The analysis

05 — The adversarial review: what was attacked, what survived, what broke

Scope: the third and last honesty problem — the one that was not caught during planning. This file documents the adversarial lane: its stance, its verdicts, the defects it found, the systemic finding that both headline verification numbers are blind to the same class of defect, and the remediation state at the time of writing.

Source root: as 01-premise-and-reaim.md §0. The review is verification/ADVERSARIAL-REVIEW.md (740 lines), written by the adversary lane on 2026-10-29.

Method: reading the review in full, plus independent re-runs of the three harnesses during the writing of this file. The remediation status in §10 was observed directly from the source tree at 2026-10-29 17:38 PT (2026-10-30T00:38Z) and is timestamped because the tree was being edited concurrently.

1. The review's stance

The review states its own rules of engagement:

"Stance: falsify, not confirm. A review that finds nothing is a failed review." — verification/ADVERSARIAL-REVIEW.md:4

"Write scope: this file only. Nothing under app/ or any top-level .md was modified. All mutation testing was done on a throwaway copy at /tmp/adv-cine; the working tree was verified unchanged afterwards." — verification/ADVERSARIAL-REVIEW.md:5-7

That second line matters for the integrity of the exercise: the reviewer did not fix what it found, and it did not mutate the artefact it was reviewing. Every defect it reports is reproducible against the tree as it stood.

The review ran 24 falsification attempts (verification/ADVERSARIAL-REVIEW.md:630-657), of which 13 succeeded in falsifying or materially qualifying a claim.

2. The verdicts

The review's verdict table (verification/ADVERSARIAL-REVIEW.md:31-40), reproduced in full:

Two CONFIRMED, four PARTIALLY CONFIRMED, one REFUTED. The distribution is the honest shape of the result: the entry's engineering held up; its documentation did not.

2.1 The bottom line

"Not safe to submit as-is. One CRITICAL defect breaks the entry's headline deliverable: installed exactly as app/README.md documents, the ComfyUI nodes do not load at all. A judge who follows the README gets an empty category, not six nodes." — verification/ADVERSARIAL-REVIEW.md:13-15

"Do not submit as-is. Fix defect #1 first — without it the entry's primary artefact does not install, and that is the failure a judge is most likely to hit first and remember." — verification/ADVERSARIAL-REVIEW.md:719-720

3. The CRITICAL defect: the install path is broken

3.1 The claim

app/README.md:33-41 documents the install as:

cd ComfyUI/custom_nodes
git clone <this-repo> ComfyUI-Cinematographer
# no dependencies — the engine is pure stdlib

"Restart ComfyUI. The nodes appear under the Cinematographer category." — app/README.md:41

3.2 What actually happens

The review simulated the documented clone exactly:

$ ls -a /tmp/fakecomfy/custom_nodes/ComfyUI-Cinematographer
.  ..  00-BREAKDOWN.md  01-SPEC.md  02-DEMO-SCRIPT.md  03-SUBMISSION.md
analysis  app  raw-articles  verification

$ test -f .../ComfyUI-Cinematographer/__init__.py && echo YES || echo NO
NO

— verification/ADVERSARIAL-REVIEW.md:489-494

The pack root has no __init__.py. ComfyUI's loader requires one.

3.3 The loader evidence

The review read ComfyUI's own loader rather than assuming:

# ComfyUI nodes.py:2261-2276, quoted at verification/ADVERSARIAL-REVIEW.md:499-508
elif os.path.isdir(module_path):
    sys_module_name = module_path.replace(".", "_x_")
...
    module_spec = spec_from_file_location(sys_module_name, os.path.join(module_path, "__init__.py"))
    module_dir = module_path
module = importlib.util.module_from_spec(module_spec)
module_spec.loader.exec_module(module)

Two facts from that source, both load-bearing:

1. The directory branch resolves <pack_root>/__init__.py via spec_from_file_location (:504).
2. init_external_custom_nodes iterates os.listdir(custom_node_path) — top level only, no recursion (verification/ADVERSARIAL-REVIEW.md:512-513).

The second fact is why app/cinematographer/__init__.py does not save it. ComfyUI never looks two levels down.

