Wave 2: PanelProfile outcome/dose shape extensions + autonomous-guide worked examples

[igerber]

Summary

• Extend PanelProfile with outcome_shape: Optional[OutcomeShape] (numeric outcomes) and treatment_dose: Optional[TreatmentDoseShape] (continuous treatments). New fields populate distributional facts that gate WooldridgeDiD (QMLE) and ContinuousDiD prerequisites pre-fit.
• Add three end-to-end worked examples to llms-autonomous.txt (binary-staggered with never-treated, continuous-dose with zero baseline, count-shaped outcome). New §5 between the existing §4 reasoning and §5 post-fit validation; existing §5–§8 renumbered to §6–§9.
• Top-level exports: OutcomeShape, TreatmentDoseShape from diff_diff. Both new dataclasses are frozen and JSON-serializable via PanelProfile.to_dict().

Methodology references (required if estimator / math changes)

• Method name(s): profile_panel is descriptive (not an estimator); the worked examples reference CallawaySantAnna, WooldridgeDiD, ContinuousDiD. No estimator math is changed in this PR.
• Paper / source link(s): no new methodology claims beyond what is already in docs/methodology/REGISTRY.md.
• Any intentional deviations from the source: None. The is_time_invariant field uses exact distinct-count on observed non-zero doses (matching the documented contract and ContinuousDiD.fit()'s exact equality check). The is_count_like and is_integer_valued fields are explicitly labeled heuristics in the docstring + guide.

Validation

• Tests added/updated: 17 new unit tests in tests/test_profile_panel.py covering each shape heuristic (count-like Poisson, binary-as-not-count-like, continuous normal, bounded unit, categorical returning None, skewness/kurtosis gating, JSON roundtrip, time-invariant dose, time-varying dose, no-zero-dose, binary/categorical treatment returning None, frozen invariants on both new dataclasses, sub-1e-8 dose-precision regression). 2 new content-stability tests in tests/test_guides.py guard the §5 worked-examples block and the new field references.
• Backtest / simulation / notebook evidence: N/A (descriptive utility + documentation).

Security / privacy

• Confirm no secrets/PII in this PR: Yes

[github-actions]

Overall assessment

⚠️ Needs changes — one unmitigated P1 methodology/contract mismatch in the new ContinuousDiD pre-fit gating surface.

Executive summary

• P1: The new TreatmentDoseShape fields are documented as the ContinuousDiD prerequisites, but they do not match what ContinuousDiD.fit() actually enforces. That can mark ineligible panels as eligible pre-fit.
• The mismatch is internal as well as methodological: the existing autonomous guide still says ContinuousDiD requires has_never_treated and treatment_varies_within_unit == False, while the new section shifts the gate to row-level has_zero_dose and nonzero-only is_time_invariant.
• P2: The new worked example uses WooldridgeDiD(family="poisson"), but the public API is WooldridgeDiD(method="poisson"). The new guide test hard-codes the wrong call.
• The new tests do not cover the two cases that would have exposed the P1: 0,0,d,d within-unit dose paths, and row-level zeros without any never-treated units.
• I did not find a mitigating entry for the P1 in TODO.md. This was a static source review; I could not execute the code in this environment.

Methodology

• Severity: P1
  Impact: ContinuousDiD is still implemented as requiring unit-level never-treated controls (first_treat == 0) and full per-unit dose constancy, but the new profile surface defines has_zero_dose as “any zero row” and is_time_invariant on nonzero doses only, then states those new fields are the prerequisites. That is an undocumented mismatch between the shipped pre-fit guidance and the estimator contract. A panel with pre-period zeros and no never-treated units, or a 0,0,d,d dose path, can satisfy the new guide’s gate while ContinuousDiD.fit() still raises.
  Concrete fix: Either align _compute_treatment_dose() to the actual estimator contract, or keep the fields descriptive and stop claiming they alone gate ContinuousDiD. In either case, keep has_never_treated and full-path time-invariance as the authoritative prerequisites unless you also change ContinuousDiD.fit() and document that deviation in REGISTRY.md.
  Locations: diff_diff/profile.py:L656-L710, diff_diff/continuous_did.py:L222-L227, diff_diff/continuous_did.py:L348-L360, diff_diff/guides/llms-autonomous.txt:L115-L140, diff_diff/guides/llms-autonomous.txt:L198-L211, diff_diff/guides/llms-autonomous.txt:L690-L723, docs/methodology/REGISTRY.md:L719-L745

Code Quality

• No findings in the changed Python implementation beyond the methodology contract issue above.

Performance

• No findings.

Maintainability

• No separate findings.

Tech Debt

• No separate findings.
• The P1 above is not mitigated by TODO.md; I did not find an existing tracked entry for it.

Security

• No findings.

Documentation/Tests

• Severity: P2
  Impact: The new autonomous guide tells users/agents to call WooldridgeDiD(family="poisson"), but the estimator constructor accepts method, not family. Anyone following the example will get a runtime error, and the new guide test now locks that bad example in place.
  Concrete fix: Change the guide and the stability test to WooldridgeDiD(method="poisson").
  Locations: diff_diff/guides/llms-autonomous.txt:L767-L771, diff_diff/wooldridge.py:L273-L275, tests/test_guides.py:L62-L64

• Severity: P2
  Impact: The new tests only cover “constant dose in every row” and “changing nonzero dose,” so they miss the exact contradictory cases introduced by the new guidance: 0,0,d,d within-unit paths and zero rows without any never-treated units. That leaves the P1 above unguarded.
  Concrete fix: Add regression tests asserting consistency between profile_panel guidance and ContinuousDiD.fit() for:

  1. balanced panel with 0,0,d,d dose paths;
  2. panel with row-level zeros but no never-treated units.
     Locations: tests/test_profile_panel.py:L1028-L1105

Path to Approval

1. Make the new TreatmentDoseShape contract consistent with ContinuousDiD.fit() and the Methodology Registry, or explicitly scope it as descriptive-only and restore the existing prerequisites (has_never_treated, full-path time invariance) in the guide.
2. Add regression tests for the two contradictory continuous-dose scenarios: 0,0,d,d paths and row-level zeros without never-treated controls.
3. Fix the autonomous guide and its stability test to use WooldridgeDiD(method="poisson").

