diff --git a/changelog.d/20260905-deaner-ku-paper-review.md b/changelog.d/20260905-deaner-ku-paper-review.md
new file mode 100644
index 000000000..78514b573
--- /dev/null
+++ b/changelog.d/20260905-deaner-ku-paper-review.md
@@ -0,0 +1,6 @@
+### Documentation
+- Added a full-paper and appendix review of Deaner and Ku's *Causal Duration
+  Analysis with Diff-in-Diff* (arXiv 2405.05220v2), including identification,
+  bootstrap inference, diagnostics, deferred extensions, and a pinned reference-code
+  audit. Records requirements for a prospective two-group `DurationDiD` estimator;
+  no estimator or public API is introduced.
diff --git a/docs/methodology/papers/deaner-ku-2026-review.md b/docs/methodology/papers/deaner-ku-2026-review.md
new file mode 100644
index 000000000..36aba8649
--- /dev/null
+++ b/docs/methodology/papers/deaner-ku-2026-review.md
@@ -0,0 +1,1098 @@
+# Paper Review: Causal Duration Analysis with Diff-in-Diff
+
+**Authors:** Ben Deaner and Hyejin Ku
+
+**Citation:** Deaner, B., & Ku, H. (2026). *Causal Duration Analysis with Diff-in-Diff*. arXiv:2405.05220v2, working paper revision. [Version record](https://arxiv.org/abs/2405.05220v2), [original PDF](https://arxiv.org/pdf/2405.05220v2).
+
+**PDF reviewed:** `.workflow/sources/deaner-ku-2026-v2.pdf`
+
+**Review date:** September 5, 2026
+
+The reviewed version was submitted May 25, 2026; its title page is dated May 26,
+2026. The original arXiv submission was May 8, 2024. The author's
+[research page](https://bendeaner.wordpress.com/research/), checked September 5,
+2026, lists the paper as revise and resubmit at *Quantitative Economics*.
+This is not evidence of acceptance or journal publication; the citation above
+deliberately identifies the reviewed working paper revision.
+
+This document is a methodology foundation, not a shipped estimator. The
+maintainer-approved first implementation is a two-group `DurationDiD` with common
+treatment timing, common dynamics and proportional hazards, individual bootstrap
+inference, and pre-treatment diagnostics. Covariate adjustment and staggered
+adoption are deferred from that implementation. They, and all other appendices,
+are covered below. Library recommendations are labeled separately from paper
+results and reference-software behavior.
+
+## Source provenance and coverage
+
+All 50 PDF pages were read through independent overlapping extractions of
+pp.1–20, 16–35, and 31–50. Overlaps were reconciled; no missing or illegible region
+was found. Original pages containing figures, tables, algorithms, and ambiguous
+equations were also rendered and visually inspected. The page references below
+use the PDF's printed page numbers, which match its page order.
+
+| Pages, inclusive | Coverage |
+|---|---|
+| 1–5 | Title, abstract, introduction, related literature |
+| 5–9 | Section 1, population, potential outcomes, ordinary DiD, Figure 1.1 |
+| 9–20 | Section 2, hazard restrictions, Theorems 1–3, Proposition 1, specification and censoring |
+| 20–24 | Section 3, estimation, weighting, bootstrap, diagnostics, asymptotic qualifications |
+| 24–28 | Section 4, application and Figures 4.1–4.4, including conclusions before References on p.28 |
+| 28–32 | Complete bibliography, including references before Appendix A on p.32 |
+| 32–38 | Appendix A.1–A.4.1: motivation, conditional identification, staggered adoption, semiparametric adjustment |
+| 38–40 | Appendix B and Algorithms 1–2, after the estimator on p.38 and before Appendix C on p.40 |
+| 40–44 | Appendix C, simulation design, Tables 1–2, Figures C.1–C.4 |
+| 45 | Appendix D, proportional-hazards application and Figures D.1–D.2 |
+| 46–50 | Appendix E, all identification proofs |
+
+The review snapshot contains seven unchanged originals. The code snapshot is
+pinned to commit `202e92ef222bf7d250c26c98e8fb42e151223c71` of the
+[authors' repository](https://github.com/ben-deaner-teaching/Duration-DiD/tree/202e92ef222bf7d250c26c98e8fb42e151223c71).
+The following paths are relative to `.workflow/sources/`; that local cache and
+its seven-entry `.workflow/sources.json` are review evidence, not public package
+assets. Each filename links to the canonical original bytes. All seven inventory
+digests were verified against the cached files.
+
+| Cached original | SHA-256 |
+|---|---|
+| [deaner-ku-2026-v2.pdf](https://arxiv.org/pdf/2405.05220v2) | `d40df031669d5ceb2a3fa39c9e904c708fd7f60e7f60f460ab4adab02945f639` |
+| [duration-did-reference-202e92ef/README.md](https://raw.githubusercontent.com/ben-deaner-teaching/Duration-DiD/202e92ef222bf7d250c26c98e8fb42e151223c71/README.md) | `d7bcfb98c9498a6701465d620cc2e3c6d1c65e0ce3377245d856228f147b9b7b` |
+| [duration-did-reference-202e92ef/UIP_application.do](https://raw.githubusercontent.com/ben-deaner-teaching/Duration-DiD/202e92ef222bf7d250c26c98e8fb42e151223c71/UIP_application.do) | `5a4ee15a3a8cbaabed725ea4973cd64f2c533581db85abb1cde45c907d09be2f` |
+| [duration-did-reference-202e92ef/UIP_application_revision_R2.m](https://raw.githubusercontent.com/ben-deaner-teaching/Duration-DiD/202e92ef222bf7d250c26c98e8fb42e151223c71/UIP_application_revision_R2.m) | `092ed6ab096a2127693221708df17f4e56ac52d5f235a82c1d2fffdd0fefa8d5` |
+| [duration-did-reference-202e92ef/durationDiD.m](https://raw.githubusercontent.com/ben-deaner-teaching/Duration-DiD/202e92ef222bf7d250c26c98e8fb42e151223c71/durationDiD.m) | `3449d12a4b9e2aca7f943493c9f303130f90d348d089fd88604ac352280f43df` |
+| [duration-did-reference-202e92ef/durationdid.ado](https://raw.githubusercontent.com/ben-deaner-teaching/Duration-DiD/202e92ef222bf7d250c26c98e8fb42e151223c71/durationdid.ado) | `6e66e9f19ae3cf384d08209d52abcbaedcc85196f97e524e168ee7eaca58aa0e` |
+| [duration-did-reference-202e92ef/montecarlo_revision.m](https://raw.githubusercontent.com/ben-deaner-teaching/Duration-DiD/202e92ef222bf7d250c26c98e8fb42e151223c71/montecarlo_revision.m) | `f7a01d3a46762c6d05eac88014c20dc60da468603367746dcb02da9fa1b5c022` |
+
+## Methodology Registry Entry
+
+This prospective entry follows [REGISTRY.md](../REGISTRY.md)'s format. Keep it
+here until the estimator ships; no public API, catalog, or production registry
+entry is added by this review.
+
+## DurationDiD
+
+**Primary source:** Deaner and Ku (2026), arXiv:2405.05220v2, cited above.
+
+**Key implementation requirements:**
+
+*Assumption checks / warnings:*
+
+- Absorbing binary outcomes, fixed populations, no anticipation before the common
+  intervention, and unaffected controls. The chosen restriction is on untreated
+  hazards, not on cumulative outcome levels.
+- At least two pre-treatment dates, including the baseline, and one post-date.
+  A nontrivial diagnostic requires at least three pre-dates. Proportional hazards
+  additionally needs informative positive control hazard increments.
+- Preserve the initial treated survival probability and whole-group denominator.
+  Check survival/log domains and counterfactual validity; never silently clip.
+- Individual bootstrap inference requires independence across individuals and
+  regular interior identification. Group intervention identification alone does
+  not justify this statistical independence assumption.
+
+*Estimator equation (Theorem 1; Equations 3.1–3.4, with the PH choice below):*
+
+```text
+S_hat[k,t] = 1 - mean(Y[i,t] | G[i]=k)
+R_hat[k,t] = -log(S_hat[k,t])
+D_hat[k,t] = R_hat[k,t] - R_hat[k,1]
+H_hat[k,t] = D_hat[k,t] / (t-1), t>1
+
+CD: c_hat = sum_t alpha[t] (H_hat[1,t] - H_hat[2,t])
+    R0_hat[1,t] = R_hat[1,1] + D_hat[2,t] + (t-1)c_hat
+PH: c_hat = sum_t alpha[t] D_hat[1,t]/D_hat[2,t]
+    R0_hat[1,t] = R_hat[1,1] + c_hat D_hat[2,t]
+
+tau_hat[t] = exp(-R0_hat[1,t]) - S_hat[1,t], t>tstar
+```
+
+Here `tstar` is the last untreated date, `alpha` sums to one over selected usable
+pre-dates, and positive `tau` means increased cumulative absorption/exit.
