0.58)·ERA` included in both rows. `held` = weeks the mark
stood above the curve. Worst single week is measured **ex-contribution** — the 1 unit/week inflow
lands before the week is priced, so a raw week-on-week change would flatter every drop by the size
of the contribution.*

### PRIMARY · 2018-12-10 → 2022-11-21 · 207 weeks · the path that ends at the bottom

| configuration | final coins | vs control | max drawdown | held wks | worst single week |
|---|---|---|---|---|---|
| CONTROL — v29, no memory | 0.01019 | — | −50.49% *(2022-11-07)* | — | **−18.76%** *(2021-05-17)* |
| 85/85 *(superseded)* | 0.01207 | +18.36% | −57.25% *(−6.76pp)* | 68 | −20.80% |
| **⚖ 90/90 — RULED, LIVE** | **0.01305** | **+28.06%** | **−59.03%** *(−8.54pp)* | **68** | **−21.33%** |

**→ 90/90's price over 85/85: +8.19% coins for 1.78pp of additional maximum drawdown.**

### peak-ending · 2018-12-10 → 2021-11-08 · 153 weeks · the path that ends at the top

| configuration | final coins | vs control | max drawdown | held wks | worst single week |
|---|---|---|---|---|---|
| CONTROL — v29, no memory | 0.00554 | — | −34.10% *(2021-07-12)* | — | **−18.76%** *(2021-05-17)* |
| 85/85 *(superseded)* | 0.00738 | +33.15% | −37.92% *(−3.82pp)* | 58 | −20.80% |
| **⚖ 90/90 — RULED, LIVE** | **0.00797** | **+43.89%** | **−38.89%** *(−4.79pp)* | **58** | **−21.33%** |

**→ 90/90's price over 85/85: +8.07% coins for 0.98pp of additional maximum drawdown.**

---

## §2 — THE WORST SINGLE WEEK, IN PORTFOLIO TERMS

**Translated to the live 100-unit book, which is the only scale the operator holds:**

| | worst week | on a 100-unit book | vs 85/85 |
|---|---|---|---|
| CONTROL | −18.76% | **18.76 units gone in one week** | — |
| 85/85 | −20.80% | **20.80 units** | — |
| 90/90 | −21.33% | **21.33 units** | **+0.53 units** |

**⚑ AND THE HONEST READING OF THAT COLUMN IS THAT IT IS THE SMALL DIFFERENCE, NOT THE LARGE ONE.**
Going from 85/85 to 90/90 costs **about half a unit of book in the worst week the record contains**
— roughly the size of one ordinary weekly deployment step. **The cost that is not small is the
SUSTAINED one: 1.78pp of additional maximum drawdown on the trough path, on top of an already
−57.25% hole.** *If the worst week is what the decision feels like, 90/90 is nearly free; if the
long grind is what it feels like, it is not. The table gives both because they point different
ways.*

**⚠ TWO FACTS ABOUT THAT WEEK THAT CHANGE HOW IT SHOULD BE READ.**

**1. The worst single week is NOT at the bottom.** It is **2021-05-17** — the May 2021 crash, in
the middle of the recovery leg — while the deepest drawdown lands at **2022-11-07** on PRIMARY and
**2021-07-12** on peak-ending. **The worst week and the deepest hole are different events**, and a
configuration chosen on one is not thereby chosen on the other.

**2. It is the same week and the same figure on both windows** — peak-ending is a prefix of
PRIMARY, so they share it. **The worst-week column therefore discriminates between configurations
but NOT between paths**, which is exactly the opposite of the max-drawdown column, where 90/90 costs
1.78pp on the trough path and 0.98pp on the recovery path. *Read them as a pair or the asymmetry
disappears.*

**3. The percentages are on a growing book, so the unit figures in the raw JSON are not comparable
across rows.** 90/90 loses more raw units in that week (93.69 vs 86.69) largely because it was
*holding more* — the book it fell from was 439 units against 417. **The comparable number is the
percentage, and the 100-unit translation above is built from the percentage, not the raw units.**

---

## §3 — WHAT THE TWO ROWS SHARE, WHICH IS WORTH SEEING BEFORE RULING

**`held wks` is IDENTICAL in both configurations — 68 on PRIMARY, 58 on peak-ending.** The plateau's
height does not change how long the memory holds a released target, only how high the held target
is. **So the "weeks of holding a target the composite has released" commitment the operator weighed
at `#24-09` is unchanged by this decision.** *This ruling is about height alone.*

**And the direction is the campaign's own monotonicity, arriving for the fourth time:** deeper buys
more coins and carries more drawdown, on every window, with no crossing. **This table cannot tell
the operator which is better and does not claim to** — it prices a risk preference, exactly as
`FINDINGS.md` ruled the whole campaign does.

---

## WHAT THIS DOES NOT DO

**It moved no constant and ranked nothing** — it priced a pair, and the operator ruled against the
pair rather than against a score. `#24-15` took the 90/90 row the next block along.
**`#24-14` item 3 made this the last calibration of the era, and the ruling has now closed the
constant: further changes come only through a new campaign under JOURNEY-R1 discipline.**

**⚠ AND THE SCOPE LIMIT INHERITED FROM `CEILING_COMPARE.md` SURVIVES THE RULING:** the conditional
drift guard is not simulated by either runner, so the coins columns are an upper bound — equally
biased across rows, so the 85-vs-90 comparison is undistorted, but the absolute figures are not
what a live path would have produced. **`#24-15` item 3 ruled the drift guard unchanged and
deferred, which means that bias is now a standing property of every figure in this file.**