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: cc91a7dad85dbb36ec713b2deedeae7d0938e3bc

---

Overall Assessment
✅ Looks good — no unmitigated P0/P1 findings.

Executive Summary

• The prior re-review P1 is resolved: the new TreatmentDoseShape surface is now explicitly descriptive-only, and the guide/tests defer to the same ContinuousDiD gates the estimator and Methodology Registry enforce: has_never_treated, treatment_varies_within_unit, and is_balanced. diff_diff/profile.py:L64-L83, diff_diff/profile.py:L658-L697, diff_diff/continuous_did.py:L222-L360, docs/methodology/REGISTRY.md:L717-L745, diff_diff/guides/llms-autonomous.txt:L198-L220, diff_diff/guides/llms-autonomous.txt:L499-L507, diff_diff/guides/llms-autonomous.txt:L732-L757, tests/test_profile_panel.py:L1057-L1115
• The prior re-review P2 is resolved: the worked example now uses WooldridgeDiD(method="poisson"), matching the public API and the Wooldridge documentation. diff_diff/guides/llms-autonomous.txt:L794-L805, diff_diff/wooldridge.py:L271-L276, docs/methodology/REGISTRY.md:L1333-L1349, tests/test_guides.py:L50-L71
• I did not find estimator-math, weighting, variance/SE, identification-assumption, or default-behavior changes in the PR; the changed surface is descriptive profiling, exports, guide content, and tests. diff_diff/profile.py:L43-L200, diff_diff/__init__.py:L253-L259, diff_diff/guides/llms-autonomous.txt:L637-L811
• P3: ROADMAP.md still overstates the new fields as if they themselves gate Wooldridge/ContinuousDiD routing, which conflicts with the corrected guide/docstrings. ROADMAP.md:L141-L141
• P3: the new top-level exports lack a matching regression in the existing top-level import-surface test. diff_diff/__init__.py:L253-L259, diff_diff/__init__.py:L503-L508, tests/test_profile_panel.py:L404-L413
• Static source review only: I could not execute tests here because this environment is missing pytest and even numpy.

Methodology

• No findings. The previous ContinuousDiD contract mismatch is fixed and now matches both the estimator implementation and REGISTRY.md. diff_diff/profile.py:L64-L83, diff_diff/continuous_did.py:L222-L360, docs/methodology/REGISTRY.md:L717-L745
• No findings. The previous Wooldridge guide/API mismatch is fixed. diff_diff/guides/llms-autonomous.txt:L794-L805, diff_diff/wooldridge.py:L271-L276

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• P3 — ROADMAP.md still says the new extensions expose facts that “gate WooldridgeDiD QMLE / ContinuousDiD prerequisites pre-fit,” which reintroduces the exact conceptual confusion the guide/docstrings just corrected; reword it to say the new fields add descriptive outcome/dose-shape context, while the authoritative ContinuousDiD gates remain has_never_treated, treatment_varies_within_unit, and is_balanced, and outcome_shape.is_count_like only informs Poisson-QMLE consideration. ROADMAP.md:L141-L141, compare diff_diff/profile.py:L64-L83, diff_diff/guides/llms-autonomous.txt:L198-L220, diff_diff/guides/llms-autonomous.txt:L499-L507

Tech Debt

• No findings. I do not see a new correctness issue that should be deferred into TODO.md.

Security

• No findings.

Documentation/Tests

• P3 — The PR adds OutcomeShape and TreatmentDoseShape to the public top-level namespace, but test_top_level_import_surface() still only covers profile_panel, PanelProfile, and Alert; extend that test to assert diff_diff.OutcomeShape, diff_diff.TreatmentDoseShape, and their presence in diff_diff.__all__. diff_diff/__init__.py:L253-L259, diff_diff/__init__.py:L503-L508, tests/test_profile_panel.py:L404-L413
• Review note: I could not run the added tests in this environment because pytest is unavailable and importing the package runtime also fails for lack of numpy.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 370915efd4056efc0cfe882a96041818a5d62239

---

Overall assessment

✅ Looks good — no unmitigated P0/P1 findings. The prior substantive re-review issues I checked are resolved.

Executive summary

• The earlier ContinuousDiD contract concern is fixed: the new treatment_dose surface is now explicitly descriptive-only, and the guide points back to the same field-based gates enforced in ContinuousDiD.fit() and documented in the registry. diff_diff/profile.py:L64-L89, diff_diff/guides/llms-autonomous.txt:L198-L220, diff_diff/guides/llms-autonomous.txt:L499-L507, diff_diff/continuous_did.py:L222-L360, docs/methodology/REGISTRY.md:L717-L745
• The earlier Wooldridge worked-example/API mismatch is fixed: the guide now uses WooldridgeDiD(method="poisson"), matching the public API and the Wooldridge registry entry. diff_diff/guides/llms-autonomous.txt:L794-L805, diff_diff/wooldridge.py:L271-L275, docs/methodology/REGISTRY.md:L1333-L1349
• The earlier top-level export/test gap is fixed: OutcomeShape and TreatmentDoseShape are exported from diff_diff and covered by the import-surface regression. diff_diff/__init__.py:L253-L259, diff_diff/__init__.py:L503-L510, tests/test_profile_panel.py:L404-L422
• P3: the new “authoritative ContinuousDiD pre-fit gates” wording now sounds exhaustive but omits the separate duplicate-cell hard stop already documented elsewhere; _precompute_structures() still resolves duplicate (unit, time) rows by last-row-wins overwrite. diff_diff/guides/llms-autonomous.txt:L330-L343, diff_diff/guides/llms-autonomous.txt:L499-L507, diff_diff/continuous_did.py:L818-L823, diff_diff/profile.py:L69-L75, CHANGELOG.md:L11-L11, ROADMAP.md:L141-L141
• Static review only: I could not run tests here because this environment is missing pytest, numpy, and pandas.

Methodology

• No P0/P1 findings. The changed guide/docstring text now matches the current ContinuousDiD and WooldridgeDiD contracts and does not introduce an undocumented estimator-methodology deviation.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings beyond the documentation item below.

Tech Debt

• No findings. I did not see a related mitigation entry needed in TODO.md.

Security

• No findings.