+
+- **Note:** PH above is the proposed mean-of-ratios estimator, supported by the
+  authors' code and Theorem 1. It is not literal Equation 3.5, nor the
+  finite-sample PH specialization of Equation 3.6. The distinctions and a
+  reproducible reciprocal-coefficient check appear below.
+- **Note:** Equal weights on eligible pre-dates and Appendix B's fixed-anchor
+  diagnostic are proposed first-version defaults. Paper/application alternatives
+  remain documented; they are not silently interchanged.
+
+*With covariates / doubly robust:*
+
+Deferred from the initial estimator. Sections 2.2/3.1 and Appendix A offer
+conditional identification, initial-survivor balancing, and semiparametric
+adjustment. None is presented as doubly robust. Their distinct assumptions and
+estimators are developed below.
+
+*Standard errors (Section 3.2; Appendix B):*
+
+- Resample whole individual histories with replacement from the pooled sample;
+  re-estimate all quantities. No wild/multiplier weight distribution applies.
+- Bootstrap SD; symmetric bands from centered absolute bootstrap deviations,
+  with post-period maxima for simultaneous bands. Recommend 1,000 draws and
+  sample SD (`ddof=1`); these are library defaults, not universal paper rules.
+- Serial dependence within a person is preserved. Arbitrary higher-level
+  clustering, survey inference, and censoring-adjusted inference are outside
+  this first version.
+
+*Edge cases:*
+
+- Invalid panel, absorption reversal, missing cells, or unsupported design:
+  explicit validation error.
+- Zero survival or PH denominators: identify unsupported periods/draws; no
+  epsilon replacement or silent sample alteration.
+- Zero/invalid SE: joint undefined test statistics, p-values, and intervals;
+  retain point estimates and failure diagnostics when meaningful.
+- **Note:** A fixed-draw failure policy and simultaneous-family validity must be
+  explicit; the conservative first-version recommendation is detailed below.
+
+*Algorithm (Theorem 1 and Algorithms 1–2):*
+
+1. Validate the panel and target; compute group survival and baseline-normalized
+   average hazards.
+2. Fit the selected hazard relationship on pre-treatment moments; impute
+   counterfactual treated survival and post-period absorption ATT.
+3. Recompute the complete calculation in each individual bootstrap draw; form
+   pointwise and simultaneous effect inference.
+4. Separately compare pre-treatment hazard gaps/ratios with their last
+   pre-treatment value and use a centered simultaneous bootstrap test.
+
+**Reference implementation(s):** Authors' MATLAB `durationDiD` and Stata
+`durationdid` at the pinned commit. Static audit only; executability and exact
+numerical equivalence have not been established.
+
+**Requirements checklist:**
+
+- [ ] Balanced individual-panel validation and explicit exit-ATT estimand.
+- [ ] CD weighted gap and PH mean-ratio coefficient with baseline normalization.
+- [ ] Whole-individual bootstrap with complete nuisance re-estimation.
+- [ ] Coherent centered-bootstrap pointwise, simultaneous, and scalar inference.
+- [ ] Separate fixed-anchor pre-treatment diagnostics and unavailable-test states.
+- [ ] Domain, invalid-curve, failed-draw, and joint-NaN inference handling.
+- [ ] `BaseEstimator`/results/serialization/event-study integration and documentation.
+- [ ] Regression scenarios and independent reference checks specified below.
+
+## Identification and estimator derivation
+
+### Population, time, and causal assumptions
+
+Section 1 (pp.5–7) follows individuals `i=1,...,n` at dates `t=1,...,T`.
+`Y_it=1` means the spell has ended by date `t`; zero means it continues. No new
+spells start after the initial observation, but some spells may already have
+ended before it. `G_i` is fixed; group 1 receives the intervention strictly after
+`tstar`, and group 2 supplies unaffected comparisons. Time can be calendar time
+or elapsed spell time. The same calendar event need not occur at the same elapsed
+duration for everyone; a common-timing analysis must justify its chosen clock.
+
+The potential outcome `Y_it^(0)` removes the group intervention. The target is
+
+```text
+tau[t] = E[Y_it - Y_it^(0) | G_i=1].                    (p.6)
+```
+
+It averages over the whole fixed treated group, including baseline-absorbed
+people, whose effects are zero under the assumptions. It is neither an ATT
+conditional on surviving to treatment nor a hazard ratio nor a mean-duration
+effect. If spells begin together, the counterfactual mean is a duration CDF on
+the observed grid. A survival-probability effect has the opposite sign. Dropping
+baseline-absorbed people changes the population and rescales the estimand.
+
+**Assumption 1** requires binary absorbing factual and counterfactual outcomes:
+once one, always one. **Assumption 2** requires factual/untreated equality for
+everyone through `tstar` and controls afterward. Within-group spillovers are
+permitted by the group-level causal model (p.7); control spillovers are excluded.
+Observed monotonicity is checkable, whereas no anticipation, control validity,
+and counterfactual hazard restrictions cannot be verified from these checks.
+
+Write `S_kt=1-E[Y_it|G=k]`, `R_kt=-log(S_kt)`, with superscript `(0)` for
+counterfactual quantities (Equations 2.6–2.7). The instantaneous hazard is the
+limit of exit probability over a short interval conditional on survival,
+divided by its length (p.9). With the continuous-duration regularity underlying
+the paper's integral identity,
+
+```text
+Delta_s R_kt = R_kt - R_k,t-s
+Delta R_kt^(0) = integral_(t-1)^t h_k^(0)(r) dr          (2.8)
+D_kt^(0) = R_kt^(0)-R_k1^(0) = integral_1^t h_k^(0)(r) dr
+H_kt^(0) = D_kt^(0)/(t-1).                             (E.1)
+```
+
+Absorption alone does not supply an absolutely continuous hazard representation
+for arbitrary mass points. Discrete observation of a continuous duration process
+is compatible with these identities; the review does not extend their proof to
+every discrete-duration hazard model. Remark 3 permits real observation dates;
+elapsed durations must replace `t-1` consistently. The initial survival level is
+unrestricted (p.10, footnote 4).
+
+### Two specifications and their identification
+
+Common dynamics (CD) says untreated hazard changes agree across groups
+(Equation 2.1), equivalently
+
+```text
+h_1^(0)(t) = h_2^(0)(t) + c.                           (2.3)
+```
+
+Proportional hazards (PH) says proportional evolution agrees (Equation 2.2),
+equivalently
+
+```text
+h_1^(0)(t) = c h_2^(0)(t), c>0,                       (2.4)
+```
+
+with strictly positive hazards as assumed in the paper. Integrating gives
+constant differences of average hazards for CD (2.11), and constant ratios of
+long increments for PH (2.12). Equations 2.9–2.10 express related restrictions
+on successive integrated hazards. They concern untreated paths; observed treated
+post-period hazards need not obey them.
+
+**Theorem 1** (p.13) identifies CD `c=H_1t-H_2t` at every `1<t<=tstar`
+(2.13). For PH, `D_1t=c D_2t` (2.15), and at least one nonzero control
+pre-increment is required. The respective counterfactuals are
+
+```text
+CD: R_1t^(0) = R_11 + D_2t + (t-1)c                   (2.14)
+PH: R_1t^(0) = R_11 + c D_2t                         (2.16)
+E[Y_it^(0)|G=1] = 1-exp(-R_1t^(0)).                   (2.17)
+```
+
+Independent rearrangement makes the normalization and estimand transparent:
+
+```text
+CD: S_1t^(0) = S_11 (S_2t/S_21) exp(-(t-1)c)
+PH: S_1t^(0) = S_11 (S_2t/S_21)^c
+tau[t] = S_1t^(0)-S_1t.
+```
+
+The treated baseline factor is outside the PH exponent. Neither equal initial
+survival nor a zero initial event fraction is required. Two pre-dates identify
+one nuisance coefficient; a third supplies overidentification (Remark 1).
+Only an extra known equal-hazards restriction, CD `c=0` or PH `c=1`, permits
+identification with just one pre-date. That restricted special case is not the
+proposed unrestricted estimator.
+
+Remark 2 offers one-period or final-window CD identification moments. These
+agree in population under the restriction, but averaging them gives different
+finite-sample estimators. Remark 4 rewrites CD as
+`R_kt^(0)=xi_k-c_k t+gamma_t` with `c_1=0` (2.18), PH as
+`R_kt^(0)=xi_k+c_k gamma_t` with `c_1=1` (2.19), and equal hazards as additive
+group/time effects in `R` (2.20). These group-specific normalizations differ
+from the two-group coefficient `c` above.
