# USD SOFR OIS Discount Curve — Methodology Document

**Document version:** demo-v1  
**Document date:** 2026-10-12  
**As-of date:** 2026-09-18; **spot date:** 2026-09-22  
**Status:** for documentation readiness assessment  
**Fictional inventory:** HER-QM-017 (simulated)  
**Source revision:** `01ae3bdac6f61ecbabcd72e0116ebb2860e93157`

> This methodology demonstration combines real historical public data, derived numerical evidence and an explicitly fictional institution dossier. It is not a live pricing service, an approved production model or a complete regulatory submission. All institutional roles, policies, records and controls are simulated; actual institutional evidence and authorization remain outside this demonstration.

The model converts a historical SOFR fixing and estimated swap rates into the present value of future USD payments and the compounded SOFR rates consistent with those values. The developer chooses method B, monotone-convex forward interpolation, principally because its projected forwards are continuous and show smaller observed changes than method A, log-linear discount factors. The tradeoff is wider sensitivity spillover and substantial long-end uncertainty: the largest observed disagreement is about USD 227,458 for one USD 100 million payment on 2050-04-18, not a portfolio gain or loss. [BRIEF, Model and choice 1; RESULTS, `df_difference_A_minus_B_monthly_grid`, `forward_smoothness`, `locality_bump_5Y_plus_1bp`.]

The curve is a single-currency, single-curve analytical construction calibrated to two deposits and twenty overnight-indexed swaps (OIS). Inputs are a published historical fixing and derived close-adjusted medians of reported executions; they are not a licensed closing curve. Harbor Example Research is fictional and proposes only sandbox explanation, historical comparison and illustrative sensitivity work. Read Section 5 for the interpolation decision, Section 7 for its measured support, and Sections 8–10 for reliance limits, assumptions and unresolved authorization. [BRIEF, Model; QUOTES, tenor table; INST, Inventory and intended use.]

## 1. Purpose and scope

Discounting translates a promised payment at a future date into a present value. A discount factor of 0.80 means that one dollar paid on that date has a model value of eighty cents at the curve's as-of date. Projection uses the same curve to infer the compounded overnight SOFR accrual between two dates; it is a rate implied by this construction rather than a forecast that future published fixings will equal it. The Secured Overnight Financing Rate (SOFR) is the underlying overnight reference rate identified in the retained New York Fed fixing. This document concerns the curve's implementation, its historical evidence and the fictional research scenario, rather than a transaction recommendation. [BRIEF, Model and Par relation; FIXING, `last_known_fixing_date`, `sofr_percent`; HW-CODE, `Ois`, `ois_par_rate`.]

The reader's decision is whether the developer's choice is understandable and sufficiently documented for a separate assessment. Both methods are intended to fit the same instruments; exact calibration alone therefore cannot choose between them. The relevant distinction is how each fills the spaces between quoted maturities and how that choice changes projected forwards and sensitivities. B is the developer's preferred interpolation for projection. A provides a transparent comparison and a more local risk pattern on the retained tests. No model validity certification follows from either statement. [BRIEF, choice 1 and smaller choice “Method A as the comparison method”; RESULTS, comparison metrics.]

The fictional commission records an explanation of historical inputs as of 2026-09-18. The proposed sandbox would support hypothetical SOFR-discounted USD cash flows and single-curve compounded-SOFR OIS; the dossier authorizes no production use or trading. There is no real portfolio, actual customer exposure or live downstream integration. The USD 100 million notional used in numerical examples is a unit exposure, not a position inventory. The numerical run occurred historically in October; its receipts are evidence about that calculation, not a record of current deployment. [INST, Inventory and intended use; ENV, stage timestamps.]

The method supplies dated discount factors through the last 30Y instrument and an explicit extrapolation beyond it. Prices can be evaluated at intermediate dates, but the information supplied by the market remains concentrated at eligible OIS tenors. A visually smooth interpolated curve does not add observations between those tenors. This is especially consequential between 20Y and 30Y, where the retained 20Y quote has six prints and only USD 0.04 billion reported notional after rounding. [QUOTES, 20Y row; RESULTS, forward comparison; BRIEF, choice 1.]

## 2. Product and portfolio coverage

The covered analytical products are USD cash flows assumed to be discounted at SOFR and OIS whose floating leg compounds SOFR daily in arrears. A single curve both discounts and projects that rate. This is an explicit modeling scope: other collateral rates or currencies would require different discounting, and other projected rates cannot be inferred merely by relabeling the SOFR curve. [BRIEF, Model, Scope and exclusions.]

Excluded uses include a separate projection/discount multi-curve framework, tenor basis, cross-currency basis, collateral-choice or credit support annex (CSA) specific discounting, term SOFR, SOFR futures and their convexity adjustment, and explicit central-bank-meeting or year-end jumps. The implementation does not model option volatility, dynamic scenarios or transaction costs. The first seven exclusions are supplied developer scope; the latter absence follows from the static cash-flow equations and retained input set, rather than evidence that such risks are immaterial. [BRIEF, Scope and exclusions, choices 5–6; BOOT, `ql_helpers`; HW-CODE, `ois_par_rate`.]

Calibration swaps run from 1M to 30Y after spot, with missing 9M and 25Y tenors. The overnight and tom-next deposits bridge the as-of date to spot. The monthly comparison grid runs from 2026-09-18 to 2056-09-18, whereas the final instrument payment node is 2056-09-26. These are different horizons. Beyond that node the curve is extrapolated; the 40Y illustration is not calibrated to a 40Y quote. [GRID, first/last rows; NODES, O/N, T/N and 30Y rows; RESULTS, `extrapolation`.]

[INSTITUTIONAL GAP: Actual usage and exposure inventory, customer positions and deployed downstream consumers are unavailable.]

[INSTITUTIONAL GAP: The actual production instrument set, collateral terms and approved convention set are unavailable.]

## 3. Model description and theory

### 3.1 Cash-flow relations and conventions

Let (Z(t)) be the dimensionless present value at 2026-09-18 of USD 1 paid on date (t), with (Z(0)=1). Curve time (T(t)) is Actual/365 Fixed from that date. The continuously compounded zero rate and instantaneous forward are

\[
r(t)=-\frac{\log Z(t)}{T(t)},\qquad
Z(t)=\exp\left[-\int_0^{T(t)} f(u)\,du\right].
\]

Zero rates at time zero are undefined in the monthly CSV, which records `nan`; this is not a missing nonzero-date discount factor. Percentage zero rates multiply the rate by 100. One basis point (bp) is 0.0001 in rate units. Changing the descriptive time basis changes quoted zero rates, not the discount factors being described. [BOOT, `T`, `CurveSet.zero`, monthly-grid writer; HW-CODE, `integral`, `discount`; BRIEF, smaller construction choices.]

A simple Act/360 deposit of rate (q), start (s), end (e), and accrual fraction (\alpha) imposes (Z(e)=Z(s)/(1+q\alpha)). Both retained deposits use 3.85%: O/N from 2026-09-18 to 2026-09-21 and T/N from 2026-09-21 to 2026-09-22. These dates span four calendar days and two SOFR business days. The input is the fixing effective 2026-09-17, published 2026-09-18, so the later fixing bearing the as-of date is not used with hindsight. [BRIEF, Instruments and conventions, choice 4; FIXING; HW-CODE, `bootstrap` deposit construction.]

For OIS period (k), let (s_k,e_k,p_k) denote accrual start, accrual end and payment date, and let (\alpha_k) be its Act/360 accrual fraction. The fixed rate equates discounted fixed and floating payments:

\[
R=\frac{\sum_k \left[Z(s_k)/Z(e_k)-1\right]Z(p_k)}
{\sum_k \alpha_k Z(p_k)}.
\]

The projected daily factors telescope to (Z(s_k)/Z(e_k)) under this single-curve construction. Payment discounting uses (p_k), not (e_k). Thus the floating leg does not generally reduce to a single difference of discount factors when there is a payment lag. For one accrual period the common pay-date factor cancels from the par ratio, which helps explain why a quote may not pin the very last payment-date discount factor directly. This relation is implemented in `ois_par_rate`; the brief describes it as a generalization of HW2006 equation (1), an attribution to the developer rather than a newly verified reading of the paper. [HW-CODE, `Ois`, `ois_par_rate`; BRIEF, Par relation.]

Spot is two business days after trade date. Fixed and floating legs use annual schedules, with one payment for maturities up to one year, Act/360 accrual, and payment two business days after accrual end using Following. The brief identifies backward accrual-date generation in twelve-month steps with Modified Following. Both methods obtain dates from QuantLib's SOFR calendar and the same helper objects. Payment dates are the nodes: the 2Y accrual maturity is 2028-09-22 but the node is 2028-09-26. The gap between accrual end and payment is interpolated. Thus equal instrument repricing does not require equal node zero rates: at 1M, A is 3.9407% and B is 3.9477%. [BRIEF, Instruments and conventions; BOOT, `ql_helpers`, `hw_instruments`; NODES, 1M and 2Y rows.]