Reproducing the loader's directory branch:

entry file: .../ComfyUI-Cinematographer/__init__.py exists: False
LOAD FAILED: FileNotFoundError [Errno 2] No such file or directory:
             '.../ComfyUI-Cinematographer/__init__.py'

— verification/ADVERSARIAL-REVIEW.md:516-519

3.4 The secondary and tertiary defects

Secondary: app/README.md:47 documents python -m pytest tests/ -q but never says to install pytest. requirements.txt lists it only as a commented-out optional, and pyproject.toml puts it behind [project.optional-dependencies] dev. On a clean machine:

$ python3 -m pytest tests/ -q
No module named pytest
EXIT=1

— verification/ADVERSARIAL-REVIEW.md:536-538

Tertiary: acceptance criterion 1 (python -c "import cinematographer") works from app/ but not from the repo root, which is where the README's install section leaves the reader (verification/ADVERSARIAL-REVIEW.md:541-542).

Quaternary: there is no pyproject.toml at the pack root, so ComfyUI's config_parser.extract_node_configuration finds nothing — the [tool.comfy] block is at app/pyproject.toml (verification/ADVERSARIAL-REVIEW.md:523-524).

3.5 Why this defect is the most damaging one

"Impact: acceptance criterion 8 ('README setup works from a clean clone') fails, and the entry's primary artefact — a ComfyUI custom node pack — is not installable by the documented route. The panel and the tests work, so a judge who only runs those will not notice." — verification/ADVERSARIAL-REVIEW.md:526-528

That last sentence is the crux. The entry's three verification harnesses all pass, and all three pass without ever loading the pack into ComfyUI. The verification story is strong precisely in the places that cannot see this defect.

This is the single most important structural lesson in the whole build: the entry verified everything it could verify without ComfyUI, and the one thing it could not verify is the one thing that was broken.

4. What the review falsified

4.1 Attempt 1 — can the differential check fail at all? (FAILED to falsify)

| 1 | Reproduce the 289-check baseline | FAILED to falsify — 289, exit 0, exactly as claimed | — verification/ADVERSARIAL-REVIEW.md:634

The check is real. It was re-run during the writing of this file and reproduced exactly:

$ PYTHONDONTWRITEBYTECODE=1 python3 verification/differential_check.py verification/go-reference-dump.json
differential checks run: 289
ALL FIXTURES AGREE with the original Go implementation.

4.2 Attempt 2 — what does it miss? (SUCCEEDED)

11 of 30 injected defects were missed, and they are all the same kind:

"11 of 30 injected defects were MISSED, and they are all the same kind: a threshold or a boundary constant moved slightly, where no fixture sits near the boundary." — verification/ADVERSARIAL-REVIEW.md:134-135

The full table (verification/ADVERSARIAL-REVIEW.md:137-153):

Root cause, measured:

"The fixture set has near-zero boundary diversity. For the 9 framing fixtures, in_frame has exactly one distinct value (True) and level has exactly one distinct value (0 / True). No fixture exercises a boundary, so moving a boundary is invisible. The check is strong on arithmetic and blind to thresholds." — verification/ADVERSARIAL-REVIEW.md:155-158

And why it matters:

"This matters because thresholds are where the cinematography actually lives. 0.4 <= frac <= 0.75 is the entire medium-shot rule; 0.12 is the entire rule-of-thirds tolerance. Both can be wrong and the entry's strongest claim still prints 289 / ALL FIXTURES AGREE." — verification/ADVERSARIAL-REVIEW.md:160-162

4.3 Attempt 3 — does it compare everything it says it compares? (SUCCEEDED)

No. Three surfaces are compared only partially, and the skipped fields really do differ (verification/ADVERSARIAL-REVIEW.md:164-201).

SceneFramingExtras — differential_check.py:233-246 compares rule, pass, and (except for size_importance) value. It never compares target, weight, or correction:

Go     extras_big[0].correction = "a prop out-sizes the key subject — resize/reposition so the subject dominates"
Python extras_big[0].correction = "a prop out-sizes the key subject - resize/reposition so the subject dominates"

— verification/ADVERSARIAL-REVIEW.md:173-174

Em dash vs hyphen. The framing rules get this right (geometry.py:368-371 uses \u2014); the scene rules were rewritten to ASCII and nothing notices (verification/ADVERSARIAL-REVIEW.md:177-179).

EvaluateContinuity — differential_check.py:203-210 compares rule, pass, shot_a, shot_b. It never compares detail:

Go     30_degree: "camera angle changed 3° (need ≥30° to avoid a jump cut)"
Python 30_degree: "camera angle changed 3 deg (need >=30 deg to avoid a jump cut)"

Go     180_axis:  "axis crossed — the subject's screen direction flips between the cuts"
Python 180_axis:  "axis crossed - the subject's screen direction flips between the cuts"

— verification/ADVERSARIAL-REVIEW.md:185-189

Three of the four continuity_go_test details differ (app/cinematographer/continuity.py:121, :140-141).

size_importance — the slot is counted in the 289, then the value is skipped (differential_check.py:243-244). So 2 of the 289 headline checks compare pass and nothing else (verification/ADVERSARIAL-REVIEW.md:194-196).