+
+### Sample estimators and the PH ambiguity
+
+Equation 3.1 replaces each group expectation by its sample mean with fixed
+denominator `n_k`. Equations 3.2–3.4 give the CD estimator in the prospective
+registry entry. The text calls time weights positive and normalized; the
+application explicitly uses zero early weights. Thus nonnegative weights over
+a declared selected set reconcile the implementation descriptions. Suggested
+alternatives emphasize periods near treatment or minimize asymptotic variance;
+the paper supplies no universal optimal-weight procedure.
+
+Equation 3.5 prints, under the natural squared-increment reading,
+
+```text
+c_hat_printed = sum_t alpha[t] D_hat[1,t]D_hat[2,t]
+                / sum_t alpha[t] D_hat[1,t]^2
+R0_hat[1,t] = R_11 + c_hat_printed D_hat[2,t].           (3.5, printed)
+```
+
+The exponent placement also leaves squared increment versus increment of a
+squared `R` typographically ambiguous. Either way, the expression cannot be
+used uncritically. Under exact `D_1t=c D_2t`, the displayed slope reading gives
+`1/c`, while (2.16) requires `c`. The empirical baseline also needs a hat.
+Replacing the denominator by control squared increments repairs that slope's
+direction. It does **not** uniquely select a finite-sample PH estimator:
+
+| PH estimator | Coefficient | Evidence / meaning |
+|---|---|---|
+| Mean ratios, proposed core choice | `sum alpha[t] D1[t]/D2[t]` | Direct moments from Theorem 1; pinned MATLAB/Stata implement a mean of usable ratios |
+| Repaired cumulative-increment slope | `sum alpha[t] D1[t]D2[t] / sum alpha[t] D2[t]^2` | Directional repair of printed 3.5 |
+| Average-hazard slope | `sum alpha[t] H1[t]H2[t] / sum alpha[t] H2[t]^2` | PH specialization of 3.6 with zero intercept |
+
+All equal the true ratio under exact population PH and nondegenerate moments.
+They differ under sampling noise, weighting, and zero denominators. The second
+and third coincide only with suitable elapsed-time-squared reweighting, not
+generally with the same `alpha`. The recommended first implementation uses
+equal weights over original-sample eligible pre-dates with positive `D2` and
+finite logs. Freeze that set for bootstrap draws; a failed denominator then
+creates a failed draw, rather than silently changing the estimator in that draw.
+This eligibility/failure convention is library policy, not a claim of exact
+reference-software parity. Require a positive fitted PH ratio for its strict
+positive-hazard interpretation; a zero ratio is a boundary case. Report selected
+and excluded fitting dates with their reasons, including unavailable estimation
+when the eligible set is empty.
+
+### General affine hazards and constrained estimators
+
+Equation 2.5 allows
+`h_1^(0)=beta_1+sum_(k=2)^K beta_k h_k^(0)`, with known constraint set `B`.
+**Theorem 2** identifies coefficients from
+
+```text
+H_1t = beta_1 + sum_(k=2)^K beta_k H_kt, 1<t<=tstar,    (2.21)
+```
+
+provided the solution in `B` is unique. For unrestricted `K` coefficients,
+`tstar>K` is necessary and generically sufficient, not a guarantee against
+collinearity. Duration triple differences fixes weights `1,1,-1` on three
+controls and leaves an intercept; duration synthetic control fixes zero
+intercept and nonnegative control weights summing to one (p.10).
+
+The coherent weighted least-squares reading of Equations 3.6–3.7 is
+
+```text
+beta_hat = argmin_(beta in B) sum_(t=2)^tstar alpha[t]
+  (H_hat[1,t]-beta_1-sum_(k=2)^K beta_k H_hat[k,t])^2
+R0_hat[1,t] = R_hat[1,1]+(t-1)beta_hat[1]
+             +sum_(k=2)^K beta_hat[k] D_hat[k,t].
+```
+
+The printed objective starts at `t=1`, where `H=0/0`. Exclude the row before
+evaluation; multiplying it by zero does not define it. These extensions are
+outside the core scope. Binding constraints also require separate inference:
+ordinary bootstrap normal approximations need not remain valid at their
+boundaries (Section 3.3).
+
+## Bootstrap inference and pre-treatment diagnostics
+
+### Algorithm 1: whole-individual bootstrap
+
+Section 3.2 and Appendix B (pp.22–24,38–40) assume independent individuals with
+arbitrary serial dependence within each history. For a declared post-period
+family `P={tstar+1,...,T}`, the reconciled Algorithm 1 is:
+
+1. Compute `tau_hat[t]` for `t in P` on the original data.
+2. For each `b=1,...,B`, independently sample `n` indices uniformly with
+   replacement from all individuals. Carry their complete outcomes, groups,
+   and any covariates together. Repeated indices represent separate sampled
+   individuals, not duplicate records to discard.
+3. Recompute group means, log survivals, all coefficients, and effects.
+   In adjusted extensions, re-estimate weights or nonlinear first stages too.
+4. Calculate `sigma_hat[t]=SD_b(tau_star[b,t])` and centered absolute pivots
+   `Z_star[b,t]=abs(tau_star[b,t]-tau_hat[t])/sigma_hat[t]`.
+5. Pointwise criticals are `Q_(1-alpha){Z_star[b,t]}`; the simultaneous critical
+   is `Q_(1-alpha){max_(t in P) Z_star[b,t]}`.
+6. Form symmetric bands `tau_hat[t] +/- q sigma_hat[t]` using the appropriate
+   pointwise or simultaneous critical.
+
+There is one bootstrap SD per coordinate, not a nested resampling/studentizing
+loop. The intervals are centered-bootstrap absolute-deviation bands, not
+percentile intervals of the raw estimates. Original SD cancels from the
+pointwise width, but provides the relative scaling in the simultaneous maximum.
+The bootstrap is pooled pairs resampling, not a stratified, residual, wild,
+multiplier, or group-cluster bootstrap.
+
+Algorithm 1 inconsistently includes `tstar` in some loops/maxima but defines
+SEs and pointwise intervals only after it. Table 2 explicitly defines coverage
+over `tstar+1,...,T`, supporting the post-only family above. A reference-period
+ATT is causally zero and has undefined studentized inference; pre-fit residuals
+must not be mixed into this post-period family.
+
+**Proposed numerical conventions:** use `ddof=1`, an explicit inverse empirical
+CDF quantile, and `numpy.random.Generator` with recorded seed. For a scalar null
+`theta=0`, use the empirical fraction of centered absolute draws at least as
+large as `abs(theta_hat)`, with equality counted conservatively. For a
+simultaneous null, compare bootstrap maxima with
+`max(abs(theta_hat[t]/sigma_hat[t]))`. These p-values are an inversion-style
+library companion to the paper's bands, not an additional paper theorem.
+Finite-bootstrap discreteness and ties should be reported consistently.
+
+Pooled draws can lose a group, exhaust survival, or eliminate a PH increment.
+The paper does not specify a failure/retry rule. The conservative first-version
+recommendation is fixed `B` without retries: record each failure and reason;
+if any draw needed for an inferential family is invalid, retain meaningful
+point estimates but mark that family's inference unavailable. Successful draws
+may be retained as diagnostics, without treating a filtered distribution as
+the original bootstrap. Use the same complete-draw mask across a simultaneous
+family; separate per-column filtering does not define its joint maximum.
+Do not silently stratify, redraw, clip, or substitute tiny positive SEs.
+
+### Algorithm 2: separate specification tests
+
+Appendix B's Algorithm 2 (p.40) compares each pre-treatment average-hazard gap
+or ratio with the last pre-treatment value:
+
+```text
+CD: delta_hat[t] = (H_hat[1,t]-H_hat[2,t])
+                  -(H_hat[1,tstar]-H_hat[2,tstar])
+PH: delta_hat[t] = D_hat[1,t]/D_hat[2,t]
+                  -D_hat[1,tstar]/D_hat[2,tstar].
+```
+
+Compute them for `2<=t<=tstar`, but studentize/test only
+`J={2,...,tstar-1}`: the final anchor is identically zero and baseline `H` is
+undefined. For each pooled individual draw, recompute these contrasts with the
+same normalization. Calculate
+
+```text
+sigma_delta[t] = SD_b(delta_star[b,t])
+M_star[b] = max_(t in J) (abs(delta_star[b,t]-delta_hat[t])/sigma_delta[t])
+q = Q_(1-alpha){M_star[b]}
+diagnostic_band[t] = delta_hat[t] +/- q sigma_delta[t].