### 3.2 Interpolation and solution

Write (t_i=T(\text{node}_i)), (h_i=t_i-t_{i-1}), and (d_i=-\log[Z(t_i)/Z(t_{i-1})]/h_i). Method A linearly interpolates log discount factors: inside interval (i), (\log Z(t)=\log Z(t_{i-1})-d_i(t-t_{i-1})). Its instantaneous forward is constant on each interval. A's QuantLib implementation is `PiecewiseLogLinearDiscount`; the separately written `LogLinearDiscount` class represents the same construction. The rates can jump at a node because adjacent (d_i) need not coincide. [BOOT, `CurveSet`; HW-CODE, `LogLinearDiscount`.]

Method B constructs node forwards from adjacent discrete forwards. At an interior node,

\[
f_i=\frac{h_i d_{i+1}+h_{i+1}d_i}{h_i+h_{i+1}},\qquad
f_0=d_1-\tfrac12(f_1-d_1),\qquad
f_n=d_n-\tfrac12(f_{n-1}-d_n).
\]

The optional positivity collar clips the first and final forwards to the interval from zero to twice the adjacent discrete forward, and each interior forward to zero through (2\min(d_i,d_{i+1})). On each interval, (x=(t-t_{i-1})/h_i), (g_0=f_{i-1}-d_i), (g_1=f_i-d_i), and (f(t)=d_i+g(x)). Sector (i) uses

\[
g(x)=g_0(1-4x+3x^2)+g_1(-2x+3x^2).
\]

The other sectors replace inappropriate shapes with flat pieces joined to quadratics, or two quadratics meeting at an interior value. The `_sector`, `_g` and `_G` functions specify the complete branches and their analytic integrals. Integration gives (Z(t)=\exp[-\int_0^t f]) and preserves each interval's discrete forward; merely interpolating node forwards without that integral constraint would be a different method. The chosen implementation is the developer's unameliorated Hagan–West account, with equations identified in the code comments as HW2008 (22)–(34). No general theorem about every admissible input is established by the retained snapshot. [HW-CODE, `MonotoneConvex`, `_sector`, `_g`, `_G`; BRIEF, choice 1.]

Deposit intervals are fixed analytically. The remaining discrete forwards are solved simultaneously so all twenty OIS par residuals vanish. SciPy `root(method="hybr")` uses each swap rate as an initial forward guess, tolerance (10^{-14}), and a maximum 20,000 function evaluations; the code raises an error if the largest remaining par-rate residual exceeds (10^{-11}) in rate units. These are solver settings, distinct from the fictional house's 0.01 bp instrument criterion. Neighboring intervals influence B's interpolation, which explains the developer's simultaneous solve and its broader sensitivity. There is no general convergence proof in this evidence. [HW-CODE, `bootstrap`; HOUSE, Modeling choices and tests.]

Beyond the last node, each method holds its own final instantaneous forward constant. Thus (Z(t)=Z(t_n)\exp[-f(t_n)(t-t_n)]). A holds the last interval forward; B holds its end-rule value. This is an explicit continuation assumption, not information from a traded long-maturity instrument. [HW-CODE, both `integral` methods; BRIEF, choice 7.]

## 4. Inputs and data

### 4.1 Lineage, timing and transformations

**Table 1. Input classes for the 2026-09-18 snapshot. Units and date roles distinguish source observations, transformations and scenario facts.**

| Input | Meaning, timing and unit | Transformation and quality limit | Source |
|---|---|---|---|
| Retained NY Fed fixing | 3.85% effective 2026-09-17, published 2026-09-18; retained volume USD 2,992 billion | Assumed constant to spot; subsequent 2026-09-18 and 2026-09-21 fixings are ex-post evidence only | FIXING, named fields |
| DTCC cumulative rates report | Report date 2026-09-18; execution times converted to New York | Public trade prints, not executable offers or a closing fixing; raw archive omitted | QUOTES, filter account; PROVENANCE |
| Close-adjusted OIS estimates | Median rate (%) at hour 16 New York; IQR in bp | One common intraday path and tenor loading; statistical estimate from filtered executions | QUOTES, tenor/path/loading tables; DERIVE |
| Instrument dates | Spot 2026-09-22; accrual/pay/node dates | Shared QuantLib SOFR calendar, Act/360 accrual and Act/365F curve time; independent calendar reconciliation absent | BOOT, helpers/instruments; NODES |
| Fictional house and dossier | HER-scenario-1, authored 2026-10-12 | Simulated roles, thresholds and records; no operating evidence | HOUSE; INST |

Source: `sofr_fixing.json`, `quotes.md`, `derive_quotes.py`, `bootstrap.py`, `institution.md`. Values in this table are supplied, not new market observations.

The developer selects NEWT/TRAD USD SOFR OIS executions on the as-of New York date, effective at spot, without package flags or upfront payments, with an A004 fixed-leg code and annual or expiry-only schedules on both legs. Maturities within four calendar days of a standard tenor are matched. The report has 23,563 rows and 4,258 new USD SOFR OIS trades; 663 spot prints pass the filters, 529 match standard tenors, and the twenty eligible tenor sets supply 523 calibration-tenor prints. The six-print difference is the four 9M and two 25Y observations excluded by the count rule. A tenor requires at least five prints and adjusted IQR no greater than 2 bp; dispersion excludes none on this date. [QUOTES, filter paragraphs and tenor rows; DERIVE, `MIN_TRADES`, `MAX_IQR_BP`, tenor eligibility condition.]

Across all effective dates the filter account excludes 76 monthly-paying and 309 otherwise qualifying prints carrying another payment. In the separate population of 4,051 as-of executions including packages, 1,661 have another payment and 2,390 do not; 26% versus 3% have fixed rates on round 5 bp steps. This supports the developer's concern that standard coupons with upfronts need not be par rates. These denominators differ from 663 spot prints and should not be combined. Of 4,258 new trades, 4,211 use A004; 47 with other or blank codes are excluded rather than converted. The interpretation of A004 as Actual360 is the brief's account of the historical CFTC specification. [QUOTES, filter and code-count paragraphs; BRIEF, choice 3 and smaller selection choices.]

Twenty of the 529 matched prints remain more than 5 bp from their adjusted tenor medians, with deviations 5.7–154.3 bp. They are retained to avoid tuning an exclusion rule; the developer reports that removing them moves no used quote more than 0.124 bp. Median robustness does not prove every retained print is a valid par execution, particularly at a thin tenor. A count floor of ten would remove 1M, 2M, 18M, 8Y, 9Y, 12Y, 15Y and 20Y, illustrating the tradeoff between coverage and depth. [QUOTES, off-market paragraph; BRIEF, choice 3.]

### 4.2 Derived close and tenor evidence

For execution (j) at hour (h) and tenor (m), the estimate is (r_j^{close}=r_j-\lambda_m P_h), where (P_{16}=0). Median regression (`QuantReg`, quantile 0.5) fits a common path on tenors at least 2Y with loading one. The retained path is −6.8 bp at hour 00, −2.8 bp at hour 08 and zero at hour 16; its minimum displayed value is −8.2 bp at hour 02, so the movement is not a monotonic rise. The developer's “about 7 bp” describes the start-to-close level difference. The close is the last traded hour, 16:00–16:59, not an instantaneous published fixing. [DERIVE, path fit and `close_adj`; QUOTES, common intraday path.]

Short-tenor loadings require at least eight prints in five distinct hours; otherwise they are interpolated. Fitted loadings include 3M 0.08, 4M 0.15, 6M 0.14, 1Y 0.49 and 18M 0.86. The 1M fitted −0.04 is clipped to zero; 2M is interpolated at 0.04. All loadings are constrained to [0,1]; fitted long-tenor values range 0.97–1.08 but are used as one. Forward-starting trades take the loading of time from as-of to expiration, treating a one-year swap starting in one year like a two-year maturity for this adjustment. This rule avoids a forward-start model but ignores start/length differences. [QUOTES, loading table; DERIVE, `loading_for`; BRIEF, choice 2 and smaller close-adjustment choices.]

For maturities of at least 1Y, adjusted within-tenor IQRs are 0.1–0.8 bp versus 1.7–5.8 bp raw. The raw 4Y median is 5.9 bp below 3Y after rounding, a pattern the developer attributes to execution timing. Narrowing is partly in-sample because the path fits the same prints. Leave-one-tenor-out refits move a path tenor's quote by at most 0.165 bp; that tests own-tenor path influence, not short-tenor loading estimation or an independent closing level. Alternatives are unadjusted all-day medians, late-hours-only medians, least-squares fits, or a licensed closing curve; none supplies a demonstrated superior estimate here. [QUOTES, tenor and leave-one-out tables; BRIEF, choice 2.]