Documentation/Tests

• P3 — Impact: the new ContinuousDiD prerequisite summary omits the already-documented duplicate-row hard gate, so a reader treating the new treatment_dose summary as exhaustive can still route duplicate-containing panels into a code path that silently overwrites earlier (unit, time) rows. Concrete fix: in the new TreatmentDoseShape docs and the related guide/ROADMAP/CHANGELOG wording, change “authoritative gates are ...” to “core field-based gates are ... plus no duplicate_unit_time_rows alert,” and add a regression in tests/test_guides.py that the ContinuousDiD prerequisite summary still mentions duplicates. diff_diff/guides/llms-autonomous.txt:L330-L343, diff_diff/guides/llms-autonomous.txt:L499-L507, diff_diff/continuous_did.py:L818-L823, diff_diff/profile.py:L69-L75, CHANGELOG.md:L11-L11, ROADMAP.md:L141-L141
• Verification note: I could not execute the new tests in this environment.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 22ff5dc161534d087d0aa191b74f04efd99eaa43

---

Overall Assessment

⚠️ Needs changes — I found two unmitigated P1 methodology issues in the new count-outcome guidance/routing.

Executive Summary

• The prior duplicate-row documentation issue appears resolved: the ContinuousDiD prerequisite summary now keeps duplicate_unit_time_rows in scope.
• P1: The new Wooldridge Poisson guidance says the estimator returns a multiplicative/proportional effect, but the implementation and reporting contract expose an ASF-based ATT difference on the outcome scale.
• P1: The new OutcomeShape.is_count_like heuristic does not exclude negative support, so it can steer negative integer outcomes toward WooldridgeDiD(method="poisson"), which hard-rejects negative outcomes.
• P2: The added tests pin section presence and API spelling, but they do not guard the nonlinear Wooldridge target-parameter wording or the negative-count edge case.
• Static review only: I could not run the test suite because pytest is not installed in this environment.

Methodology

• Severity: P1. Affected method: WooldridgeDiD(method="poisson"). Impact: The new guide tells agents that Poisson ETWFE “estimates the multiplicative effect directly” and that they should “report the multiplicative effect (proportional change),” but the library’s own methodology contract is ASF-based ATT, and the Poisson path computes E[exp(η_1)] - E[exp(η_0)], i.e. an outcome-scale difference, not a ratio. An agent following the new worked example can therefore misreport overall_att on the wrong scale. Concrete fix: Rewrite the §4.11 / §5.3 wording so Poisson ETWFE is described as an ASF-based ATT on the natural outcome scale; if you want to mention proportional effects, label them as a separate derived interpretation, not the estimator’s overall_att target. Refs: diff_diff/guides/llms-autonomous.txt:L621-L629, diff_diff/guides/llms-autonomous.txt:L801-L811, docs/methodology/REGISTRY.md:L1335-L1357, diff_diff/wooldridge.py:L1191-L1225, diff_diff/_reporting_helpers.py:L262-L281.
• Severity: P1. Affected surface: profile_panel outcome-shape routing for WooldridgeDiD(method="poisson"). Impact: is_count_like currently requires integer-valued data, zeros, positive skew, and >2 distinct values, but it does not require non-negative support. That means a right-skewed integer outcome with zeros and some negative values can still set is_count_like=True, after which the new guide treats it as a Poisson candidate even though the Poisson fitter raises on any negative outcome. This is a missing edge-case guard on the new routing signal. Concrete fix: Add a non-negativity requirement to is_count_like (value_min >= 0 or equivalent), document it, and add a regression test with negative integer outcomes plus zeros expecting is_count_like=False. Refs: diff_diff/profile.py:L621-L645, diff_diff/guides/llms-autonomous.txt:L621-L629, diff_diff/wooldridge.py:L1105-L1108.

Code Quality

No findings.

Performance

No findings.

Maintainability

No findings.

Tech Debt

No findings. I did not see these P1 items mitigated in TODO.md, and they are not deferrable under the review policy.

Security

No findings.

Documentation/Tests

• Severity: P2. Impact: The new tests verify the worked-examples section exists and that the Poisson example uses method= rather than family=, but they do not verify the nonlinear Wooldridge estimand wording or the negative-outcome exclusion. Both P1 issues above can therefore pass the current suite. Concrete fix: Add guide-content assertions that §4.11/§5.3 do not describe overall_att as “multiplicative” / “proportional change,” and add a profile_panel regression test for a right-skewed integer outcome with zeros and some negative values. Refs: tests/test_guides.py:L62-L83, tests/test_profile_panel.py:L887-L907.

Path to Approval

1. Fix the Poisson ETWFE wording in diff_diff/guides/llms-autonomous.txt so it matches the ASF-based ATT contract already documented in REGISTRY.md and _reporting_helpers.py.
2. Tighten OutcomeShape.is_count_like to require non-negative support before it can steer users toward Poisson.
3. Add regression tests for both the estimand wording and the negative-integer edge case.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: b444300244662048617cd21fc6357d5cae0b1e32

---

Overall Assessment

⚠️ Needs changes — the two prior Wooldridge-related P1s look resolved in the provided diff, but there is one P1 [Newly identified] in the new ContinuousDiD pre-fit routing/docs.

Executive Summary

• The prior Poisson issues appear fixed in the provided diff: WooldridgeDiD(method="poisson") is now described as ASF/outcome-scale, and outcome_shape.is_count_like now excludes negative outcomes.
• P1 [Newly identified]: the new treatment_dose / ContinuousDiD prerequisite summaries omit the estimator’s strictly-positive-treated-dose restriction. A panel can satisfy every documented gate and still hard-fail in ContinuousDiD.fit() on negative dose support.
• The added tests cover the two prior Poisson regressions, but there is still no regression around negative-dose ContinuousDiD routing.
• No findings in code quality, performance, security, or tracked-tech-debt handling.
• Static review only: pytest is not installed in this environment, so I could not run the suite.

Methodology

Re-review note: the previous Poisson estimand-wording and negative-count-routing findings appear resolved in the provided diff.