+```
+
+Reject if any band excludes zero. An equivalent proposed test statistic is
+`M=max_(t in J) abs(delta_hat[t]/sigma_delta[t])`, with the empirical tail
+fraction of `M_star` as p-value. Algorithm 2 incorrectly prints `tau_hat` at
+the band center and calls the maximum-based critical pointwise; its preceding
+steps and rejection target require `delta_hat` and a simultaneous critical.
+
+At least three pre-dates are required for nonempty `J`. PH diagnostic ratios
+need positive control increments at every tested date and the anchor, a
+stronger requirement than identification from one informative moment. The
+recommended behavior is to report the full diagnostic unavailable if those
+required ratios or SEs fail, while retaining a separately valid point estimate.
+Diagnostic curves have their own units and are not pre-treatment outcome ATTs.
+Failure to reject cannot establish post-treatment identification or rule out
+low power.
+
+**Main text versus appendix:** Section 3.2, p.23 instead prints contrasts of
+the initial window with an equal-length final window ending at `tstar`:
+
+```text
+CD: (D_1t-D_2t)/(t-1)
+    -[Delta_(t-1)R_1,tstar-Delta_(t-1)R_2,tstar]/(t-1)
+PH: D_1t/D_2t
+    -Delta_(t-1)R_1,tstar/Delta_(t-1)R_2,tstar.
+```
+
+The lag operator on p.12 means the final window starts at `tstar-(t-1)`.
+These are also valid population null contrasts, supported by Remark 2; they
+are not arithmetic errors. They are distinct finite-sample diagnostics from
+Algorithm 2's fixed-baseline anchor. The proposed first implementation chooses
+Algorithm 2, matching the application description. Never replace a lagged
+final-window numerator with a baseline-to-`tstar` difference while retaining
+the shorter elapsed-time denominator.
+
+### Scope of statistical justification
+
+Section 3.3 invokes GMM/sequential-GMM arguments with `n` increasing and `K,T`
+fixed, unique identification, and standard regularity. It explicitly omits a
+formal asymptotic theorem and proof. The appendix theorems prove identification,
+not bootstrap validity, efficiency, or finite-sample unbiasedness. Binding
+inequality constraints can break differentiability and bootstrap coverage;
+covariate weights must be bounded, requiring overlap bounded away from zero.
+Near-zero survival, weak PH denominators, dependence across people, increasing
+numbers of tiny cohorts, and censoring require additional work. No general
+analytic variance, degrees-of-freedom correction, or universal draw count is
+provided by the paper.
+
+## Complete extension and proof coverage
+
+### Appendix A.1: why outcome-level parallel trends can mislead
+
+Equation 1.1 imposes a constant untreated event-probability gap. Appendix A.1
+(p.32), for a positive gap `c`, positive control hazard, and
+`E[Y_t^(0)|G=2]<1-c`, derives
+
+```text
+h_1^(0)(t)/h_2^(0)(t)
+  = (1-E[Y_t^(0)|G=2])/(1-E[Y_t^(0)|G=2]-c).
+```
+
+As control absorption approaches `1-c`, the hazard ratio diverges. Thus a
+constant nonzero CDF gap can impose increasingly different survivor hazards.
+The p.8 numerical example increases the required exit-probability ratio from
+two to three as outcome shares move from `.8/.6` to `.9/.7`. Figure 1.1 shows
+converging untreated outcome means with parallel time-average hazards; naive
+DiD can reverse the true effect's sign. This motivates hazard restrictions;
+it does not prove that every use of ordinary DiD with a duration outcome fails.
+
+### Covariate-conditional identification: Section 2.2 and Appendix A.2
+
+For time-invariant `X`, Equations 2.22–2.24 allow conditional additive gaps
+`c(x)`, ratios `c(x)`, or affine coefficients `beta(x)`. Appendix A.2 repeats
+the general restriction as A.1. **Theorem A.1** (pp.32–33) applies Theorem 2
+separately at each `x`:
+
+```text
+H_1t(x) = beta_1(x)+sum_(k=2)^K beta_k(x) H_kt(x)       (A.2)
+R_1t^(0)(x) = R_11(x)+(t-1)beta_1(x)
+              +sum_(k=2)^K beta_k(x)D_kt(x).
+```
+
+Unique solution in the known set `B` is required per stratum; the theorem
+suppresses `x` on some coefficient symbols but does not establish constancy
+across strata. Transform to `1-exp(-R_1t^(0)(x))`, then integrate over the
+treated covariate distribution to get the unconditional treated mean.
+The source omits a separate proof because Theorem 2 applies to the stratum.
+Nonparametric conditional-mean estimation is suggested but can be impractical
+with high-dimensional covariates or sparse cells. No smoothing/selection
+algorithm or special high-dimensional inference guarantee is specified.
+
+### Initial-survivor balancing: Sections 2.2.1 and 3.1
+
+This is a distinct route, balancing covariates among **baseline survivors**:
+
+```text
+omega_k(x) = P(X=x|Y_1=0,G=1)/P(X=x|Y_1=0,G=k)        (2.25)
+p_k(x) = P(G=k|Y_1=0,X=x)
+omega_k(x) = p_1(x)P(G=k|Y_1=0)
+             /[p_k(x)P(G=1|Y_1=0)].                  (2.26)
+```
+
+Use densities for continuous covariates. The treated weight is one. Overlap
+requires positive comparison propensity on relevant treated support; bounded
+weights matter for inference. Balancing the full population or current
+post-treatment risk sets would define a different procedure.
+
+The weighted hazard (2.27) is the derivative of weighted cumulative event
+mass divided by weighted survivor mass. Define
+
+```text
+R_tilde[k,t] = -log E[omega_k(X)(1-Y_it)|G=k].
+```
+
+The identifying restrictions are directly on weighted untreated hazards:
+`h_tilde_1^(0)=h_tilde_2^(0)+c` or `c h_tilde_2^(0)` (2.28), and their affine
+generalization (2.29). **Theorem 3** identifies coefficients from (2.30),
+`H_1t=beta_1+sum beta_k H_tilde_kt`, and imputes
+`R_1t^(0)=R_11+(t-1)beta_1+sum beta_k D_tilde_kt`. The resulting mean
+conditions on treated group, having averaged over `X`; it is not a
+stratum-specific mean despite the source's occasional terminology.
+
+**Proposition 1** supplies a sufficient primitive condition for weighted CD:
+conditional CD with a gap `c(x)=c` constant in `x`. Equal conditional hazards
+and certain additive-separable hazard models are examples. It does not show
+that arbitrary conditional CD or PH collapses to the same marginal weighted
+restriction. Selection into survivor risk sets matters.
+
+For discrete covariates, let `N^S_k(x)` count group-k baseline survivors in
+stratum `x`, and `N^S_k=n_k(1-Ybar_k1)`. Equation 3.8 is
+
+```text
+omega_hat_k(x) = N^S_k N^S_1(x)/(N^S_1 N^S_k(x)).
+```
+
+For continuous/many-valued `X`, estimate `p_k(x)` among baseline survivors,
+for example by logistic regression, and use
+`omega_hat_k(x)=N^S_k p_hat_1(x)/(N^S_1 p_hat_k(x))`. Estimate survivor mass
+as `sum_(i:G=k) omega_hat_k(X_i)(1-Y_it)/n_k`, then substitute its negative
+log in the estimator. In general `E[omega(1-Y)]` is not `1-E[omega Y]`;
+normalization cannot be dropped casually. A time-constant multiplicative
+scaling cancels in long differences, but does not justify inconsistent
+baseline treatment or a different estimand. Re-estimate weights in every
+bootstrap draw. Missing support is an identification problem; trimming
+changes the population and needs an explicit target.
+
+### Appendix A.3: staggered adoption
+
+Pages 33–35 redefine `G=k` as treatment after date `k`; `G=T` includes
+never-treated units within the observation window. **Assumption 2*** gives
+untreated equality through each group's last pre-date. Equations A.3–A.4
+impose common untreated hazard changes or proportional evolution across all
+cohorts. For `1<s<=t`, their average hazards obey
+
+```text
+CD: H_kt^(0)-H_ks^(0)=c_ts
+PH: H_kt^(0)/H_ks^(0)=c_ts.
+```
+
+Not-yet-treated cohorts `k>=t` make these observable (A.5–A.6); at least one
+such cohort must exist, `P(G>=t)>0`. For treated target `k`, select
+`1<s<=k<t`, naturally `s=k`, and impute
+
+```text
+CD: R_kt^(0)=R_k1+(t-1)c_ts+(t-1)H_ks
+PH: R_kt^(0)=R_k1+(t-1)c_ts H_ks.