**Table 2. Derived OIS quote evidence, as-of 2026-09-18, spot 2026-09-22. Trades and within-1-bp values are counts; rates are percent, dispersion is bp, notional is USD billion. These are historical estimates.**

| Tenor | Trades | Quote: close-adjusted median (%) | IQR adjusted (bp) | Within 1 bp of quote | Raw median (%) | IQR raw (bp) | Notional (USD bn) | Used |
|---|---:|---:|---:|---:|---:|---:|---:|:---:|
| 1M | 8 | 3.8974 | 0.1 | 8 | 3.8974 | 0.1 | 9.67 | yes |
| 2M | 5 | 3.9683 | 0.1 | 5 | 3.9678 | 0.1 | 3.39 | yes |
| 3M | 17 | 4.0319 | 0.3 | 17 | 4.0295 | 0.3 | 8.81 | yes |
| 4M | 10 | 4.1031 | 0.2 | 10 | 4.0991 | 0.4 | 4.32 | yes |
| 6M | 10 | 4.2057 | 0.3 | 10 | 4.2007 | 0.7 | 0.92 | yes |
| 9M | 4 | 4.3699 | 0.5 | 4 | 4.3492 | 1.1 | 0.18 | no |
| 1Y | 41 | 4.4987 | 0.7 | 35 | 4.4778 | 3.3 | 5.31 | yes |
| 18M | 9 | 4.5868 | 0.3 | 8 | 4.5691 | 4.5 | 0.23 | yes |
| 2Y | 51 | 4.6358 | 0.3 | 44 | 4.6060 | 5.6 | 4.00 | yes |
| 3Y | 25 | 4.6342 | 0.5 | 24 | 4.6206 | 4.5 | 1.43 | yes |
| 4Y | 33 | 4.6107 | 0.6 | 31 | 4.5612 | 5.2 | 0.62 | yes |
| 5Y | 76 | 4.5968 | 0.6 | 69 | 4.5803 | 4.2 | 4.63 | yes |
| 6Y | 12 | 4.5917 | 0.5 | 12 | 4.5778 | 4.3 | 0.39 | yes |
| 7Y | 42 | 4.5923 | 0.7 | 38 | 4.5732 | 5.1 | 1.90 | yes |
| 8Y | 8 | 4.5964 | 0.5 | 8 | 4.5411 | 5.8 | 0.11 | yes |
| 9Y | 8 | 4.6107 | 0.3 | 8 | 4.5897 | 2.4 | 0.13 | yes |
| 10Y | 99 | 4.6208 | 0.8 | 88 | 4.6067 | 2.2 | 2.76 | yes |
| 12Y | 5 | 4.6582 | 0.2 | 4 | 4.6455 | 1.7 | 0.09 | yes |
| 15Y | 7 | 4.7085 | 0.5 | 6 | 4.6982 | 2.1 | 0.08 | yes |
| 20Y | 6 | 4.7477 | 0.1 | 4 | 4.7278 | 3.7 | 0.04 | yes |
| 25Y | 2 | 4.7189 | 0.0 | 2 | 4.7146 | 0.2 | 0.01 | no |
| 30Y | 51 | 4.6637 | 0.8 | 47 | 4.6495 | 2.2 | 0.60 | yes |

Source: QUOTES, `quotes.md` tenor table and `quotes.csv`, originally calculated by `derive_quotes.py`. Retained displayed precision is preserved; “Used” applies to the bootstrap eligibility rule.

## 5. Method choices and rationale

### 5.1 Principal choice: measured disagreement and tradeoff

Define the largest observed present-value disagreement as

\[
\max_{t\in G} 100{,}000{,}000\,|Z_A(t)-Z_B(t)|,
\]

where (G) is the 361-date SOFR-calendar-adjusted monthly grid from 2026-09-18 through 2056-09-18 inclusive. The retained maximum is USD 227,458.20 on 2050-04-18, from a dimensionless difference 0.0022745819731463945. Displayed discount factors are 0.33209463 for A and 0.32982005 for B; A minus B is positive at that date. This is the disagreement for one specified USD 100 million cash flow, not profit, an observed loss, portfolio risk, or proof that either curve is correct. The exact currency conversion uses retained JSON precision; the rounded CSV values alone should not support cents. [RESULTS, `df_difference_A_minus_B_monthly_grid`; GRID, 2050-04-18 row; BOOT, monthly grid and `NOTIONAL`.]

**Table 3. A/B method comparison, as-of 2026-09-18. Historical measures use shared quotes, fixing, instrument dates and conventions; no general superiority threshold is supplied for shape or locality.**

| Criterion and population | A: log-linear DF | B: monotone convex | Measure and locator |
|---|---:|---:|---|
| Largest instrument repricing error (bp), all 22 inputs | 3.8601e-9 | 2.7554e-10 | Absolute model-minus-input rate; RESULTS `repricing_max_abs_error_bp`; criterion ≤0.01 bp each |
| Largest successive-business-day forward move (bp), as-of through last node | 59.71 | 1.68 | Maximum absolute change of Act/360 overnight forward; `forward_smoothness` |
| Moves above 2 bp (count), same grid | 18 | 0 | Diagnostic threshold, not acceptance rule; same fields |
| Largest five-business-day forward change (bp), same grid | 59.73 | 5.34 | Absolute endpoint difference, not cumulative variation; same fields |
| Total forward variation (bp), same grid | 421.3 | 365.0 | Sum of absolute successive changes; `forward_total_variation_bp` |
| 5Y +1 bp: largest absolute forward change outside [4Y,6Y] (bp) | 0.0305 | 4.1475 | Dates strictly outside 2030-09-25 through 2032-09-24; `locality_bump_5Y_plus_1bp` |
| 4.5Y payment: share of absolute PV sensitivity outside 4Y/5Y (%) | 5.86 | 11.94 | Sum of absolute individual quote-bump PV changes outside / sum across twenty OIS inputs; `pv_ladder_4p5y` |
| Zero rate on maximum-disagreement date (%) | 4.6715 | 4.7006 | Continuous Act/365F; GRID 2050-04-18 |
| Extrapolated discount factor at 40Y (dimensionless) | 0.16426043 | 0.16853911 | No 40Y quote; `extrapolation` |

Source: `results.json`, `df_grid_monthly.csv`, calculated historically by `bootstrap.py` and `hagan_west.py`. Rounded displays retain their source precision; thresholds are from HOUSE where stated.

The developer's reason for B is its projection shape, with a smaller benefit for some short broken dates. B's continuity does not imply uniform smoothness: sectors (ii) and (iii) insert flat stretches, and after the 3Y node its overnight forward falls from 4.4705% to 4.4317% in three business days. The no-jumps-above-2-bp statistic alone would miss that short ramp. B uses sectors i/ii/iii/iv on 10/4/3/5 intervals and the positivity collar binds at zero nodes on this date. [BRIEF, choice 1; RESULTS, `forward_after_3Y_node`, `hagan_west_sectors`, `hagan_west_collar_binding_nodes`.]

A's cleaner local risk can matter for instrument-by-instrument hedging. B's 5Y bump modifies the 4Y–5Y discrete forward, which changes the node forward at 4Y and bends the preceding 3Y–4Y interval; the largest response outside the chosen window is just after the 3Y node. This causal explanation follows the neighboring-forward formula, not an inference that the 3Y input changed. The test uses a mid-curve 5Y tenor with 76 prints; 10Y, with 99, is the most-traded path tenor. It is one quote bump and one cash-flow ladder, not a stability norm over every cash flow and input. [BRIEF, smaller test choices; QUOTES, 5Y/10Y rows; HW-CODE, node formula; BUMP, rows near 2029-09-26.]

The long end exposes the interpolation assumption most strongly. The low 20Y–30Y discrete forward lowers B's 20Y node forward and influences the previous interval from about 18Y; the end condition then places the final forward below the interval's average. B's Act/360 overnight forward moves from 4.6127% at 20Y to 3.9957% at 30Y; its business-day average over the interval is 4.2013%. The end is about 20.6 bp below that average and about 67 bp below the 30Y par rate, 4.6637%. A holds approximately 4.2248%. The average refers to starts on business days from 2046-09-26 to the day before 2056-09-26; it is not the Act/365F discrete forward or a par rate. The developer describes the final-node condition as imposing zero instantaneous-forward slope at 30Y, extending the preceding curve shape. Sparse inputs and that end rule determine this shape; price evidence does not identify it uniquely. [RESULTS, comparison table; QUOTES, 30Y; BRIEF, choice 1; BOOT, long-end average calculation.]

![Historical discount-factor comparison](figures/fig_discount_factors.png)

*Figure 1. As-of 2026-09-18 discount factors (dimensionless) against years from as-of (Act/365F), comparing A and B on the historical monthly grid, with the time-zero point omitted by the plotting code. Frozen figure `fig_discount_factors.png`, from `bootstrap.py` using `df_grid_monthly.csv`, `quotes.csv` and `sofr_fixing.json`. The curves overlap visually over much of the domain; this scale does not establish numerical equality or eliminate the single-payment disagreement above.*

![Historical overnight-forward comparison](figures/fig_forwards.png)

*Figure 2. As-of 2026-09-18 Act/360 simple overnight forwards (%) against years from as-of (Act/365F), on SOFR business-day starts through one year beyond the last node. Frozen `fig_forwards.png`, `bootstrap.py`, `forward_daily.csv`. “Overnight” accrues to the next business day and can span a weekend. The plot displays A's staircase and B's continuous construction and flat extrapolation; it does not show realized future rates or establish a preferred shape in an untested regime.*

### 5.2 Alternatives and remaining construction choices

QuantLib's convex-monotone default blend, quadraticity 0.3 and monotonicity 0.7, converges but is not the selected paper setting. Its maximum monthly-grid absolute DF difference from B is 8.01175e-5 and largest daily forward move 20.56 bp. The paper-setting variant (0,1) fails in QuantLib's iterative bootstrap at the 3M instrument, with “root not bracketed”; that failed solver attempt is not a successful alternative curve and does not establish that B's simultaneous system is impossible. The uncollared B variant converges and is identical in the retained result because the collar does not bind. These variants have only the reported subset of tests; cubic splines were rejected by the developer for positivity/locality concerns attributed to HW2008 Table 1 and were not built. [RESULTS, `interpolation_variants`; BRIEF, choice 1.]