• Severity: P1 [Newly identified]. Impact: the PR now presents has_never_treated, treatment_varies_within_unit == False, is_balanced, and no duplicate_unit_time_rows alert as the authoritative pre-fit gates for ContinuousDiD, and §5.2 says that if those pass the estimator is “in scope.” That omits the implementation’s hard support restriction that treated doses must be strictly positive; ContinuousDiD.fit() raises on negative treated dose support, and the methodology note is written on D > 0 support. An autonomous agent can therefore route a balanced, never-treated, time-invariant panel with treatment_dose.dose_min < 0 into a deterministic fit-time error. Concrete fix: extend the new routing contract so ContinuousDiD is only in scope when nonzero doses are strictly positive as well; at minimum, document treatment_dose.dose_min > 0 as a hard gate, and preferably expose an explicit negative-dose alert/boolean. Refs: diff_diff/profile.py:L65-L85, diff_diff/guides/llms-autonomous.txt:L737-L751, diff_diff/continuous_did.py:L287-L294, docs/methodology/continuous-did.md:L180-L184.

Code Quality

No findings.

Performance

No findings.

Maintainability

No findings.

Tech Debt

No findings. I did not find the ContinuousDiD gate omission documented as a registry deviation or tracked in TODO.md, so it is unmitigated.

Security

No findings.

Documentation/Tests

• Severity: P2. Impact: the new tests now guard the prior Poisson regressions, but they still do not cover the expanded ContinuousDiD routing surface for negative-dose panels. The treatment-dose tests only cover zero-baseline/no-zero/binary/categorical cases, and the guide tests do not assert any negative-dose exclusion. Concrete fix: add a regression with a balanced, never-treated, time-invariant continuous panel containing a negative nonzero dose and assert the routing surface rejects ContinuousDiD (via explicit alert/boolean or guide wording keyed to dose_min < 0). Refs: tests/test_profile_panel.py:L1037-L1187, tests/test_guides.py:L62-L105.

Path to Approval

1. Update the new TreatmentDoseShape / autonomous-guide prerequisite summaries so ContinuousDiD is only “in scope” when nonzero doses are strictly positive in addition to being zero-supported, time-invariant, balanced, and duplicate-free.
2. Add a targeted regression for a negative-dose continuous panel that otherwise passes the existing documented gates, so this support restriction is covered going forward.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: ed57c7cffa81eec1685695f3ad11894729e41372

---

Overall Assessment

⚠️ Needs changes — the prior negative-dose / Wooldridge Poisson wording issues look resolved, but there is one unmitigated P1 in the new ContinuousDiD routing contract.

Executive Summary

• The earlier dose_min > 0 omission appears fixed: the guide now reflects ContinuousDiD’s negative-dose rejection, and the Wooldridge Poisson text now matches the ASF outcome-scale estimand.
• P1 [Newly identified]: the PR repeatedly calls a five-item PanelProfile checklist the full ContinuousDiD pre-fit gate set, but ContinuousDiD.fit() still has separate first_treat validity/consistency gates and still requires P(D=0) > 0 regardless of control_group.
• The new OutcomeShape / is_count_like heuristic is aligned with the Poisson fitter’s non-negativity requirement and does not look methodologically off.
• Static review only: pytest is not installed in this environment, so I could not execute the added tests.

Methodology

• Severity: P1 [Newly identified]. Impact: the new docs present has_never_treated, treatment_varies_within_unit == False, is_balanced, no duplicate_unit_time_rows, and treatment_dose.dose_min > 0 as the full ContinuousDiD gate set. That is not the estimator’s full contract. ContinuousDiD.fit() also validates the separate first_treat column, rejects NaN/negative values, drops first_treat > 0 units with dose == 0, zeroes first_treat == 0 rows with nonzero dose, and enforces P(D=0) > 0 even under control_group="not_yet_treated". An agent can therefore mark a panel as “in scope” from PanelProfile alone even though the actual first_treat input will deterministically warn/error or change the sample. Concrete fix: stop calling the five profile-derived facts the “full” gate set; describe them as the profile-side screening surface and add an explicit caveat that ContinuousDiD.fit() still requires a valid/consistent first_treat column. Refs: diff_diff/guides/llms-autonomous.txt:L203-L235, diff_diff/guides/llms-autonomous.txt:L514-L526, diff_diff/guides/llms-autonomous.txt:L758-L788, diff_diff/profile.py:L65-L99, diff_diff/continuous_did.py:L230-L360, docs/methodology/REGISTRY.md:L719-L745, docs/methodology/continuous-did.md:L182-L184, ROADMAP.md:L141, CHANGELOG.md:L12.

Code Quality

No findings.

Performance

No findings.

Maintainability

• Severity: P3. Impact: the new internal commentary is inconsistent. Public docs now say treatment_dose.dose_min > 0 is part of the profile-side ContinuousDiD screen, but _compute_treatment_dose() and test_treatment_dose_does_not_gate_continuous_did() still say treatment-dose fields “do not gate” ContinuousDiD. That inconsistency will make future regressions more likely. Concrete fix: align the helper/test docstrings with the public contract (“most fields descriptive; dose_min > 0 is the extra screen”). Refs: diff_diff/profile.py:L65-L99, diff_diff/profile.py:L689-L694, tests/test_profile_panel.py:L1110-L1128.

Tech Debt

No findings. The P1 above is not mitigated by TODO.md tracking or a registry deviation note.

Security

No findings.

Documentation/Tests

• Severity: P2. Impact: the new tests pin the negative-dose wording and ASF outcome-scale wording, but they do not protect the missing first_treat caveat that causes the P1 above. Concrete fix: add a guide regression asserting that the ContinuousDiD section describes the five PanelProfile facts as screening checks only and explicitly mentions separate first_treat validation/consistency. Refs: tests/test_guides.py:L37-L69, diff_diff/guides/llms-autonomous.txt:L758-L788, diff_diff/continuous_did.py:L230-L360.

Path to Approval

1. Replace every “full ContinuousDiD pre-fit gate set” claim in CHANGELOG.md, ROADMAP.md, diff_diff/profile.py, and diff_diff/guides/llms-autonomous.txt with “profile-side screening checks,” and document the separate first_treat contract from ContinuousDiD.fit().
2. In llms-autonomous.txt §5.2, rewrite the has_never_treated == False counterexample so it states that the current implementation requires P(D=0) > 0 under both control_group options, not just the default.
3. Add a regression in tests/test_guides.py that requires the ContinuousDiD section to mention separate first_treat validation/consistency; optionally pair it with a small tests/test_continuous_did.py scenario where the five profile-side checks pass but inconsistent first_treat still warns/errors.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 8601b611fa2c21684b91ecb58bf4b5df314557d6

---

Overall Assessment

Looks good — the prior P1 is resolved, and I did not find any new unmitigated P0/P1 issues in the changed code/docs.