+```
+
+The PH population line on p.34 prints `R_kt` instead of `R_k1`; A.6 and
+the same page's sample formula require the baseline correction shown above.
+Estimate each `c_ts` by a normalized weighted average of eligible cohorts'
+differences or ratios, then obtain cohort-time absorption ATT. The paper does
+not specify an overall aggregation or joint inference across cohort-time cells.
+
+Already-treated cohorts are ineligible controls. No comparison remains once
+everyone is treated; `k=1` lacks a usable target pre-increment; PH needs
+positive denominators. Nonlinear log transformations create small-cohort bias.
+The authors suggest downweighting tiny cohorts and different methods if all
+cohorts are small, leaving bias adjustment for future work. This is not a
+justification for large-`K` asymptotics with tiny cohorts or for using existing
+staggered-DiD aggregators without a new derivation.
+
+### Appendix A.4: semiparametric covariate adjustment
+
+Pages 35–38 impose a known positive link, for example `phi(z)=exp(z)`, and
+group-specific coefficients:
+
+```text
+h_k^(0)(t,x) = phi(x'gamma_k) hbar_k^(0)(t)             (A.8)
+hbar_1^(0)(t) = beta_1+sum_(k=2)^K beta_k hbar_k^(0)(t). (A.9)
+```
+
+The unrestricted baseline hazard `hbar` is not the unconditional hazard.
+Together these imply the conditional affine relationship (A.7), with
+intercept `phi(x'gamma_1)beta_1` and coefficients
+`beta_k phi(x'gamma_1)/phi(x'gamma_k)`. For two groups, equal `gamma` and
+`beta_2=1` give conditional CD; zero intercept gives conditional PH.
+
+Let `r_kt=integral_1^t hbar_k^(0)(s) ds`. Then
+
+```text
+E[Y_it^(0)|Y_i1^(0)=0,G=k,X=x]
+  = 1-exp(-phi(x'gamma_k)r_kt).                       (A.10)
+```
+
+Equation A.11 makes this observable in untreated cells: all control periods
+and all groups' pre-periods. **Theorem 4** assumes a strictly positive,
+strictly increasing link, full covariate support, the baseline-hazard relation,
+and unique coefficients in `B`. Its statement/proof differ in risk-set support
+and use of “and” versus “and/or”; the operative proof uses baseline survivors
+and the union of untreated cells. An informative positive increment and
+covariate normalization are also essential to recover `gamma`.
+
+The baseline moment system and its implication are
+
+```text
+r_1t/(t-1) = beta_1+sum_(k=2)^K beta_k r_kt/(t-1)      (A.12)
+r_1t = (t-1)beta_1+sum_(k=2)^K beta_k r_kt.
+```
+
+The latter corrects the printed `(1-t)beta_1`. With group conditioning
+restored in A.13, the counterfactual mean is
+
+```text
+E[Y_it^(0)|G=1] = E[Y_i1|G=1]
+ + E[(1-Y_i1){1-exp(-phi(X_i'gamma_1)r_1t)}|G=1].      (A.13)
+```
+
+Appendix A.4.1 proposes nonlinear least squares over untreated cells `U`:
+
+```text
+min_(gamma,r) sum_((k,t) in U) sum_i
+  1{G_i=k}(1-Y_i1)[Y_it-1+exp(-phi(X_i'gamma_k)r_kt)]^2.
+```
+
+It mentions Cox partial likelihood, but prefers NLS for ties from discrete
+observation. A common `gamma` across groups is optional. Fit baseline-hazard
+relations using estimated `r`, then form the sample analogue of A.13. The
+corrected empirical ATT is
+
+```text
+Ybar_1t-Ybar_11
+ -(1/n_1)sum_i 1{G_i=1}(1-Y_i1)
+                  [1-exp(-phi(X_i'gamma_hat_1)r0_hat_1t)].
+```
+
+The p.38 formula prints current-survivor weight `(1-Y_it)` and an incorrect
+summation index. Baseline weighting is forced by A.13 and the proof: current
+survival is treatment affected and changes the estimand. No division by the
+current number at risk repairs it.
+
+The source does not settle optimizer, starts, tolerances, monotonicity or
+nonnegative constraints on fitted increments, or existence of a finite-sample
+solution. Covariates should exclude an intercept under the stated baseline
+normalization (footnote 10); zero baseline increments cannot identify `gamma`
+through the proof's inverse link. A future extension needs these additional
+conditions, complete nuisance re-estimation, and separate boundary-inference
+analysis. It is not an implementation shortcut for the unadjusted core.
+
+### Censoring, repeated cross-sections, and all Appendix E proofs
+
+Remark 3 (p.14) notes that identification uses only group means, so repeated
+cross-sections can suffice if they represent the same populations over time.
+That does not supply panel-style resampling for independently changing samples.
+Section 2.3 suggests substituting Kaplan–Meier group survival under independent
+right-censoring (`C_it` independent of `Y_it` conditional on group), and other
+survival estimators for dependent censoring. It provides no complete
+censoring-adjusted algorithm or bootstrap proof. Missing panel rows cannot be
+silently encoded as continued survival. Administrative end of a complete common
+observation window, with remaining survivors, is different from dropout.
+
+Appendix E (pp.46–50) gives the following implementation-relevant arguments:
+
+- **Theorems 1–2:** integrate affine hazards from the initial date (E.1–E.2),
+  replace pre-period counterfactuals using no anticipation (E.3), identify the
+  unique coefficient vector, and use unaffected controls after treatment.
+  Theorem 1 is the restricted two-group case. The treated initial `R_11`
+  remains essential; the proof's stray post-period treated equality is not an
+  additional no-effect assumption.
+- **Proposition 1:** absorption gives
+  `(1-Y_t^(0))(1-Y_1^(0))=1-Y_t^(0)`. Initial-survivor density ratios transfer
+  covariate expectations; a constant conditional additive gap factors out of
+  survival integrals and yields the weighted hazard gap. Positive baseline
+  survival is required. The printed `exp(tc)` should be `exp((t-1)c)`; its
+  log derivative and thus the final proposition remain unchanged.
+- **Theorem 3:** the weighted hazard is minus the derivative of log weighted
+  survivor mass. Integrate its assumed affine relation, use `omega_1=1`, and
+  apply the same identification logic. This proves the weighted model result,
+  not automatic collapsibility of conditional hazards.
+- **Theorem 4:** E.4–E.6 connect conditional survival among baseline survivors
+  with `phi(x'gamma_k)r_kt`. Evaluate at `x=0` to recover
+  `r_kt=-log(E[1-Y_it|G=k,X=0,Y_i1=0])/phi(0)`, then invert the link and use
+  support to identify `gamma`. The printed denominator `-phi(0)` has the
+  wrong sign. Integrating A.9 and averaging over group-1 baseline survivors
+  gives the corrected A.12–A.13 formulas above. No positive increment means
+  the inverse-link division cannot identify `gamma`.
+
+## Reported simulations and empirical evidence
+
+The results in this section are transcribed or summarized from the paper.
+No Monte Carlo experiment or author application was executed for this review.
+
+### Appendix C: design and complete numerical results
+
+Appendix C (pp.40–44) uses the untreated hazard shape
+
+```text
+f(t) = 1+sqrt(t/T)-0.5(t/T-0.5)^2
+h_2^(0)(t)=f(t)/(T-1)
+h_1^(0)(t)=[f(t)+c]/(T-1)
+h_1(t)=[f(t)+c+beta 1{t>=tstar}]/(T-1).
+```
+
+Transitions over `(t,t+1]` have probability
+`1-exp(-integral_t^(t+1) h_k(s) ds)`, evaluated numerically by the authors.
+The actual additive untreated hazard gap is `c/(T-1)`. The instantaneous
+change at `tstar` affects subsequent observed transitions, so `tstar+1`
+is the first affected discrete outcome. All Table 1 parameters are:
+
+| T | tstar | Initial treated event share | Initial control event share | c | beta |
+|---:|---:|---:|---:|---:|---:|
+| 20 | 11 | 0.4 | 0.2 | 0.5 | 1 |
+
+The paper uses equal time weights, 1,000 datasets, 1,000 bootstrap draws,
+and 95% inference. It labels sample size `n`; its appendix does not specify
+allocation, while the pinned script uses the listed count **per group**.