The developer keeps A as a transparent mistake-finding reference, attributing that practice to HW2008; A is not demonstrated to be the best competitor. Using shared dates isolates interpolation comparisons conditional on those dates, at the cost of shared schedule defects. Placing nodes at payment dates follows the selected helper pillars; using final accrual ends instead was considered, but changes the few interpolated days beyond each instrument's accrual. Act/365F time and continuous compounding are descriptive choices; Act/360 time was an alternative. The positivity collar is intended to preserve positive forwards with positive discrete inputs, while flattening can occur where it binds; a no-collar alternative was actually measured. Simultaneous solving replaces possible node-by-node fixed-point iteration because of neighboring-interval dependence, without a general convergence guarantee. [BRIEF, Smaller choices—Construction; HW-CODE, collar, solution and interpolation functions.]

The other principal choices are deliberately narrow. OIS-only calibration avoids absent futures settlement data and the developer's cited need for a convexity adjustment, at the cost of front-end detail. Single-curve construction avoids a broader multi-curve framework and projects only SOFR. The last known fixing bridges to spot without pre-swap interpolation; starting at spot or using a later published fixing were alternatives, but the latter would introduce unavailable-at-close information. Flat final-forward extrapolation follows the developer's attribution to HW2006 p.109; a long-rate assumption or institution-specific policy would be different alternatives and is not supplied. [BRIEF, choices 4–7.]

Selection choices also affect the curve: a four-calendar-day maturity allowance absorbs weekends/holidays but may admit an unusual stub; exact-date matching or recomputed business-adjusted dates would be alternatives. New York dating aligns executions with the chosen close instead of the report's UTC day, but can include evening executions. Keeping packages, converting other schedules, converting non-A004 day counts, raising the print count floor, removing off-market prints or dropping the dispersion test are alternative designs, not results of this model. Each has a supplied reason for rejection or deferral; none is silently assumed to have been tested. [BRIEF, choice 3 and Smaller choices—Print selection.]

### 5.3 Estimation and test-design decisions

The developer describes public dissemination as the available free, date-specific citable source for this example; the stronger claim that it is the only such source is the developer's rationale, not an independently established market-data survey. Packages are excluded because a leg's rate can reflect a package-wide bargain. Five sufficiently tight prints are considered by the developer enough to locate a median within 1 bp; this is a judgment rather than an established sampling-error guarantee. Replacing the derived inputs with the institution's own reconciled market data is proposed for any deployment. [BRIEF, choices 2–3; INST, Data lineage.]

**Table 3a. Additional estimator decisions for the 2026-09-18 snapshot. Reasons are developer statements, with alternatives and limits preserved.**

| Decision | Developer's reason | Alternative | Limitation |
|---|---|---|---|
| Median regression | Align robust path fitting with median quotes | Least squares, possibly after trimming | No least-squares comparison is supplied |
| Common path fitted at 2Y and longer with loading one | Separate fitted loadings are near one; shorter tenors move less | Fit each tenor's loading | Small long-tenor movement differences are ignored |
| At least eight prints across five hours to fit a loading | Sparse time coverage cannot separate a level from movement | Fit every tenor | Interpolated 2M loading is an assumption |
| Clip loadings to zero through one | Negative or above-long-tenor movement has no economic reading for this day in the developer's view | Unconstrained fit | 1M's −0.04 estimate becomes zero |
| Last traded hour defines close | Latest observed hour among path tenors | A fixed clock time such as 15:00 | Hour average/median reference is not a published closing fixing |
| Forward-start loading follows expiration distance | Simple rule without a forward-start model | Depend on both start and swap length | One-year-forward one-year swap receives a two-year-maturity loading |

Source: BRIEF, Smaller choices—Close adjustment; QUOTES, path/loading tables; DERIVE, regression and `loading_for` code. Print counts are counts, hours are distinct New York clock hours, loadings are dimensionless.

The selected 5Y +1 bp perturbation is motivated by a liquid mid-curve quote; the developer attributes the 1 bp convention to HW2008 p.78 and the 4.5Y cash flow to HW2006 p.126. Bumping every input and reporting a stability norm would answer a broader question. The maximum single-cash-flow money measure expresses an interpolation choice in desk currency units, while a diversified portfolio comparison would require an actual portfolio and could offset differences. Leave-one-tenor-out tests whether a tenor drives its own path adjustment; cross-validation over hours is an alternative, and fitted short-tenor loadings remain outside that check. [BRIEF, Smaller choices—Tests.]

The 1 bp outside-input tolerance is a developer judgment relative to observed quote dispersion; a tolerance by maturity bucket was an alternative. The paired sign test avoids a residual-size model and counts the closer curve; t-tests and Wilcoxon tests were alternatives. Agreement tolerances at rounding scale, half printed paper precision, numerical quadrature scale and daily-grid approximation scale address different numerical questions; tightening or loosening them would require a stated reason. None becomes a production acceptance policy merely by being used in the historical tests. [BRIEF, Check tolerances and other test choices; CHECKS, named checks.]

## 6. Implementation notes

The original source revision identifies all copied calculation code. `derive_quotes.py` reads the raw report and historical rate snapshots, filters and adjusts prints, and writes quote/fixing summaries. `bootstrap.py` supplies the common QuantLib instrument construction, builds A and B, evaluates grids, bumps and convention variants, and writes figures. `hagan_west.py` supplies B's interpolation, analytic integral, simple deposit/OIS representations, simultaneous calibration and the separately written log-linear class. `verify.py` supplies developmental checks and pricing outside the input set. `summarize_implementations.py` and `summarize_residuals.py` provide the retained supplements. `run_recorded.py` records stage execution and package checks. [WE, Calculation source; corresponding public script functions; ENV.]

**Table 4. Historical calculation receipts for the 2026-09-18 inputs. Times are UTC on 2026-10-02 and are supplied receipt values.**

| Stage | Start (UTC) | Finish (UTC) | Retained state / exit code (integer) |
|---|---|---|---|
| Quote derivation | 04:15:17.701293 | 04:15:18.632378 | completed / 0 |
| Bootstrap | 04:15:18.737172 | 04:15:20.209686 | completed / 0 |
| Verification | 04:15:20.299618 | 04:15:21.088870 | completed / 0 |
| Residual supplement | 04:15:21.165869 | 04:15:21.171377 | completed / 0 |
| Implementation supplement | 05:00:42.422897 | 05:00:49.979153 | completed / 0 |

Source: ENV, `evidence/environment/{derive-quotes,bootstrap,verify,residual-summary,implementation-summary}.json`, `started_utc`, `finished_utc`, `state`, `exit_code`. These identify historical executions only.

Historical Python is 3.12.14, QuantLib 1.43 and SciPy 1.18.1; stage receipts also identify NumPy 2.5.3, pandas 3.0.6, statsmodels 0.15.0 and matplotlib 3.11.2 where used. The residual summary uses Python's standard library. These are reported historical package identities, not a declaration that the same environment runs in an actual institution. Original script and input hashes are retained in the provenance manifest and receipt JSON. [ENV, `run_info.json`, `required_packages`, `installed_packages`; DISTANCE, metadata.]

Reproducibility has two levels. The selective public bundle allows inspection of instrument-level residuals, monthly DF differences, the 7,482-date implementation series, distance-bin arithmetic, and the code's conventions without executing a model. It cannot support a complete offline raw-data rerun: the original DTCC ZIP, raw extracted trade tables, original NY Fed snapshots and third-party PDFs are omitted. The original DTCC SHA-256 is `03a09a0ec9c5c5d1e233c2b663cb21d7c36b64573ef9ad5d587a874c5a7ba036`; original rate-snapshot hashes are listed in `original-raw-input-hashes.txt`. Reacquiring raw inputs, checking their identities and preparing the numerical environment would be separate work. A matching output hash confirms bytes, not the truth of every claim or independent model validation. [WE, Evidence scope and exclusions; PROVENANCE; RAW-HASHES.]

## 7. Testing and benchmarking

### 7.1 Calibration and developmental checks

**Table 5. All 22 calibration instruments, as-of 2026-09-18. Errors are implied minus input par rate (bp); the selected house criterion is absolute error ≤0.01 bp for every row in each method.**

| Instrument | Input rate (%) | A error (bp) | B error (bp) | Within criterion, A/B |
|---|---:|---:|---:|---|
| O/N | 3.850000 | -9.090e-12 | -9.090e-12 | yes / yes |
| T/N | 3.850000 | 5.238e-10 | -2.755e-10 | yes / yes |
| 1M | 3.897400 | 1.395e-11 | -1.263e-11 | yes / yes |
| 2M | 3.968300 | -4.191e-11 | -3.261e-12 | yes / yes |
| 3M | 4.031900 | -4.642e-11 | -2.498e-12 | yes / yes |
| 4M | 4.103100 | -1.467e-10 | -2.567e-12 | yes / yes |
| 6M | 4.205700 | -3.860e-09 | -2.082e-13 | yes / yes |
| 1Y | 4.498700 | -9.506e-12 | 1.388e-12 | yes / yes |
| 18M | 4.586800 | -6.245e-13 | -2.082e-13 | yes / yes |
| 2Y | 4.635800 | 9.190e-10 | 9.714e-13 | yes / yes |
| 3Y | 4.634200 | 6.939e-14 | 4.163e-13 | yes / yes |
| 4Y | 4.610700 | 4.372e-12 | 6.939e-14 | yes / yes |
| 5Y | 4.596800 | 4.857e-13 | 2.776e-13 | yes / yes |
| 6Y | 4.591700 | -2.776e-13 | 4.857e-13 | yes / yes |
| 7Y | 4.592300 | -2.082e-13 | -6.939e-14 | yes / yes |
| 8Y | 4.596400 | -5.551e-13 | 3.469e-13 | yes / yes |
| 9Y | 4.610700 | -5.412e-12 | 2.776e-13 | yes / yes |
| 10Y | 4.620800 | -1.527e-12 | 4.163e-13 | yes / yes |
| 12Y | 4.658200 | -3.483e-11 | 3.469e-13 | yes / yes |
| 15Y | 4.708500 | -1.388e-13 | 2.082e-13 | yes / yes |
| 20Y | 4.747700 | 0.000e+00 | 1.388e-13 | yes / yes |
| 30Y | 4.663700 | 3.518e-11 | 2.082e-13 | yes / yes |

Source: REPRICE, `repricing.csv`, all instrument rows; historical `bootstrap.py`. The criterion comparison is a current arithmetic check of frozen error columns, not a new calibration run.