Executive Summary

• Prior re-review P1 is resolved: the ContinuousDiD language now consistently frames the five PanelProfile facts as a profile-side pre-flight screen and separately documents first_treat validation/coercion/drop behavior (diff_diff/profile.py:L65-L115, diff_diff/guides/llms-autonomous.txt:L203-L250, diff_diff/guides/llms-autonomous.txt:L530-L548, diff_diff/guides/llms-autonomous.txt:L780-L812, CHANGELOG.md:L12, ROADMAP.md:L141).
• The new OutcomeShape.is_count_like routing is aligned with Wooldridge Poisson fit behavior by excluding negative-support outcomes, and the guide now correctly describes Poisson overall_att as an ASF-based outcome-scale difference (diff_diff/profile.py:L669-L683, docs/methodology/REGISTRY.md:L1333-L1348, diff_diff/wooldridge.py:L1102-L1109, diff_diff/wooldridge.py:L1191-L1245).
• No estimator math, weighting, variance/SE, identification assumptions, or defaults changed in the implementation; this PR is limited to descriptive profiling fields, exports, guide content, and tests.
• One minor documentation drift remains: the changelog summary omits the non-negativity clause from the is_count_like heuristic (CHANGELOG.md:L12).
• Added tests cover the new profile fields and the prior review regression surface (tests/test_guides.py:L37-L185, tests/test_profile_panel.py:L889-L1313), but I could not execute them here because pytest and runtime deps such as numpy are unavailable.

Methodology

No findings. The changed docs now match the ContinuousDiD contract in docs/methodology/REGISTRY.md:L717-L745 and diff_diff/continuous_did.py:L222-L327,L348-L360, and the Wooldridge Poisson wording matches docs/methodology/REGISTRY.md:L1333-L1348 and diff_diff/wooldridge.py:L1191-L1245.

Code Quality

No findings.

Performance

No findings.

Maintainability

No findings.

Tech Debt

No findings.

Security

No findings.

Documentation/Tests

• Severity: P3
  Impact: CHANGELOG.md:L12 describes is_count_like without the value_min >= 0 / non-negative-support gate. Readers relying on the release notes alone may expect negative-valued outcomes to route as count-like even though the implementation intentionally suppresses that to stay compatible with WooldridgeDiD(method="poisson").
  Concrete fix: Update the changelog bullet to include the non-negativity clause, matching diff_diff/profile.py:L669-L683, diff_diff/guides/llms-autonomous.txt:L188-L198, and tests/test_profile_panel.py:L930-L971.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: bc13abb50011b5a5495d6448d458cff7dfa157e4

---

Overall Assessment

✅ Looks good

Executive Summary

• Re-review status: the prior changelog drift is resolved; I did not find any new unmitigated P0/P1 issues in the changed files.
• Methodology check passed for the affected guidance surfaces: the new ContinuousDiD profile-side screening language matches the estimator contract on zero-dose controls, full-path dose constancy, balanced-panel requirement, duplicate-row caveat, negative-dose rejection, and separate first_treat validation (diff_diff/guides/llms-autonomous.txt:L203-L251, diff_diff/guides/llms-autonomous.txt:L507-L548, diff_diff/continuous_did.py:L222-L360, docs/methodology/REGISTRY.md:L717-L745).
• Methodology check passed for the Wooldridge Poisson path: OutcomeShape.is_count_like now gates on non-negative support, and the guide correctly describes overall_att as an ASF-based outcome-scale difference rather than a multiplicative ratio (diff_diff/profile.py:L631-L695, diff_diff/wooldridge.py:L1102-L1278, docs/methodology/REGISTRY.md:L1333-L1349).
• No estimator math, weighting, variance/SE, identification assumptions, or defaults changed; the implementation is confined to descriptive profile fields, public exports, guide content, and tests (diff_diff/profile.py:L43-L442, diff_diff/init.py:L253-L259, diff_diff/init.py:L503-L508).
• One minor docs inconsistency remains: §4.7 still opens with “Three eligibility prerequisites” for ContinuousDiD, then later in the same section correctly expands the profile-side screen to five checks by adding duplicate-row absence and treatment_dose.dose_min > 0 (diff_diff/guides/llms-autonomous.txt:L507-L548).
• I could not execute the added tests here because this environment lacks pytest and runtime dependencies such as pandas and numpy.

Methodology

No findings. Cross-checking the changed profile/guide language against docs/methodology/REGISTRY.md:L717-L745 and docs/methodology/REGISTRY.md:L1333-L1349 found no undocumented deviation in estimator behavior.

Code Quality

No findings.

Performance

No findings.

Maintainability

No findings.

Tech Debt

No findings.

Security

No findings.

Documentation/Tests

• Severity: P3
  Impact: §4.7 contains two different summaries of the ContinuousDiD pre-fit contract. The opening bullet says there are only three eligibility prerequisites, while the later paragraph correctly adds the duplicate-row hard stop and the negative-dose screen. A reader who stops at the first summary can miss two load-bearing gates and route a duplicate-containing or negative-dose panel into a failed or silent-overwrite fit path (diff_diff/guides/llms-autonomous.txt:L507-L548, diff_diff/continuous_did.py:L222-L338, diff_diff/continuous_did.py:L800-L823).
  Concrete fix: Update the opening ContinuousDiD bullet in §4.7 to enumerate all five profile-side checks, or relabel the first three as “core estimator prerequisites” and explicitly cross-reference the fuller five-check screen immediately below.

Static review only: the added coverage in tests/test_profile_panel.py:L887-L1313 and tests/test_guides.py:L37-L182 looks appropriate, but I could not run it in this environment.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: a8880b46a47e39fb3c89a60718cc9b80a7bc2994

---

Overall Assessment

✅ Looks good

Executive Summary

• Prior re-review issue is resolved: the autonomous guide now consistently names all five ContinuousDiD profile-side checks and the separate first_treat validation surface.
• No estimator implementation, weighting, variance/SE computation, identification assumption, or default behavior changed in this PR; the code changes are confined to descriptive PanelProfile metadata, exports, guide prose, and tests.
• One methodology-facing docs issue remains: the new count-outcome guidance justifies Poisson QMLE by saying linear DiD asymptotic SEs assume “normal-shaped errors,” which does not match the library’s documented cluster-robust OLS inference contract.
• One minor docs clarity issue remains: the new ContinuousDiD worked example is written like a zero-pre/post-dose adoption path even though the estimator, the profile gate, and the same section later reject 0,0,d,d dose paths.
• Static review only: I could not run the added tests here because this environment does not have pytest, numpy, or pandas.