The consequence:

"verification/README.md:12-13 claims the port 'reproduces the original Go cinematography implementation value-for-value'. For the numeric values, that is true and well-evidenced. For the report's user-facing strings — which are what the panel and the ComfyUI nodes actually display — it is false for 2 of the 14 rules and for both continuity rules." — verification/ADVERSARIAL-REVIEW.md:198-201

4.4 Attempt 4 — is eyeline verified at all? (SUCCEEDED — it is not)

"eyeline (rule 12 of 14) is emitted only when len(scene.characters) >= 2 (app/cinematographer/scene.py:231). Both differential extras fixtures have exactly one character, so eyeline is never produced on the Go side and never compared" — verification/ADVERSARIAL-REVIEW.md:205-207

The dump confirms it:

extras_big -> [('size_importance', False, 10.125), ('foreground_occlusion', True, 0)]
extras_ok  -> [('size_importance', True, 0),      ('foreground_occlusion', True, 0)]

— verification/ADVERSARIAL-REVIEW.md:210-211

One of the fourteen rules is not differentially verified at all, and the entry lists it among the fourteen without disclosing that. The README counts it (app/README.md:78), the spec counts it (01-SPEC.md:81), and the UNVERIFIED table does not mention it (verification/README.md:117-126).

4.5 Attempt 5 — the combined mutation matrix (SUCCEEDED)

| 5 | Combined mutation matrix (both checks) | SUCCEEDED — 12 constants unconstrained by either | — verification/ADVERSARIAL-REVIEW.md:638

This is the systemic finding, developed in §6.

4.6 Attempt 8 — the 1800-call solver sweep (FAILED to falsify)

| 8 | 1800-call solver invariant sweep | FAILED to falsify — zero anomalies | — verification/ADVERSARIAL-REVIEW.md:641

The convergence claim survived the strongest attack in the review. See 04-convergence-and-build.md §1.5.

4.7 Attempt 17 — structural mutations against the validator (FAILED to falsify)

| 17 | 6 structural mutations vs the workflow validator | FAILED to falsify — all 6 caught | — verification/ADVERSARIAL-REVIEW.md:650

— verification/ADVERSARIAL-REVIEW.md:390-397

The validator is a real check.

4.8 Attempt 21 — the trademark grep (FAILED to falsify)

| 21 | Grep the whole tree for Higgsfield / MiniMax H3 / closed APIs | FAILED to falsify — clean | — verification/ADVERSARIAL-REVIEW.md:654

4.9 Attempt 20 — simulating the install (SUCCEEDED)

| 20 | Simulate ComfyUI's load_custom_node against a clean clone | SUCCEEDED — FileNotFoundError; the install path is broken | — verification/ADVERSARIAL-REVIEW.md:653

4.10 The discarded findings

The review records its own false positives, which is a mark of a careful reviewer:

"Note on discarded findings. My first workflow-input audit reported ~15 'UNDECLARED inputs' (clip, text, device, codec, audio_vae, and the CrossViewWarp optional block). These were artifacts of a crude regex extractor that ignored "optional" blocks and old-style INPUT_TYPES. I re-verified each against the real sources: every one is declared. No input-name defect exists. They are recorded here only so a later reader does not rediscover the false positives and report them." — verification/ADVERSARIAL-REVIEW.md:659-664

That paragraph is worth as much as the defect register. It prevents a future reviewer from re-reporting a non-defect, and it documents the reviewer's own error rate.

5. The workflow graph findings

5.1 The validator passes

$ PYTHONDONTWRITEBYTECODE=1 python3 app/workflows/validate_graphs.py
PASS  reshoot_ltx_crossview.json  (35 nodes, 49 links)
PASS  reshoot_wan22_fun_control.json  (25 nodes, 33 links)
All structural checks passed. This does NOT mean the graphs run: they have never been executed.
EXIT=0

— verification/ADVERSARIAL-REVIEW.md:381-385; re-run during the writing of this file with the same result.

5.2 The slot-count audit

The review did not trust the table. It downloaded the real upstream sources and parsed every class_type with ast, handling both RETURN_TYPES = (...) and the new io.Schema(outputs=[...]) form (verification/ADVERSARIAL-REVIEW.md:401-402).