Both methods satisfy the fictional criterion on every retained calibration instrument. The tiny nonzero errors measure numerical solver tolerance, not market error. Identical rounded implied/market percentages alone would not establish this result; the error columns preserve finer information. [REPRICE, all rows; RESULTS, `repricing_max_abs_error_bp`; HOUSE, Modeling choices and tests.]

**Table 6. Retained developmental checks, as-of 2026-09-18 unless a paper example or flat synthetic curve is specified. All “pass” results below are historical `verify.py` records, not tests performed anew.**

| Check and exact scope | Historical measured result (unit) | Supplied tolerance / interpretation |
|---|---|---|
| Paper example attributed to HW2006 p.102, factors versus printed discrete forwards | 4.55e-5 rate maximum difference | ≤5e-5 rate, half printed precision |
| Same example, B instantaneous forwards on (0,30] | Minimum 0.0087 rate; recovered discrete-forward error 0 | Positive; interval recovery ≤1e-14 rate |
| Paper example attributed to p.107 | 9.09e-6 rate printed-forward difference; minimum 0.0366 rate; interval error 0 | Same precision/recovery thresholds |
| B average-forward recovery, each interval, midpoint integration with 2,000 steps | 6.26e-10 rate maximum | ≤1e-7 rate; numerical quadrature check |
| B minimum instantaneous forward, curve of record | Approximately 0.039 rate | >0; differs from Act/360 overnight-forward quote |
| B continuity, either side of each node at 1e-9 years | 6.06e-11 rate maximum jump | ≤1e-6 rate |
| B intervals breaking local monotonicity | 0 intervals | ≤0 |
| Flat 4% curve priced into deposits/OIS and calibrated back | A 9.07e-11 bp; B 5.97e-12 bp maximum zero error | ≤1e-6 bp, monthly grid through 30Y |
| Separately written OIS par formula versus QuantLib helper on A | 2.08e-13 bp maximum | ≤1e-6 bp; common dates |
| B valued by QuantLib swap engine via daily business-day DFs | 1.27e-11 bp maximum input par error | ≤1e-4 bp; log-linear interpolation between those dates |
| B interpolation versus QuantLib (0,1), same nodes/discrete forwards | 3.47e-13 bp maximum forward difference | ≤1e-6 bp; collar inactive on this input |
| Anchor assumed versus subsequently realized fixing, 2026-09-18 and 2026-09-21 | 0 bp maximum | Exact retained match; ex-post check only |

Source: CHECKS, `checks` entries by full check name; `verify.py`, `paper_examples`, `properties`, `two_implementations`, `repricing`. Rounded measurements are supplied by `verify.md`.

The developer uses the p.102 example's capitalisation factors because the brief reports a printed 4Y yield of 4.40% inconsistent with the factor 1.22140276 and forward, which imply 5.00%. Choosing the printed yield would fail the retained check; the underlying paper-page interpretation has not been established from a paper copy in this evidence review. Thus this is evidence of an implemented example check with an explicit developer interpretation, not independent confirmation of the paper's alleged misprint. Developer tolerances reflect rounding, half printed precision or the numerical approximations used; tighter or looser criteria were possible and are not institutionally validated. [BRIEF, Smaller choices—Tests; CHECKS, paper-example entries; HW-CODE; VERIFY, `paper_examples`.]

### 7.2 Separately written implementation comparison

**Table 7. Method A only: retained implementation agreement, as-of 2026-09-18. The grid is exactly 7,482 SOFR-business-day observations from 2026-09-18 through 2056-09-26 inclusive.**

| Implementations compared | Maximum absolute DF difference (dimensionless) | Mean absolute DF difference (dimensionless) |
|---|---:|---:|
| QuantLib A versus `hagan_west.py` log-linear bootstrap | 1.925126724700e-13 | 1.917074862633e-14 |

Source: IMPL, `summary`, `implementation-comparison.csv`, and `summarize_implementations.py`. The mean averages magnitudes; it is not a signed mean. Both share quotes, fixing, schedules and financial conventions.

This checks two implementations of A, and its agreement is far smaller than the A-versus-B method disagreement. It does not cover B as a full separately scheduled implementation, calendar correctness, data truth or an institution-specific benchmark. The interpolation-only B comparison feeds the same nodes to both algorithms and the collar is inactive; the brief's account that QuantLib's positivity option floors endpoints rather than applying the selected interior collar remains a source-code attribution, not a proof for other regimes. [IMPL, limits; CHECKS, B interpolation entry; BRIEF, Worked example.]

### 7.3 Pricing outside the bootstrap input set

Define a signed residual as close-adjusted print rate minus curve par rate, in bp. A negative value means the curve rate is above the print estimate. Absolute residuals remove that direction; the sign of ( |r_A|-|r_B| ) indicates which method is closer. The retained 983 rows comprise 523 spot-start prints at used input tenors, 140 spot-start prints at other maturities, and 320 forward-start prints. Only the last two cohorts, 460 rows, are held out of the bootstrap. They still share the fitted intraday adjustment; this is not independent licensed benchmark validation. [CHECKS, `out_of_sample`, `paired_comparison`; RESIDUALS, `group` and residual columns; VERIFY, `out_of_sample`.]

The historical combined held-out upper-middle absolute residual is 0.459942 bp for A and 0.404662 bp for B (about 0.460 and 0.405 bp), each below the selected 1 bp median-absolute criterion. The statistic sorts 460 magnitudes and selects zero-based index 230 rather than averaging the two center observations. Ordinary two-center medians computed from the same rows are 0.458535 and 0.404222 bp; this convention difference does not affect the threshold finding. The criterion was chosen by the developer because adjusted within-tenor IQRs were smaller than 1 bp, not derived from independent loss or execution-cost evidence. [VERIFY, `med_a`, `med_b`; CHECKS, held-out check; RESIDUALS; BRIEF, smaller testing choices.]

**Table 8. Historical signed residual summaries, as-of 2026-09-18. Residual = adjusted print minus model par rate (bp), except the explicitly unadjusted late row. All numeric medians/quartiles use `verify.py`’s nearest-index rule.**

| Prints | n (count) | A median (bp) | A p25 / p75 (bp) | A within 1 bp (%) | B median (bp) | B p25 / p75 (bp) | B within 1 bp (%) |
|---|---:|---:|---:|---:|---:|---:|---:|
| spot start, input tenor | 523 | +0.00 | -0.28 / +0.24 | 91% | +0.00 | -0.28 / +0.25 | 91% |
| spot start, other maturity | 140 | -0.21 | -0.69 / +0.17 | 76% | -0.18 | -0.51 / +0.14 | 86% |
| forward start | 320 | -0.15 | -0.74 / +0.24 | 79% | -0.17 | -0.77 / +0.16 | 78% |
| all | 983 | -0.04 | -0.49 / +0.24 | 85% | -0.04 | -0.45 / +0.21 | 86% |
| held out, last two hours, unadjusted | 49 | -0.76 | -1.27 / +0.02 | 55% | -0.47 | -1.21 / +0.04 | 59% |

Source: CHECKS, `out_of_sample`, `verify.md`; RESIDUALS; `verify.py` `stats` selects sorted index `int(p*(n-1)+0.5)`. The input-tenor and all-print rows are descriptive, not held-out populations.

The 49 held-out prints in the last two traded New York hours are additionally compared at their unadjusted rates, with median signed residual −0.76 bp for A and −0.47 bp for B. Because the close adjustment is at most 0.2 bp there, this provides a limited robustness view not driven by applying the intraday adjustment to the observed print. The small sample remains from the same day, and rates used to calibrate the curves were still derived with the adjustment. Last-hour-only or last-three-hour choices were possible; the former would be thinner and the latter admit larger adjustment. [CHECKS, late row; VERIFY, `LATE_HOUR=15`; BRIEF, Late-print row.]