Methodology

• Severity: P2. Impact: The new count-outcome routing text tells agents that linear DiD SEs are questionable because they “assume normal-shaped errors,” then uses that as the rationale for preferring WooldridgeDiD(method="poisson"). That does not match the library’s documented inference surface, which is cluster-robust sandwich OLS for linear ETWFE and QMLE sandwich for nonlinear ETWFE; the current wording can steer estimator choice for the wrong reason even though the Poisson path itself is valid. Location: diff_diff/guides/llms-autonomous.txt:L188-L197, diff_diff/guides/llms-autonomous.txt:L666-L680, diff_diff/guides/llms-autonomous.txt:L882-L907, docs/methodology/REGISTRY.md:L1343-L1349. Concrete fix: rewrite the rationale in support/functional-form terms (or efficiency/model-fit terms) rather than claiming robust linear DiD inference depends on normal-shaped errors.

Code Quality

No findings.

Performance

No findings.

Maintainability

No findings.

Tech Debt

No findings.

Security

No findings.

Documentation/Tests

• Severity: P3. Impact: The new ContinuousDiD worked example is titled “zero baseline” and describes “block-style adoption in year 3,” but the shown profile has treatment_varies_within_unit=False, and the same section later correctly says a 0,0,d,d path is out of scope. That mixed framing can confuse an agent about the estimator’s full-path dose-invariance contract. Location: diff_diff/guides/llms-autonomous.txt:L759-L781, diff_diff/guides/llms-autonomous.txt:L822-L828, diff_diff/continuous_did.py:L222-L228. Concrete fix: rename the example to something like “continuous-dose panel with zero-dose controls,” and state explicitly that positive-dose units keep the same dose in every observed period while first_treat carries the timing information separately.

No additional test-coverage findings in the diff. Testing note: I could not execute tests/test_profile_panel.py or tests/test_guides.py in this environment because the required Python packages are not installed.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 0d711f1c334cb3fb6cccad02ab6eecfbcacac0ad

---

Overall Assessment

✅ Looks good

Executive Summary

• Re-review scope: the prior continuous-dose example issue appears resolved. §5.2 now keeps positive-dose units time-invariant across the full observed path and carries timing separately in first_treat, which matches ContinuousDiD.fit()’s contract in diff_diff/guides/llms-autonomous.txt:L769-L858 and diff_diff/continuous_did.py:L222-L360.
• No estimator implementation, weighting, variance/SE computation, identification assumption, or default behavior changed in this PR. The code changes are limited to descriptive PanelProfile metadata, exports, guide prose, and targeted tests.
• The ContinuousDiD screening narrative is now materially better aligned with the implementation: it explicitly names the duplicate-row hard stop, the dose_min > 0 gate, and the separate first_treat validation surface in diff_diff/guides/llms-autonomous.txt:L203-L251, diff_diff/guides/llms-autonomous.txt:L539-L556, and diff_diff/continuous_did.py:L287-L360.
• One methodology-facing docs issue remains from the prior review: the new outcome_shape.is_count_like field reference and the §4.11 preamble still imply count outcomes make linear DiD asymptotic SEs suspect, which conflicts with the registry’s cluster-robust OLS inference contract.
• Static review only: python -m pytest tests/test_profile_panel.py tests/test_guides.py -q could not be run here because pytest is not installed.

Methodology

• Severity: P2. Impact: The previous count-outcome docs finding is only partially resolved. The detailed §4.11/§5.3 discussion now correctly reframes Poisson-vs-linear as a functional-form/support/efficiency choice, but the §2 outcome_shape.is_count_like field reference still says raw-outcome OLS has “questionable asymptotic SEs,” and the §4.11 preamble still says asymptotic SEs “may mislead.” That conflicts with the documented OLS cluster-robust sandwich inference surface and can still steer estimator choice for the wrong reason. Location: diff_diff/guides/llms-autonomous.txt:L188-L193, diff_diff/guides/llms-autonomous.txt:L663-L671, docs/methodology/REGISTRY.md:L1343-L1349. Concrete fix: replace the remaining SE-based rationale with the same functional-form/support language already used later in §4.11/§5.3.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings.

Tech Debt

• No findings.

Security

• No findings.

Documentation/Tests

• No findings in the diff beyond the methodology item above.
• Static review note: the added guide/profile tests look well targeted on inspection, but I could not execute them here because pytest is unavailable.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 31e887523fb6cd1f31a5c3acc4be8b07330dfa42

---

Overall Assessment

⚠️ Needs changes

Executive Summary

• P1 The new agent-facing ContinuousDiD guidance overstates PanelProfile.has_never_treated and treatment_dose.dose_min > 0 as necessary/authoritative pre-fit gates, but ContinuousDiD.fit() actually keys off first_treat, zeroes nonzero dose on never-treated rows, and only rejects negative doses among treated units.
• That mismatch is now repeated across the public guide, docstrings, changelog/roadmap prose, and regression tests, so an autonomous agent can wrongly route valid panels away from ContinuousDiD.
• The prior count-outcome wording issue appears resolved: the updated Wooldridge Poisson guidance now correctly frames overall_att as an ASF-based outcome-scale difference rather than a multiplicative ratio.
• No estimator math, weighting, variance, or SE implementation changed in this diff.
• Static review only: python -m pytest tests/test_profile_panel.py tests/test_guides.py -q could not be run here because pytest is not installed.