+Keep that distinction when designing a future replication. All Table 2 values
+are reproduced below:
+
+| Method | n | Absolute bias | Mean squared error | Uniform coverage | Average pointwise coverage | Pre-test rejection |
+|---|---:|---:|---:|---:|---:|---:|
+| Duration DiD | 100 | 0.00333 | 0.00176 | 0.962 | 0.957 | 0.058 |
+| Duration DiD | 500 | 0.00111 | 0.00034 | 0.949 | 0.951 | 0.041 |
+| Duration DiD | 1000 | 0.00024 | 0.00017 | 0.955 | 0.953 | 0.051 |
+| Duration DiD | 5000 | 0.00008 | 0.00003 | 0.945 | 0.946 | 0.058 |
+| Duration DiD | 10000 | 0.00010 | 0.00002 | 0.960 | 0.956 | 0.052 |
+| Standard DiD | 100 | 0.070 | 0.008 | 0.680 | 0.739 | 0.038 |
+| Standard DiD | 500 | 0.067 | 0.006 | 0.077 | 0.266 | 0.079 |
+| Standard DiD | 1000 | 0.068 | 0.005 | 0.004 | 0.083 | 0.166 |
+| Standard DiD | 5000 | 0.068 | 0.005 | 0.000 | 0.000 | 0.576 |
+| Standard DiD | 10000 | 0.068 | 0.005 | 0.000 | 0.000 | 0.804 |
+
+Uniform coverage means all post-period effects are covered within a simulated
+dataset; pointwise coverage is averaged over post-dates. Duration DiD has
+small bias, roughly nominal coverage and diagnostic size in its correctly
+specified CD design. Standard DiD's persistent bias produces collapsing
+coverage; its pretest can have little power at smaller sample sizes. These
+results do not validate PH, covariate, staggered, censoring, clustered,
+constrained, or boundary variants.
+
+Figure C.1 repeats Figure 1.1's converging untreated CDFs and parallel average
+hazards. C.2 shows estimated versus true Duration DiD effects with uniform
+bands for example datasets at `n=500,5000,10000`; C.3 shows biased,
+sign-reversed standard DiD estimates and narrowing wrong intervals. C.4
+shows average hazards and nonrejecting CD diagnostic bands for those examples.
+They are illustrations, not extra simulation estimates. The prose calls them
+A.1–A.4, while the actual captions are C.1–C.4.
+
+### Section 4 and Appendix D: Austrian unemployment insurance
+
+The August 1, 1989 Austrian reform extended potential benefit duration from
+30 weeks/210 days to 39 weeks/273 days for eligible unemployed people aged
+40–49 with at least 312 employed weeks in the preceding ten years. Existing
+eligible spells were included. The application restricts to people eligible
+for this extension but not the simultaneous replacement-rate change.
+
+With spell start `s_i` relative to reform, treated starts satisfy
+`-210<s_i<=92`, controls `-575<s_i<=-273`. Figure 4.1 shows these 302-day
+windows offset by 365 days. Initial counts are 5,311 treated and 5,390 control.
+Analysis time is elapsed unemployment duration; comparison of calendar
+cohorts at that duration is not the Appendix A.3 adoption design. To permit
+one week of anticipation before day 210, the authors use `tstar=202`.
+Earlier anticipation could bias their estimates.
+
+The application balances spell-start day of year via Equation 3.8 and removes
+40 of 291 dates with fewer than two untreated people, leaving 4,511 treated
+and 5,364 controls. It uses equal weights on the last 50 pre-dates, 153–202,
+and zero earlier weights. This does not change the initial origin of cumulative
+average hazards. Adding 50 earlier fitting dates reportedly changes little.
+The proposed unadjusted core does not automatically inherit this adjusted
+application's identifying argument or its trimmed target population.
+
+Figure 4.2 displays cumulative reemployment and weighted time-average hazards;
+the control hazard increases sharply at day 210. Figure 4.3 shows the CD
+counterfactual following control dynamics with the fitted shift, and the
+imputed cumulative exit curve above factual treated exit. Its 95% pointwise
+and uniform bands concern **counterfactual means**, which are different from
+ATT bands because ATT also includes factual-treated uncertainty/covariance.
+The application uses 10,000 whole-history bootstrap draws.
+
+Figure 4.4a reports negative absorption effects after day 210, visually near
+`-0.03` around days 240–260 and closer to `-0.01` by day 360; these are
+approximate visual readings, not numerical table estimates. Effects attenuate
+after treated expiry at day 273. The main-text conclusion on p.28 says both
+bands include zero on days 203–209, without significant evidence of one-week
+anticipation. Figure 4.4b's **60%** uniform diagnostic band contains zero and
+the reported p-value is `0.643`. Coverage 60% is not a significance level of
+60%, and nonrejection is not proof of identification. The p.27 sentence
+mislabels the solid expiry line: its day 273 corresponds to treated expiry;
+controls expire at day 210.
+
+Appendix D (p.45) shows the PH analysis with the same weighted hazards.
+Figure D.1 gives imputed hazards and counterfactual cumulative means; D.2(a)
+gives negative effects with 95% pointwise and uniform bands. The legend of
+D.2(b) specifies **60% uniform diagnostic intervals**, matching the main CD
+diagnostic; the band contains zero. The curves are close to CD because
+pre-treatment weighted hazards are close in this application. There is no new
+numerical effect table or universal
+specification selection rule; the two models need not agree in other settings.
+
+## Reference implementation audit
+
+The six pinned originals were inspected, not executed. No `stata`, `stata-mp`,
+`stata-se`, MATLAB, or Octave executable was found on PATH; the required
+`restud_data.csv` application data are absent. The cache has no LICENSE file,
+and no licensing grant was found in its six files. Public availability is not
+a copying permission. This review vendors no author code; an implementation
+should derive independently from the equations unless reuse rights are settled.
+
+| Evidence in pinned files | Consequence |
+|---|---|
+| MATLAB `durationDiD.m` begins again in concatenated `endfunction out = durationDiD(...)` at line 187; Stata `durationdid.ado` repeats `program define durationdid` at lines 6 and 616 and duplicates Mata definitions | Static structural defects prevent treating the snapshot as a certified executable oracle; no runtime error or repair was tested |
+| MATLAB lines 45/49 and 234, Stata lines 339–353 and 947 | CD averages hazard gaps; PH averages usable ratios, not either printed/repaired WLS slope |
+| Stata lines 339–346 sets baseline average hazards to zero and includes them with default `burnin=1`; MATLAB excludes the undefined baseline via `nanmean` | With two informative pre-increments and true gap `c`, Stata's default produces `2c/3`; proposed implementation excludes baseline before constructing moments |
+| MATLAB `durationDiD.m` lines 110–124 and 294–308 | SDs and centered absolute deviations with post-only maxima support Appendix B's inference structure |
+| Stata lines 495–498 replaces zero SD with `1e-100` | Do not copy this into valid inference; retain zero SE and joint unavailable inference |
+| MATLAB lines 132/316; Stata lines 528/1110 use `abs(max(delta/SD))` | Required two-sided statistic is `max(abs(delta/SD))`; `[-4,1]` gives 1 versus 4, and the incorrect statistic is sign asymmetric |
+| Reference diagnostics include baseline/reference coordinates; Stata quantiles use `ceil(level*B)` while MATLAB uses its quantile default | Exclude undefined/zero anchors; match time family, resamples, weights, and empirical-quantile convention before expecting numerical parity |
+| MATLAB core line 31 uses `periods>=absorbed_time`; Mata survival uses `absorbed_time>t` at `durationdid.ado` lines 305–306 | First absorption at `t` means `Y_t=1`; `UIP_application_revision_R2.m` line 61 instead uses a strict comparison for its raw plot |
+| README uses one row per individual and first absorption time, sentinel beyond horizon, and default horizon equal to maximum absorption time | A sentinel can accidentally become the horizon and force absorption; any future duration adapter must require an explicit observation horizon and distinguish censoring |
+| MATLAB lines 159–186 and Stata around 218–270 construct discrete covariate weights, drop small control cells, and rescale | Deferred adjustment changes target under trimming; MATLAB's `absorbed_time>=1` at lines 159–160 and Stata's `at :>= 1` at line 218 both include people absorbed exactly at baseline, unlike the required `>1` |
+| MATLAB first branch line 23 uses sampling weights but bootstrap line 76 resets to ones; Stata lines 457–467 resamples weights | No consistent survey-inference or sampling-weight parity claim is supported; global normalization can also make group survivor mass exceed one |
+| `UIP_application_revision_R2.m` line 86 requests 14 outputs, while cached functions return one struct; cohort construction and draw counts differ from `UIP_application.do` | Cached scripts do not specify one identical verified replication; MATLAB uses treatment cutoff 202 in selection and 1,000 draws, Stata uses initial benefit expiry 210 and 10,000 draws |
+| Both application scripts set `burn_in=152`; estimator indexing includes `152:202` | This fits 51 dates, unlike the paper's zero weights on the first 152 dates and equal weights on 153–202; window parity requires an explicit choice |
+| `montecarlo_revision.m` has 1,000 simulations, 1,000 bootstrap draws, counts per group, and a 100,000-point integration grid | Static design evidence only; its converted `t_star=floor(50000/5000)=10` differs from Table 1's 11; the standard-DiD bootstrap stratifies groups while duration bootstrap pools them, and comparator diagnostic anchoring differs between original and draws |
+
+README software defaults include CD, `burnin=1`, 1,000 draws, confidence
+coverage `.95`, diagnostic coverage `.6`, and seed `12345`. They are not
+paper mandates. The first estimator should use explicit library conventions
+instead of inheriting the baseline-row problem, 60% diagnostic coverage, or
+an implicit observation horizon. Original files remain unchanged, and no
+table/figure reproduction or cross-language numerical certification is claimed.