**Table 9. Historical paired comparisons, as-of 2026-09-18, separated by cohort and Act/365F years from as-of to expiration. Buckets are (0,1], (1,10], (10,last node], and (last node,99] years; “under 1Y” includes exactly 1Y in the code. P-values are dimensionless and two-sided.**

| Prints | n (count) | B closer (count) | A closer (count) | Sign-test p (two-sided) | Median of abs(A) minus abs(B) (bp) | Within 1 bp: A (count) | Within 1 bp: B (count) |
|---|---:|---:|---:|---:|---:|---:|---:|
| spot start, other maturity, under 1Y | 41 | 30 | 11 | 0.004 | +0.277 | 23 | 34 |
| spot start, other maturity, 1Y to 10Y | 70 | 37 | 33 | 0.720 | +0.003 | 63 | 64 |
| spot start, other maturity, 10Y to the last node | 28 | 11 | 17 | 0.345 | -0.041 | 21 | 23 |
| spot start, other maturity, beyond the last node | 1 | 1 | 0 | 1.000 | +3.668 | 0 | 0 |
| forward start, under 1Y | 27 | 11 | 15 | 0.557 | -0.000 | 23 | 20 |
| forward start, 1Y to 10Y | 172 | 101 | 70 | 0.022 | +0.001 | 138 | 135 |
| forward start, 10Y to the last node | 96 | 60 | 36 | 0.018 | +0.038 | 70 | 71 |
| forward start, beyond the last node | 25 | 18 | 7 | 0.043 | +0.082 | 22 | 22 |
| all held-out prints | 460 | 269 | 189 | <0.001 | +0.007 | 360 | 369 |
| spot start, other maturity, under 1Y, one per expiration date | 22 | 16 | 6 | 0.052 | +0.241 | 14 | 19 |
| spot start, other maturity, under 1Y, raw rates | 41 | 35 | 6 | <0.001 | +0.353 | 11 | 17 |
| spot start, other maturity, 6M to 1Y (exploratory) | 28 | 24 | 4 | <0.001 | +0.804 | 13 | 26 |

Source: CHECKS, `paired_comparison`; `verify.py`, `paired`, bucket selection and robustness rows. Positive median of absolute A minus absolute B favors B on that magnitude statistic. Even-count medians select the upper middle. The final three rows are separate robustness/exploratory populations, not added to the eight primary buckets.

B is closer on 30 of 41 noninput spot maturities under one year and A on 11, with historical median absolute-residual difference +0.277 bp. Those trades span 22 expiration dates: treating each date as one observation yields 16 versus 6 and two-sided sign-test p=0.052, substantially weaker than the trade-level p=0.004. On raw rates B is closer on 35 of 41, but fewer observations are within 1 bp for either curve. The 6M–1Y group gives B 24 of 28 and median difference +0.804 bp; it was selected after seeing the results and is exploratory. It coincides with the missing 9M calibration tenor, consistent with a curvature benefit in that gap rather than a universal advantage. [CHECKS, specified paired rows; BRIEF, choice 1.]

The other multi-print maturity groups have median differences from −0.041 to +0.082 bp. All held-out counts of B/A closer are 269/189 with two ties under the code's (10^{-9}) bp comparison tolerance; aggregate median difference is only +0.007 bp. The sign test measures which is closer, not how much; eight primary buckets and repeated expiration dates weaken a simple significance interpretation. A paired t-test or Wilcoxon test was an alternative, not a retained result. [CHECKS, paired comparison; VERIFY, `paired`; BRIEF, Paired sign test.]

**Table 10. Descriptive end-date distance bins, as-of 2026-09-18. All residual statistics are absolute residuals in bp and counts are prints; the two held-out cohorts remain separate.**

| Cohort | End-date distance | Prints (count) | A median absolute residual (bp) | A IQR (bp) | B median absolute residual (bp) | B IQR (bp) |
|---|---|---:|---:|---:|---:|---:|
| spot start, other maturity | 0–30 days | 37 | 0.407 | 0.464 | 0.292 | 0.373 |
| spot start, other maturity | 31–90 days | 72 | 0.548 | 0.914 | 0.364 | 0.602 |
| spot start, other maturity | >90 days | 31 | 0.564 | 0.477 | 0.354 | 0.253 |
| forward start | 0–30 days | 90 | 0.578 | 0.820 | 0.598 | 0.841 |
| forward start | 31–90 days | 163 | 0.428 | 0.704 | 0.388 | 0.732 |
| forward start | >90 days | 67 | 0.443 | 0.560 | 0.366 | 0.475 |

Source: DISTANCE, `residual-distance.json` `summary`, `residual-distance.md`, historically calculated by `summarize_residuals.py` from RESIDUALS and NODES. New arithmetic inspection of those frozen rows reproduces these medians and IQRs; no financial model was rerun.

Distance is the absolute number of calendar days between a print's expiration and the nearest used OIS payment pillar; O/N and T/N are excluded. Bins are 0–30 inclusive, 31–90 inclusive, and greater than 90. Interquartile range (IQR) is the 75th percentile minus the 25th of absolute residuals using inclusive linear interpolation of empirical order statistics. These descriptive bins preserve spot and forward starts separately and are not the same intervals as the maturity buckets. In particular, B's median is higher in the forward-start 0–30-day bin despite its lower aggregate median. Nearness of an expiration does not represent start-date differences, financial risk, independent observations or generalization. [DISTANCE, `distance_definition`, `quartiles`, `summary`; `summarize_residuals.py`, `summarize`.]

### 7.4 Sensitivity, locality and extrapolation

**Table 11. Historical convention perturbations versus the base curve, as-of 2026-09-18, maximum absolute DF change on the 361-date monthly grid (dimensionless). Each row rebuilds with the specified alternative convention in the retained calculation.**

| Change versus base | A maximum absolute DF change | B maximum absolute DF change |
|---|---:|---:|
| Payment lag 0 rather than 2 business days | 1.32579e-5 | 4.79181e-6 |
| Quarterly rather than annual payments | 0.00661621 | 0.00661321 |
| Fixed leg interpreted Act/365F rather than Act/360 | 0.00500070 | 0.00500142 |
| Overnight anchor raised 1 bp | 1.10710e-6 | 1.31062e-6 |

Source: RESULTS, `convention_sensitivities`; BOOT, convention variants. These are scenario sensitivities to changed assumptions, not evidence that the alternative convention is correct or a market risk distribution.

The schedule and day-count effects exceed the observed A/B maximum on this grid, so comparing interpolators while assuming schedules correct can miss a larger error source. Payment lag and annual-versus-quarterly schedules were chosen as convention sensitivities because the developer considered them insufficiently fixed by prints; roll convention and calendar variants were not measured. The Act/365F fixed-leg sensitivity is implemented by scaling fixed rates by 365/360 in otherwise Act/360 instruments. [BOOT, `act365`, `qtr`, `lag0`; BRIEF, Assumptions and Convention sensitivities.]

![Historical 5Y quote-bump response](figures/fig_bump_5y.png)

*Figure 3. As-of 2026-09-18 response of overnight forwards, bumped minus base (bp), to +1 bp in the 5Y OIS quote, against years from as-of. Frozen `fig_bump_5y.png`, `bootstrap.py`, `bump_5y_forward_change.csv`. The inclusive inside window uses actual 4Y/6Y payment nodes 2030-09-25 and 2032-09-24. B's larger outside response demonstrates less locality on this perturbation; it does not define a full risk-system bump policy.*

![Historical cash-flow sensitivity ladder](figures/fig_delta_ladder.png)

*Figure 4. As-of 2026-09-18 PV change (USD), bumped minus base, for USD 100 million paid 2031-03-24 (4.5Y from spot), under a separate +1 bp bump to each of twenty OIS inputs. Frozen `fig_delta_ladder.png`, `bootstrap.py`, `delta_ladder_4p5y.csv`. Signed bars are PV changes, not hedge notionals. Section 5's outside share sums their absolute magnitudes, so offsetting bars are not netted.*

The unbumped cash-flow DFs are 0.81397156 (A) and 0.81397425 (B), giving A-minus-B PV −USD 269.46 from full-precision results. Beyond 30Y the extrapolated 35Y overnight forwards are 4.2248% and 3.9957%; at the last node the displayed jump is 0.00 bp for both, with A's retained numerical difference about (8.0\times10^{-10}) bp. Rounded zero is not exact arithmetic identity. Continuity there says nothing about whether the continued level is supported by market data. [RESULTS, `off_node`, `extrapolation`, `forward_smoothness`.]

## 8. Limitations and compensating controls

Every control below is a **fictional proposed** design from HER-scenario-1, with no operating-effectiveness evidence. A control's existence in the dossier is not proof that it operates. The proposed responses withhold or restrict use and record exceptions; they do not retrospectively authorize the historical model. [INST, Proposed controls and monitoring thresholds.]