Methodology

• P1 Impact: the PR documents the five profile-side ContinuousDiD checks as “necessary” and treats has_never_treated as the authoritative gate, with dose_min > 0 presented as a hard pre-fit screen in diff_diff/profile.py:L64-L114, diff_diff/guides/llms-autonomous.txt:L206-L254, diff_diff/guides/llms-autonomous.txt:L510-L558, diff_diff/guides/llms-autonomous.txt:L807-L860, CHANGELOG.md:L12-L12, ROADMAP.md:L141-L141, tests/test_guides.py:L45-L82, and tests/test_profile_panel.py:L1110-L1212. But the actual estimator contract is different: ContinuousDiD.fit() defines controls from first_treat == 0, zeroes nonzero dose on those rows, and only rejects negative dose among treated units first_treat > 0 in diff_diff/continuous_did.py:L276-L327 and diff_diff/continuous_did.py:L348-L360; that coercion is explicitly documented in docs/methodology/REGISTRY.md:L743-L745. So a panel can fail the new profile-side gate while still warning-and-fitting. Concrete fix: reword these fields as conservative “clean-data / no-coercion preflight checks” rather than necessary eligibility conditions, or add a profiling helper that also accepts first_treat so the pre-fit gate matches the estimator’s real contract; update the new tests accordingly.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings beyond the methodology contract issue above.

Tech Debt

• No findings.

Security

• No findings.

Documentation/Tests

• No separate findings beyond the methodology item above.
• Static review note: tests were not executed here because pytest is unavailable in the environment.

Path to Approval

1. Remove “necessary” / “authoritative gate” wording for has_never_treated and treatment_dose.dose_min > 0 anywhere this PR added it, and replace it with contract-accurate wording tied to first_treat-based coercion/validation.
2. Add a regression that covers first_treat == 0 rows carrying nonzero dose (including negative dose on never-treated rows), asserting the documented warning/coercion behavior in ContinuousDiD.fit(), then align the guide/profile tests with that behavior.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 25a1197a1d7f7f73fff10107bb5e49366c3cceb9

---

Overall Assessment

⚠️ Needs changes

Executive Summary

• Re-review focus: the earlier ContinuousDiD “authoritative gate” problem is mostly fixed. The new text now correctly treats has_never_treated and dose_min > 0 as standard-workflow, first_treat-dependent preflight checks rather than universal fit-time gates.
• No estimator math, weighting, variance/SE, or default-fit behavior changed in this diff.
• P1 The new agent-facing docs now go beyond the documented coercion note and present deliberate first_treat == 0 relabeling of nonzero-dose units as an agent option to “manufacture controls” / still fit ContinuousDiD. That is not supported by the Methodology Registry or the estimator’s own requirement that actual P(D=0) > 0 exist.
• P2 Some new wording also says duplicate-row failures are “hard fit-time stops,” but ContinuousDiD still silently applies last-row-wins on duplicate (unit, time) cells.
• Static review only: python -m pytest tests/test_profile_panel.py tests/test_guides.py tests/test_continuous_did.py -q could not be run because pytest is not installed.

Methodology

• P1 Impact: the prior mismatch is not fully closed because the new docs treat intentional relabeling of positive-dose / nonzero-dose units to first_treat == 0 as an available fallback for ContinuousDiD (diff_diff/guides/llms-autonomous.txt:L241-L247, diff_diff/guides/llms-autonomous.txt:L893-L900, diff_diff/profile.py:L88-L97, ROADMAP.md:L141-L141, CHANGELOG.md:L13-L13). The registry and implementation only document this as inconsistent-input coercion with a warning, while still requiring actual P(D=0) > 0 because Remark 3.1 is unimplemented (docs/methodology/REGISTRY.md:L719-L745, diff_diff/continuous_did.py:L319-L360). An autonomous agent following the worked example could therefore alter the data to create synthetic controls without any documented methodological basis. Concrete fix: remove/rewrite the “relabel to first_treat == 0” fallback everywhere; keep the coercion note as implementation behavior for bad/inconsistent inputs, not as a valid routing option. When has_never_treated == False or dose_min <= 0, direct users to re-encode treatment or choose a different estimator unless REGISTRY is updated with a documented deviation.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings beyond the documentation inconsistency noted below.

Tech Debt

• No findings. I did not see this issue tracked in TODO.md.

Security

• No findings.

Documentation/Tests

• P2 Impact: some new text now says the duplicate-row check is a “hard fit-time stop” (diff_diff/profile.py:L117-L122, diff_diff/guides/llms-autonomous.txt:L864-L866), but the estimator still silently overwrites duplicate (unit, time) rows in _precompute_structures() (diff_diff/continuous_did.py:L820-L823). That contradiction can make readers rely on fit-time protection that does not exist. Concrete fix: change this wording to “hard preflight veto / must deduplicate before fit because ContinuousDiD otherwise uses last-row-wins,” matching the actual implementation and the other accurate duplicate-row warnings already present in the guide.
• Static note: tests were not executed here because pytest is unavailable in the environment.

Path to Approval

1. Remove the new “manufacture controls / relabel to first_treat == 0” fallback from the guide, profile docstring, changelog, and roadmap, and replace it with wording that matches REGISTRY.md and ContinuousDiD.fit() exactly.
2. Update the §5.2 continuous-dose worked example and related prose so the no-never-treated / negative-dose counterexamples route agents to re-encoding or alternative estimators, not deliberate relabeling.
3. Add or update guide/profile regression tests so they lock in the corrected wording and do not normalize relabeling as an admissible methodology.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 307f424c9580d2739ab4b05313faaec4bb3b6f2e

---

Overall Assessment

⚠️ Needs changes

Executive Summary

• Re-review: the earlier user-facing “manufacture controls” guidance appears fixed in the guide/profile prose.
• P1 The replacement ContinuousDiD preflight story is still wrong: it tells agents to derive first_treat from the same dose column that profile_panel() inspects, but the method/library contract uses a time-invariant unit dose plus a separate timing variable.
• The contradiction is internal to the new §5.2 example itself: it first describes constant per-unit dose with separate first_treat, then bases the screening logic on inferring first_treat from that same dose column.
• The same assumption is now duplicated in diff_diff/profile.py and new tests/test_profile_panel.py regression prose, so the mismatch will persist unless those surfaces are corrected together.
• No estimator math, weighting, variance, or inference code changed in this diff.
• Static review only: this environment lacks numpy/pytest, so I could not execute the targeted test files.