+
+## Implementation Notes
+
+### Data structure, behavior, and result contract
+
+The proposed initial input is a balanced long individual panel with column
+arguments `unit`, `time`, `treatment`, and `outcome`; `treatment` is a fixed
+0/1 group indicator, not a time-varying received-treatment variable. Require
+unique `(unit,time)` cells, both groups, finite binary absorbing outcomes,
+the same ordered equally spaced dates for every individual, and an explicitly
+identified last pre-date. Normalize the common time origin/scale and use actual
+elapsed durations consistently. Accept baseline absorption and survival through
+the common administrative horizon. Reject late entry, missing cells, recurrent
+spells, covariates, staggered adoption, survey weights, and arbitrary clustering
+in this first interface rather than silently interpreting them.
+
+Use `DurationDiD(BaseEstimator)` and `DurationDiDResults(BaseResults)` under
+current 3.x conventions; no interface is implemented in this PR. The time path
+is primary. A proposed headline scalar is the uniform average of explicitly
+reported post-period absorption ATTs, with those same weights applied within
+every bootstrap draw. This is an average of probability effects across dates,
+not a paper-defined unique overall ATT or a mean-duration effect.
+
+Expose specification, time origin, fitting dates/weights, group counts, fitted
+coefficient, observed/counterfactual survival, ATT path, bootstrap covariance,
+pointwise intervals, simultaneous bands, and separate diagnostic results/status.
+Use canonical `att`, `se`, `t_stat`, `p_value`, `conf_int` semantics for headline
+inference, plus `summary()`, `to_dict()`, and `to_dataframe()`. Store the actual
+confidence level and replicate/failure counts. When adapting to
+`EventStudyResults`, first treated observation has event time zero and the
+last untreated date is reference `-1`; reference inference is undefined.
+Hazard diagnostic contrasts must not masquerade as pre-treatment ATT rows.
+
+Evaluate logs only where needed. Both groups need positive survival at baseline
+and fitted pre-dates; control survival is needed at requested post-dates.
+Factual treated post-survival may equal zero: its ATT uses survival directly,
+so no treated post-log is required. Do not drop baseline-absorbed individuals.
+Near-zero survival should be flagged as weak numerical support, without
+inventing a paper-derived cutoff.
+
+CD can extrapolate negative/decreasing cumulative hazards or survival outside
+`[0,1]`, especially with a large negative gap and small control hazard
+(Section 2.3). Report the offending periods and raw extrapolation as invalid,
+with no valid causal inference for that family. Do not silently clip,
+monotonize, or change the horizon and still label it the same estimator.
+PH with a positive ratio preserves control monotonicity in population, but
+sample boundaries and failed resamples remain possible. A shorter supported
+horizon must be an explicit refit choice, not automatic outcome-dependent
+selection.
+
+### Relation to Existing diff-diff Estimators
+
+`BaseEstimator` already supplies constructor-based parameter introspection and
+transactional `set_params`; extend the existing estimator roster tests later.
+`results_base.py` supplies results/event-study conventions; serialization and
+adapter tests must cover the new type when it ships. Follow `_require_fit_alpha`
+where applicable so results cannot silently change fitted inference levels.
+
+Use `safe_inference()`/batch equivalents to compute the joint inference tuple
+and enforce its NaN gate, then coherently replace valid p-values and intervals
+with centered-bootstrap values. `apply_bootstrap_event_study_overrides` is an
+existing pattern for preserving that gate. Do not combine normal p-values
+with bootstrap intervals while claiming one inference procedure. Invalid or
+zero SE means all downstream t/p/interval/band fields are undefined, not tiny
+positive variance.
+
+Existing `stratified_bootstrap_indices` is not the paper's pooled bootstrap;
+wild/multiplier generators and influence-function chunking are not drop-in
+replacements. `compute_bootstrap_pvalue` uses uncentered sign tails and
+`compute_percentile_ci` raw percentile endpoints, so neither directly implements
+Appendix B. Failure-rate reporting can be reused. Whole-history index resampling
+with NumPy and a dedicated centered-bootstrap summary needs no new dependency.
+
+The core estimates group means and moments, with no least-squares solver or
+hazard likelihood. If a later affine extension solves regressions, use shared
+`linalg.solve_ols` and its rank policy. Ordinary `DifferenceInDifferences` on
+`Y`, or on individual `log(1-Y)`, is inappropriate; the latter is undefined
+after absorption. A group-log-survival regression omitting the baseline or
+CD time drift is also a different estimator. Existing synthetic, staggered,
+GLM, and sensitivity methods have different identifying restrictions; matching
+an event-study container does not establish HonestDiD compatibility.
+
+### Computational considerations
+
+For `n` people, `T` dates, and `B` draws, direct panel aggregation costs
+`O(nT)` per estimate and `O(BnT)` overall; storing the panel plus replicate
+effect paths costs `O(nT+BT)`. These are algorithmic estimates, not paper
+benchmarks or measured performance. Draws can run independently, but deterministic
+RNG streams, fixed eligible moments, and common simultaneous families must be
+preserved. No Rust work is necessary for this documentation foundation.
+
+### Tuning Parameters
+
+All defaults in this table are recommendations for subsequent estimator design,
+not parameters currently available in diff-diff.
+
+| Parameter / decision | Type | Proposed default | Basis / restriction |
+|---|---|---|---|
+| Specification | CD or PH | CD | Explicit choice; author software default, no universal paper preference |
+| Last pre-date | Observed date | Required | Intervention occurs afterward; never infer from absorption |
+| Fitting dates | Pre-date set | All eligible dates after baseline | Fixed original-sample set; PH requires positive control increments |
+| Time weights | Normalized vector | Equal on fitting set | 3.2 permits alternatives; no universal optimal rule |
+| PH coefficient | Moment rule | Mean ratios | Theorem 1 and author code; distinguish the two WLS alternatives |
+| Bootstrap draws | Integer | 1,000, at least 2 | Simulation/software precedent; application uses 10,000 |
+| Confidence significance `alpha` | Float in `(0,1)` | `.05` | 95% effect and diagnostic inference; source application diagnostic differs |
+| Seed | Integer or `None` | `None`, recorded when set | Library RNG policy; reproducibility examples must set it |
+| Bootstrap SD / quantile | Conventions | `ddof=1` / inverse empirical CDF | Explicit finite-sample policy; source SD divisor unspecified |
+| Diagnostic | Contrast family | Algorithm 2 fixed anchor | Interior pre-dates; unavailable if insufficient/invalid |
+| Draw failures | Policy | Fixed draws, no retries; suppress affected family inference | Conservative handling outside paper's regular interior argument |
+| Headline aggregate | Time weights | Uniform across reported post-dates | Proposed probability-effect average; same weights in every draw |
+
+### Independent checks and future acceptance scenarios
+
+The review executed an independent NumPy algebra script at
+`.workflow/validation/algebra_checks.py`; its output is preserved in
+`.workflow/validation/algebra-checks.txt`. It checked CD baseline/sign identities,
+the PH reciprocal error and three finite-sample choices, max-absolute testing,
+fixed-anchor normalization, whole-history resampling, and pointwise/uniform
+critical ordering. This is neither estimator testing nor author-code execution.