**Table 12. Reliance limitations, proposed fictional responses and residual gaps for the 2026-09-18 demonstration.**

| Limitation and consequence | Supplied proposed control and owner | Residual limitation / effectiveness |
|---|---|---|
| Calibration/solver failure can prevent reliable curve construction | C-01: every input ≤0.01 bp; developer holds/diagnoses breach, coordinator considers reassessment | One successful snapshot is no convergence guarantee; no operating evidence |
| Thin/noisy estimates, surviving off-market prints and unusual matched stubs | C-02: ≥5 qualifying prints, IQR ≤2 bp; Market Data Controls Lead marks unsupported tenor unavailable | Count/dispersion rule does not certify a par print or stale-data tolerance |
| Derived public execution medians are mistaken for a licensed close | C-03: label estimates and require licensed benchmark before production; model owner | No executable benchmark, feed entitlement or reconciliation record |
| Pricing outside inputs shares close adjustment and exploratory selection | C-04: median absolute residual ≤1 bp with separate spot/forward summaries; developer investigates breach | Passing cannot remove common adjustment, one-date dependence or multiplicity |
| Sparse 20Y–30Y input, propagated end-rule shape | C-05: model owner reviews shape, concentration and A/B disagreement each future candidate | No quantitative production exposure limit or independently observed long-end shape |
| Beyond-node values rely solely on extrapolation | C-06: label sandbox extrapolation, prohibit proposed production beyond node pending policy; model owner | No actual enforcement or authorized long-rate assumption |
| B sensitivity spillover and partial locality test | C-07: developer reviews 5Y bump and full 4.5Y ladder, proposes multi-date all-input campaign | One cash flow and one forward-bump window do not cover a portfolio |
| One positive-rate date; negative-rate and stressed regimes untested | C-08: stop extending this snapshot's claims to negative rates; model owner commissions separate study | No stress evidence, regime threshold or negative-rate admissibility assessment |
| Shared calendar, payment and accrual defects | C-09: coordinator obtains separate schedule/calendar reconciliation | Both methods may agree on wrong dates; holiday-list equivalence unverified |
| Code, data transformation, convention or library changes | C-10: developer retains immutable change record and impact assessment | No actual release or recovery evidence |
| Document/evidence drift | C-11: coordinator matches document, manifest and assessment hashes | A hash does not itself verify claims or transfer an assessment to changed bytes |

Source: INST, C-01–C-11; model limitations in BRIEF choices and Assumptions; numerical evidence in Sections 5 and 7.

Central-bank meeting and year-end event steps are excluded. No dedicated proposed event-step control is supplied; reliance where those discontinuities matter is outside scope. Multi-curve, alternative collateral and other-currency uses are also outside scope rather than compensated by a demonstrated control. The hourly close model ignores multiple intraday factors and finer-time variation; C-03's proposed data replacement and C-04's residual review only partly address that issue. No historical control-effectiveness record is supplied for any of these risks. [BRIEF, Scope and exclusions, choice 2; INST, controls and data lineage.]

Monitoring is proposed at each candidate build for C-01–C-04 and C-11, quarterly for limitations/dependencies, and on convention, data, model or regime changes. Escalation records the exception and withholds the proposed use. A missing required input, stale feed input or incompatible convention holds a candidate; a feed-specific service agreement and actual stale-quote tolerance remain unspecified. A changed snapshot needs new calibration and assessment of eligibility, not automatic reuse of the historical passes. [INST, Data lineage and proposed daily process; Proposed controls and monitoring thresholds.]

[INSTITUTIONAL GAP: Actual monitoring implementation, breach and resolution records, control effectiveness and real operational ownership are unavailable.]

[INSTITUTIONAL GAP: Actual fixed-close snapshot policy, feed-specific service agreement and stale-quote tolerances are unavailable.]

[INSTITUTIONAL GAP: Actual extrapolation authorization and production policy beyond the last node are unavailable.]

[INSTITUTIONAL GAP: The full risk-system bump methodology and approved exposure limits are unavailable.]

## 9. Assumptions register

**Table 13. Assumptions behind the historical 2026-09-18 curve and quote derivation. Monitoring references are fictional proposed controls, not observed operation.**

| Assumption | Supplied reason and evidence | Consequence if false | Monitoring or unresolved work |
|---|---|---|---|
| SOFR is both projection and discount rate for the covered USD cash flows | Scope choice, BRIEF 6; implemented par formula | Wrong collateral/basis treatment | Enforce stated scope; actual collateral/use inventory absent |
| Last known 3.85% fixing holds to spot | Published anchor; ex-post retained next fixings match, CHECKS anchor entry | Front-end DFs change; +1 bp sensitivity in Table 11 | C-03 data replacement/reconciliation; real fixing availability policy absent |
| A004 denotes Actual360 for selected fixed legs | BRIEF attributes definition to CFTC Appendix C p.55; 4,211 coded trades in QUOTES | Up to about 0.005 DF shift under alternative in retained sensitivity | C-09 independently reconcile conventions; primary specification claim unchecked here |
| Both legs follow annual/single-payment schedules | Reported frequency filters, BRIEF and QUOTES | Quarterly sensitivity about 0.0066 DF; cash-flow model differs | C-02 eligibility; C-09 schedule reconciliation |
| Two-business-day payment lag is suitable | BRIEF attribution to ARRC p.17; implemented BOOT helper setting | Table 11 lag sensitivity; par cancellation differs by schedule | C-09; actual production convention evidence absent |
| QuantLib SOFR calendar matches intended business days | Shared helper convention; developer cites the SIFMA-determined Treasury repo calendar and QuantLib implementation; direct NY Fed holiday-list use was the alternative | Both curves can misdate the same accrual/payments | C-09 separate holiday/calendar reconstruction; not performed |
| Non-package, no-upfront filtered fixed rates are suitable par observations | Developer economic reasoning and upfront-rate-rounding evidence | Residual contamination and biased quote medians | C-02/C-03; 20 off-market matched prints remain |
| One common hourly factor times bounded loading captures tenor drift | Path and loading tables; long-tenor fitted loadings 0.97–1.08 | Biased close levels and shared errors in held-out comparison | C-03/C-04; leave-one-tenor-out tests path tenors only |
| Sparse short-tenor loadings may be interpolated and forwards use expiration loading | At least eight prints/five hours requirement; simple forward-start rule | Level/move not separable, start/length variation ignored | Separate loading/forward-start analysis not supplied; C-03 replacement proposed |
| Five prints and IQR ≤2 bp make a usable estimate | Developer eligibility judgment; no IQR rejection this date | Thin median unstable despite rule, long-end concentration | C-02/C-05; no sampling-error guarantee |
| Chosen interpolation is acceptable between observations | B shape rationale versus A; Table 3 | Off-node prices and sensitivities depend on unsupported shape | C-05/C-07; broader competitor, date and risk tests unresolved |
| Positive discrete forwards are the relevant regime for B's collar | Developer collar rationale; zero binding nodes historically | Negative-rate shape behavior not established | C-08 separate study before extending conclusions |
| Final forward can continue beyond the last node | Developer's HW2006 attribution; explicit `integral` code | Extrapolated DFs differ and lack quote support | C-06; actual authorization absent |

Source: BRIEF, Assumptions and named choices; QUOTES; RESULTS; BOOT; HW-CODE; INST. Neither a stated rationale nor a historical pass demonstrates that an assumption remains valid on another date.

## 10. Governance and institutional record

All facts in this section describe the fictional Harbor Example Research dossier. Inventory HER-QM-017 (simulated), revision `sofr-snapshot-2026-09-18`, is Tier 2 under a fictional policy because it is a material quantitative method limited to sandbox use with no production exposure. Actual institutional tiering would depend on real reliance, exposures and controls. The scenario status is proposed sandbox methodology awaiting documentation assessment; approved uses are none. Proposed consumers are notebook `curve-research` and report `rates-methodology`, with no actual integration. [INST, Inventory and intended use.]

Fictional Maya Chen represents the Rates Research Lead/model owner, Daniel Ortiz the Curve Research Analyst/developer, Priya Shah the Model Review Lead/review coordinator, and Elias Brooks the Market Data Controls Lead/monitoring owner. No review has been performed by Priya Shah in the scenario. The Research Model Committee is the accountable authorization role and its decision is pending. These names do not identify actual reviewers of this document. [INST, inventory role rows.]

Scenario change HER-CR-001 concerns the first methodology document from fixed September data; model code, filters and frozen numerical outputs are not amended. The proposed system is the fictional HER research sandbox. Version history begins with this newly created demo-v1 document dated 2026-10-12; there is no supplied prior document revision. No actual implementation ticket, vendor agreement, production release or recovery test exists in the supplied evidence. [INST, Implementation and change record.]

Open fictional findings are F-01 licensed benchmark missing (blocking production consideration), F-02 multi-date/stressed assessment absent, F-03 convention/calendar reconciliation not independent, F-04 controls have no operating evidence, and F-05 document assessment pending at dossier authoring. Closing them requires, respectively, identified benchmark/tolerance/discrepancy analysis; an authorized retained campaign and separate assessment; separately sourced calendar/schedule work; real control ownership and performance records; and a separate assessment of the exact final document hash. A document readiness self-check cannot retrospectively close findings on behalf of fictional people. [INST, Open findings and disposition.]

The sign-off block is not signed. Approval conditions are unknown because the scenario committee has made no decision; resolving F-01 through F-04 and separately assessing the document are proposed conditions for consideration, not attached approval conditions. Exception handling means recording the breach, withholding the proposed use and escalating to the designated roles. Changing a tolerance to fit a result is not the supplied process. Actual institutional validation, authorization and any applicable regulatory judgment are separate from documentation readiness. [INST, Approval and required real-world gaps; Proposed controls; HOUSE, Institutional facts.]

[INSTITUTIONAL GAP: Actual institutional inventory ID, materiality tier and tier rationale are unavailable.]

[INSTITUTIONAL GAP: Actual internal market-data feed lineage, licenses, entitlements and reconciliation evidence are unavailable.]

[INSTITUTIONAL GAP: Actual model owner, developer, reviewer, committee authorization and role assignments are unavailable.]