Methodology

• P1 Impact: the new “standard workflow” in diff_diff/guides/llms-autonomous.txt:L219-L239, diff_diff/guides/llms-autonomous.txt:L530-L563, diff_diff/guides/llms-autonomous.txt:L801-L852, diff_diff/profile.py:L72-L130, and tests/test_profile_panel.py:L1111-L1199 says agents can derive first_treat from the same dose column that profile_panel() sees. That conflicts with the method contract: multi-period continuous DiD uses a time-invariant unit dose D_i and a separate timing group G_i/first_treat (docs/methodology/continuous-did.md:L65-L73), ContinuousDiD.fit() hard-rejects within-unit dose variation (diff_diff/continuous_did.py:L222-L228), and the library’s own generator emits constant dose with separate first_treat (diff_diff/prep_dgp.py:L970-L993). REGISTRY documents the untreated-group and balanced-panel requirements, but not any “derive timing from dose” workflow (docs/methodology/REGISTRY.md:L717-L723). An agent following the new guidance must either build a 0,0,d,d path to recover timing, which fit() rejects, or keep a valid constant-dose panel, in which case the dose column cannot identify first_treat at all. Concrete fix: remove the “derive first_treat from the same dose column” workflow from the guide/profile text and replace it with the actual contract: profile_panel() on the dose column alone can only provide descriptive dose-support checks; timing-dependent ContinuousDiD eligibility requires a separate first_treat column already supplied by the caller, or a new helper that validates dose and first_treat together.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• P2 Impact: PanelProfile is a public top-level type (diff_diff/__init__.py:L503-L506), but the PR adds two new required dataclass fields without defaults (diff_diff/profile.py:L179-L188). Any direct caller that instantiates PanelProfile(...) manually will now fail with TypeError, even though both additions are semantically optional. Concrete fix: make outcome_shape and treatment_dose backward-compatible additions, e.g. keyword-only fields with defaults, and add a regression test for direct construction.

Tech Debt

• No findings.

Security

• No findings.

Documentation/Tests

• P3 Impact: the new docstring in tests/test_continuous_did.py:L903-L911 still says the coercion path “lets agents legally relabel negative-dose units as first_treat == 0,” which contradicts the corrected guide/profile/registry position. The behavior under test is fine; the comment is not. Concrete fix: rewrite the docstring to say coercion is observed implementation behavior for inconsistent inputs, not a sanctioned routing option.
• Residual risk: static review only; I could not run tests/test_profile_panel.py, tests/test_guides.py, or tests/test_continuous_did.py because the environment is missing numpy/pytest.

Path to Approval

1. Remove/rewrite the “derive first_treat from the same dose column” workflow in diff_diff/guides/llms-autonomous.txt, diff_diff/profile.py, CHANGELOG.md, and ROADMAP.md.
2. Rewrite §5.2 and the new tests/test_profile_panel.py regression prose around the actual ContinuousDiD contract: time-invariant unit dose D_i, separate first_treat/G_i.
3. If predictive pre-fit screening is still desired, add a helper or optional first_treat= input that validates dose and first_treat together; otherwise explicitly narrow the docs to descriptive dose-support checks only.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 60014f5c25cc54e67380a1d0c59cf93b62dc0f72

---

Overall Assessment

✅ Looks good

Executive Summary

• Re-review: the prior P1 about telling agents to derive first_treat from the dose column is resolved on the public-facing surfaces. The updated ContinuousDiD guidance now matches the canonical contract: time-invariant unit dose D_i, separate caller-supplied first_treat, and separate fit-time validation of first_treat (diff_diff/profile.py:L65-L141, diff_diff/guides/llms-autonomous.txt:L206-L283, diff_diff/guides/llms-autonomous.txt:L810-L919, diff_diff/continuous_did.py:L222-L360).
• The new OutcomeShape.is_count_like guidance is consistent with the actual Poisson path: WooldridgeDiD(method="poisson") rejects negative outcomes, and the guide now correctly describes the reported overall_att as an ASF-based outcome-scale difference rather than a multiplicative reported estimand (diff_diff/profile.py:L700-L725, diff_diff/wooldridge.py:L1102-L1245, docs/methodology/REGISTRY.md:L1333-L1349).
• The public API propagation looks complete for this PR: the new dataclasses are exported from diff_diff, and PanelProfile keeps backward compatibility by defaulting the new fields (diff_diff/__init__.py:L252-L258, diff_diff/__init__.py:L503-L509, diff_diff/profile.py:L189-L193, tests/test_profile_panel.py:L1325-L1362).
• P3: two new regression docstrings still restate the old “derive first_treat from dose” workflow. This is test-only prose, not a runtime or methodology bug, but it leaves stale language in the tree (tests/test_profile_panel.py:L1110-L1137, tests/test_profile_panel.py:L1184-L1199).
• Static review only: I could not run the touched tests in this environment because pytest is unavailable (python -m pytest ... failed with No module named pytest).

Methodology

• No unmitigated findings. The affected methods are ContinuousDiD and WooldridgeDiD, but only through pre-fit descriptive guidance, not estimator math or inference. The updated text aligns with the ContinuousDiD contract in docs/methodology/continuous-did.md:L65-L76, docs/methodology/continuous-did.md:L265-L269, docs/methodology/REGISTRY.md:L717-L745, and diff_diff/continuous_did.py:L222-L360, and with the Poisson ASF path documented for WooldridgeDiD in docs/methodology/REGISTRY.md:L1333-L1349 and implemented at diff_diff/wooldridge.py:L1102-L1245.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings.

Tech Debt

• No findings.

Security

• No findings.

Documentation/Tests

• Severity: P3
  Impact: tests/test_profile_panel.py:L1110-L1137 and tests/test_profile_panel.py:L1184-L1199 still describe a “standard ContinuousDiD workflow” where first_treat is derived from the same dose column. That contradicts the now-corrected contract in diff_diff/profile.py:L65-L141 and diff_diff/guides/llms-autonomous.txt:L206-L283, diff_diff/guides/llms-autonomous.txt:L810-L919. It will not change runtime behavior, but it can mislead future editors and make this methodology regression easier to reintroduce.
  Concrete fix: rewrite those test docstrings/comments to the canonical formulation already used elsewhere: time-invariant per-unit dose D_i, separate caller-supplied first_treat, and profile-side screening only on the dose column.

• Residual risk: static review only; targeted pytest execution was not possible in this environment because pytest is not installed.