+The following self-contained core calculation preserves the consequential
+checks in the published document:
+
+```python
+import numpy as np
+
+t = np.arange(1, 6, dtype=float)
+elapsed = t - 1
+pre, post = np.array([1, 2]), np.array([3, 4])
+d2 = np.array([0, .10, .30, .45, .60])
+r1_base, r2_base = .50, .20  # unequal initial cumulative hazards (and hence survivals)
+r2 = r2_base + d2
+r1_cd = r1_base + d2 + .04 * elapsed
+c = np.mean((r1_cd[pre] - r1_base - d2[pre]) / elapsed[pre])
+r0 = r1_base + (r2 - r2_base) + c * elapsed
+assert np.allclose(r0, r1_cd)
+assert np.max(np.abs(np.exp(-r0) - np.exp(-r1_cd))) < 1e-14
+r1_factual = r1_cd.copy()
+r1_factual[post] -= .03 * (t[post] - 3)
+tau = np.exp(-r0[post]) - np.exp(-r1_factual[post])
+assert np.all(tau < 0)
+
+d1 = 2 * d2
+ratio = np.mean(d1[pre] / d2[pre])
+printed = np.dot(d1[pre], d2[pre]) / np.dot(d1[pre], d1[pre])
+assert ratio == 2 and printed == .5
+assert np.allclose(r1_base + ratio * (r2 - r2_base), r1_base + d1)
+x, y, dt = np.array([.1, .3]), np.array([.2, .9]), np.array([1, 2])
+choices = (np.mean(y / x), np.dot(x, y) / np.dot(x, x),
+           np.dot(x / dt, y / dt) / np.dot(x / dt, x / dt))
+assert np.allclose(choices, [2.5, 2.9, 2.6923076923076925])
+z = np.array([-4., 1.])
+assert np.max(np.abs(z)) == 4 and abs(np.max(z)) == 1
+print(f"CD coefficient={c:.6f}; PH ratio={ratio:.6f}; printed={printed:.6f}")
+print("CD post absorption effects:", np.round(tau, 8))
+print("PH mean/increment-slope/average-slope:", np.round(choices, 6))
+```
+
+```text
+CD coefficient=0.040000; PH ratio=2.000000; printed=0.500000
+CD post absorption effects: [-0.01044616 -0.01754019]
+PH mean/increment-slope/average-slope: [2.5      2.9      2.692308]
+```
+
+The later estimator PR should test these identities on admissible individual
+panels and add the following behavioral scenarios, with requirements still
+unchecked above:
+
+- No treatment effect under both models, unequal initial survival, non-unit
+  PH ratio, known exit-effect sign, and finite-sample PH choices that differ.
+- Exact timing and denominators, baseline absorption, complete survival to
+  horizon, zero treated post-survival, and invariance to consistent time rescaling.
+- Invalid binary/absorbing histories, duplicate or missing cells, changing group
+  labels, absent groups, insufficient pre-dates, PH support failure, and invalid
+  counterfactual curves with the documented outcomes.
+- Fixed resample indices preserving complete histories and duplicate draws;
+  every nuisance recomputed; seed reproducibility; omitted-group, zero-survival,
+  weak-denominator, and failed-family inference behavior.
+- Fixed bootstrap draws verifying SD/quantiles, pointwise versus simultaneous
+  widths, max-absolute sign symmetry, frozen family endpoints, joint-NaN
+  inference, and scalar aggregation using the same weights in every draw.
+- Fixed-anchor diagnostics distinct from the main-text moving-window contrast;
+  deterministic reference excluded; two pre-dates allow estimation but not a
+  nontrivial diagnostic; PH point-estimate support can exceed test support.
+- Parameter round trips, transactional invalid updates, results serialization,
+  confidence-level consistency, and event-study reference/diagnostic separation.
+
+## Gaps and Uncertainties
+
+The register below separates reproducible internal corrections from unresolved
+source choices. Corrections are review derivations supported by specified source
+identities, not author-confirmed errata. Literal copied formulas are not a safe
+implementation specification.
+
+| Location | Printed/source issue | Resolution and evidence | Confidence / consequence |
+|---|---|---|---|
+| 2.9–2.10, p.11 | Successive-increment domains include `t=2`, requiring unobserved `R_k0` | Use `t>=3` for successive one-period increments; long-difference identification starts at `t=2` | High; do not invent a pre-baseline observation |
+| 2.12, p.12 | Factual symbols used for full-horizon PH restriction | Interpret counterfactuals beyond pre-periods, by 2.4/2.15–2.16 | High; observed post-treatment equality is not required |
+| Theorem 2, p.16 | Labels the identified object an increment but displays the level including baseline | Retain `R_11` in the level, consistently with Theorem 1 and proof | High; avoid adding or removing the baseline twice |
+| 3.5, p.21 | Squared-treated-increment reading of denominator; missing baseline hat | That reading returns reciprocal under exact PH; use Theorem 1 direction, explicit mean-ratio core choice | High incompatibility; finite-sample choice is policy, not uniquely repaired by typography |
+| 3.5 versus 3.6, p.21 | Claimed special case uses cumulative versus average-hazard weights | Elapsed-time-squared reweighting needed; table/check above distinguishes all three choices | High; cannot assert identical finite-sample estimators |
+| 3.6, p.21 | Objective begins at baseline `t=1` | Start strictly after baseline, as 2.21 and 3.2 | High; avoids undefined `0/0` |
+| p.23 versus Algorithm 2, p.40 | Moving final window versus fixed-baseline anchor | Both valid null contrasts under lag definition/Remark 2; choose Algorithm 2 explicitly | High; documented convention difference, not a source arithmetic failure |
+| Algorithm 1, pp.38–39 | Some loops/maxima include `tstar`; SEs only afterward | Post-only family supported by steps 6–7 and Table 2 | High intended endpoint; exclude reference studentization |
+| Algorithm 2, p.40 | `tau_hat` band center; maximum labeled pointwise | Center on `delta_hat`, call simultaneous, by steps 1/4/7/9 | High; otherwise tests the wrong object |
+| A.3, p.34 | PH population baseline `R_kt` | `R_k1` by A.6 and same-page plug-in | High; avoids factual post-treatment contamination |
+| Theorem 4, p.37 | Baseline intercept `(1-t)beta_1` | `+(t-1)beta_1` by A.12 and integrated A.9 | High; sign-consequential deferred correction |
+| A.13, p.37 | Final expectation omits treated-group conditioning | Add `G=1`, as proof p.50 and sample group indicator | High; preserves target distribution |
+| A.4.1, p.38 | Current-survivor ATT weight and sum index | Use `1-Y_i1`, sum over `i`, by A.13 and proof | High; current risk-set selection changes estimand |
+| Theorem 2 proof, p.47 | Stray `R_1t^(0)=R_1t` after control argument | Equality at baseline only; subsequent imputation uses `R_11` | High; does not assume zero post-treatment effects |
+| Proposition 1 proof, p.48 | `exp(tc)` with origin 1 | `exp((t-1)c)`; derivative still yields `c` | High normalization correction; proposition unchanged |
+| Theorem 4 proof, p.50 | Denominator `-phi(0)` | Positive `phi(0)` by E.6 | High; printed version gives negative integrated hazard |
+| Theorem 4, pp.36/50 | Current versus baseline survivor support; untreated intersection versus union | Proof uses baseline survivors and union; retain mismatch for extension review | High algebraic interpretation; stronger implementation conditions must be stated |
+| Theorem 4 proof, p.50 | Inverse link divides by possibly zero baseline increment | Require informative positive increment to identify each fitted `gamma` | High nonidentification at zero; missing nondegeneracy detail |
+| Figures/prose, pp.13,27,41–43 | Figure 1.1 ATT text says solid red instead of treated blue; treated expiry mislabeled; A.x references for C.x figures | Use treated factual minus treated counterfactual, actual chronology, and captions | High; descriptive correction only |
+
+The core formulas, finite-sample PH recommendation, and diagnostic convention
+are sufficiently specified to support the next estimator planning session;
+there is no remaining unresolved core algebraic blocker under those explicit
+choices. Source uncertainty remains material for the deferred semiparametric
+extension, binding-constraint inference, censoring/repeated-cross-section
+resampling, and arbitrary dependence. Reference-code licensing, executability,
+application data, and numerical reproduction remain unverified limitations.
+The bibliography was covered, but cited outside papers were not independently
+re-reviewed: statements about Cox, Kaplan–Meier, GMM, or other authors above
+are limited to this paper's use of them and the stated mathematical deductions.
diff --git a/tests/test_naming_guard.py b/tests/test_naming_guard.py
index 5e9ed0d21..82444be4e 100644
--- a/tests/test_naming_guard.py
+++ b/tests/test_naming_guard.py
@@ -1080,6 +1080,10 @@ def _token_family_code_refs(tok):
     ("time", "docs/methodology/papers/goodman-bacon-2021-review.md"): (
         "canonical calendar column prose (bacon period remapping), not the M-030 overload"
     ),
+    ("time", "docs/methodology/papers/deaner-ku-2026-review.md"): (
+        "canonical calendar-period column vocabulary in the DurationDiD proposal "
+        "(including elapsed-duration periods), not the renamed 0/1 post indicator"
+    ),
     ("time", "docs/methodology/papers/wooldridge-2023-review.md"): (
         "canonical calendar column prose in the shipped-API description, not the M-030 overload"
     ),