[INSTITUTIONAL GAP: Real prior validation findings, disposition history and independently performed assessment are unavailable.]

[INSTITUTIONAL GAP: Actual approval decision and approval conditions are unavailable.]

[INSTITUTIONAL GAP: A real licensed benchmark curve, immutable snapshot, comparison tolerance and discrepancy analysis are unavailable.]

[INSTITUTIONAL GAP: Actual implementation/change ticket, production deployment record and recovery test are unavailable.]

[INSTITUTIONAL GAP: Real institutional signatures and sign-off evidence are unavailable.]

## 11. Third-party components

QuantLib 1.43 supplies A's bootstrap, SOFR calendar, deposit/OIS helpers, schedules and the comparison swap engine. The developer's B code supplies its own interpolation, integral and par calculation but consumes QuantLib's instrument dates. SciPy 1.18.1 solves the simultaneous nonlinear system. NumPy supports arrays; pandas handles trade transformation; statsmodels fits median regression; matplotlib renders historical figures. Versions in Section 6 are from the stage receipts and should not be replaced by current documentation defaults. [ENV, package fields; BOOT and HW-CODE imports and calls; DERIVE imports/path fit.]

The developer attributes B to Hagan and West (2008), the par relation and examples to Hagan and West (2006), broader multi-curve context to Ametrano and Bianchetti (2013), payment/compounding conventions to ARRC (2021), and historical futures/discounting context to CME (2020). These are external rationales recorded in the brief. Their papers were not acquired as answer sources for this document; titles and locators alone are not confirmation of contents. In particular the CME account is a historical planned transition, not evidence of eventual implementation or current practice, and no legal or supervisory applicability claim is made. [BRIEF, source-specific attributions; SOURCES, Original public sources.]

Shared dependency agreement is useful for finding coding errors but leaves shared data and schedule error unresolved. The retained default/paper-setting QuantLib variants illustrate that library settings and solver strategy matter. The developer's analytic integration `_G`, simultaneous calibration and selected interior collar are specific implementation choices, rather than an unmodified invocation of QuantLib's default interpolation. Copied developer code and authored evidence carry the bundle's MIT notice; third-party data and papers retain their own terms. [HW-CODE; RESULTS, variants; WE, license and scope paragraphs; LICENSE.]

## 12. References

Keys below identify the permitted public origins. Local copies under `inputs/` preserve the actually retrieved bytes; public origins, not private repository paths, resolve substantive references. A key plus a named function, field, row or section in the text identifies the claimed relationship. All numerical outputs concern the 2026-09-18 snapshot unless explicitly identified as a synthetic/paper check; historical receipts are dated 2026-10-02.

- **BRIEF:** [model-brief.md](https://openprose.ai/demos/sofr/v1/model-brief.md).
- **QUOTES:** [evidence/quotes.md](https://openprose.ai/demos/sofr/v1/evidence/quotes.md).
- **RESULTS:** [evidence/results.json](https://openprose.ai/demos/sofr/v1/evidence/results.json).
- **CHECKS:** [evidence/verify.json](https://openprose.ai/demos/sofr/v1/evidence/verify.json).
- **REPRICE:** [evidence/repricing.csv](https://openprose.ai/demos/sofr/v1/evidence/repricing.csv).
- **DISTANCE:** [evidence/residual-distance.json](https://openprose.ai/demos/sofr/v1/evidence/residual-distance.json).
- **IMPL:** [evidence/implementation-comparison.json](https://openprose.ai/demos/sofr/v1/evidence/implementation-comparison.json).
- **WE:** [source-manifest.md](https://openprose.ai/demos/sofr/v1/source-manifest.md).
- **ENV:** [evidence/run_info.json](https://openprose.ai/demos/sofr/v1/evidence/run_info.json).
- **HOUSE:** [house.md](https://openprose.ai/demos/sofr/v1/house.md).
- **INST:** [institution.md](https://openprose.ai/demos/sofr/v1/institution.md).
- **FIXING:** [evidence/sofr_fixing.json](https://openprose.ai/demos/sofr/v1/evidence/sofr_fixing.json).
- **NODES:** [evidence/nodes.csv](https://openprose.ai/demos/sofr/v1/evidence/nodes.csv).
- **GRID:** [evidence/df_grid_monthly.csv](https://openprose.ai/demos/sofr/v1/evidence/df_grid_monthly.csv).
- **RESIDUALS:** [evidence/oos_residuals.csv](https://openprose.ai/demos/sofr/v1/evidence/oos_residuals.csv).
- **BUMP:** [evidence/bump_5y_forward_change.csv](https://openprose.ai/demos/sofr/v1/evidence/bump_5y_forward_change.csv).
- **PROVENANCE:** [provenance.json](https://openprose.ai/demos/sofr/v1/provenance.json).
- **RAW-HASHES:** [model-source/original-raw-input-hashes.txt](https://openprose.ai/demos/sofr/v1/model-source/original-raw-input-hashes.txt).
- **BOOT:** [model-source/bootstrap.py](https://openprose.ai/demos/sofr/v1/model-source/bootstrap.py).
- **HW-CODE:** [model-source/hagan_west.py](https://openprose.ai/demos/sofr/v1/model-source/hagan_west.py).
- **DERIVE:** [model-source/derive_quotes.py](https://openprose.ai/demos/sofr/v1/model-source/derive_quotes.py).
- **VERIFY:** [model-source/verify.py](https://openprose.ai/demos/sofr/v1/model-source/verify.py).
- **SOURCES:** [sources.md](https://openprose.ai/demos/sofr/v1/sources.md).
- **LICENSE:** [LICENSE.txt](https://openprose.ai/demos/sofr/v1/LICENSE.txt).
- **Companion tables and calculation sources:** [results.md](https://openprose.ai/demos/sofr/v1/evidence/results.md), [verify.md](https://openprose.ai/demos/sofr/v1/evidence/verify.md), [quotes.csv](https://openprose.ai/demos/sofr/v1/evidence/quotes.csv), [implementation comparison series](https://openprose.ai/demos/sofr/v1/evidence/implementation-comparison.csv), [implementation supplement script](https://openprose.ai/demos/sofr/v1/model-source/summarize_implementations.py), [residual supplement script](https://openprose.ai/demos/sofr/v1/model-source/summarize_residuals.py), [recording wrapper](https://openprose.ai/demos/sofr/v1/model-source/run_recorded.py).
- **ENV stage receipts:** [quote derivation](https://openprose.ai/demos/sofr/v1/evidence/environment/derive-quotes.json), [bootstrap](https://openprose.ai/demos/sofr/v1/evidence/environment/bootstrap.json), [verification](https://openprose.ai/demos/sofr/v1/evidence/environment/verify.json), [residual summary](https://openprose.ai/demos/sofr/v1/evidence/environment/residual-summary.json), [implementation summary](https://openprose.ai/demos/sofr/v1/evidence/environment/implementation-summary.json).
- **Figure data:** [forward_daily.csv](https://openprose.ai/demos/sofr/v1/evidence/forward_daily.csv), [delta_ladder_4p5y.csv](https://openprose.ai/demos/sofr/v1/evidence/delta_ladder_4p5y.csv). The frozen PNGs are individually listed under WE.

External citations embedded in the developer's reasoning remain **attributions to BRIEF**, not independently checked paper claims: [HW2008, §6 and §6.1, equations 22–34](https://downloads.dxfeed.com/specifications/dxLibOptions/HaganWest.pdf); [HW2006, pp.92, 102, 107, 109, 124–126](https://www.deriscope.com/docs/Hagan_West_curves_AMF.pdf); [ARRC2021, pp.5, 17, 31, 35](https://www.newyorkfed.org/medialibrary/Microsites/arrc/files/2021/users-guide-to-sofr2021-update.pdf); [CFTC historical specification, Appendix C p.55](https://business.cch.com/srd/TechSpecParts4345forNPRMCLEAN12132023forOPA.pdf); [CME August 2020 methodology, §§1, 1.3, 2.1](https://www.cmegroup.com/trading/interest-rates/files/cme-sofr-discounting-and-pa-transition-whitepaper.pdf); [AB2013 abstract](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2219548); [DTCC PPD guide p.7](https://kgc0418-tdw-data-0.s3.amazonaws.com/gtr/static/gtr/docs/RT_PPD_quick_ref_guide.pdf); and [QuantLib 1.43 convex-monotone implementation](https://raw.githubusercontent.com/lballabio/QuantLib/v1.43/ql/math/interpolations/convexmonotoneinterpolation.hpp). Source-of-record origins for data provenance are [New York Fed SOFR](https://www.newyorkfed.org/markets/reference-rates/sofr) and [the original DTCC 2026-09-18 archive](https://pddata.dtcc.com/ppd/api/report/cumulative/cftc/CFTC_CUMULATIVE_RATES_2026_09_18.zip); the archive was not reacquired or reconstructed.

### SOFR data republication notice

The SOFR data is subject to the [Terms of Use posted at newyorkfed.org](https://www.newyorkfed.org/privacy/termsofuse). The New York Fed is not responsible for publication of the SOFR data by OpenProse, does not sanction or endorse any particular republication, and has no liability for your use.

OpenProse is not affiliated with the New York Fed. The New York Fed does not sanction, endorse, or recommend any products or services offered by OpenProse.