Result: 32 of 33 slot counts match upstream exactly. All 33 class types exist. (verification/ADVERSARIAL-REVIEW.md:404)

5.3 The one that does not — and the provenance claim it falsifies

app/workflows/validate_graphs.py:72:    "SamplerCustomAdvanced": 1,

comfy_extras/nodes_custom_sampler.py:1033-1035:
            outputs=[
                io.Latent.Output(display_name="output"),
                io.Latent.Output(display_name="denoised_output"),
            ],

— verification/ADVERSARIAL-REVIEW.md:411-418

The table said 1; upstream declares 2.

The count is not the important part. The claim is. The validator's docstring says:

"The output-slot table is the load-bearing part. Each count was read from the node source named in the graph's own `_README PROVENANCE block, **not guessed**." — app/workflows/validate_graphs.py:22-24`

Both graphs' _READMEs repeat the claim. One entry is a guess. The review's verdict:

| H4 | "every class_type is in a table of output-slot counts read from the node sources" / "not guessed" | validate_graphs.py:23-25, both graphs' _README | False for SamplerCustomAdvanced (table 1, upstream 2). | — verification/ADVERSARIAL-REVIEW.md:575

And it made the reviewer's top-three list:

"2. validate_graphs.py:72 — a 'read from the node sources, not guessed' table with one guess in it." — verification/ADVERSARIAL-REVIEW.md:738

This is the most interesting defect in the register, because of what it is a defect in. The entry's differentiator is honesty about measured vs. guessed. The one place it guessed, it claimed not to have guessed — and the claim was in a docstring whose entire purpose was to assert the opposite.

5.4 The graphs are honestly marked

"### The graphs are honestly marked" — verification/ADVERSARIAL-REVIEW.md:431

The review's verdict on the _README discipline is positive, and it is the one place the review calls the entry's work exemplary:

"The workflow _READMEs and app/web/README.md are exemplary." — verification/ADVERSARIAL-REVIEW.md:40

6. The systemic finding: 12 unconstrained constants

6.1 The finding

"The residual risk is concentrated in one place: the entry's two headline numbers both measure value fidelity and neither measures threshold fidelity. Twelve constants — including the entire medium-shot size band and the entire rule-of-thirds tolerance — can be wrong while 289 / ALL FIXTURES AGREE and 69 passed both print green." — verification/ADVERSARIAL-REVIEW.md:727-730

The defect register entry:

| 5 | MAJOR | 12 engine constants are unconstrained by the differential check, the unit suite, and the parity check. Both headline verification numbers miss the same class of defect. | geometry.py:368-371, framing.py:104,116,125,164,192, autoframe.py:48, continuity.py axis sign + 180 clamp | — verification/ADVERSARIAL-REVIEW.md:614

6.2 Why "both headline numbers" is the finding

The entry has three independent verification harnesses:

All three are blind to the same class of defect. They all compare values at points, and none of them places a point on a boundary.

The review states the mechanism:

"The fixture set has near-zero boundary diversity. For the 9 framing fixtures, in_frame has exactly one distinct value (True) and level has exactly one distinct value (0 / True). No fixture exercises a boundary, so moving a boundary is invisible." — verification/ADVERSARIAL-REVIEW.md:155-158

6.3 Why it looks worse than it is

"A judge who mutates one constant will find that, and it will look worse than it is, because the entry has spent so much credibility claiming otherwise. Adding boundary fixtures is a two-hour fix that removes the whole class." — verification/ADVERSARIAL-REVIEW.md:730-732

That is the sharpest sentence in the review. The entry's credibility investment in its verification story raises the cost of the gap, because the gap is in the same dimension the entry claims strength in.

6.4 The fix, and its cost

"Add boundary fixtures. One fixture per threshold, sitting exactly on the boundary. This closes the 12-constant hole in both the differential check and the parity check, and it is the highest-value verification work available." — verification/ADVERSARIAL-REVIEW.md:706-708

Two hours of work, not done. The 12 constants are still unconstrained as of the time of writing.

7. The honesty findings

The review's harshest section, because honesty is the entry's differentiator:

"This is the entry's differentiator, so it gets the harshest treatment. Verdict: PARTIALLY CONFIRMED." — verification/ADVERSARIAL-REVIEW.md:548-549

7.1 What is genuinely excellent

"This is well above the norm. The defects below are therefore more damaging, not less — they are inconsistencies in a document set that otherwise earns its credibility." — verification/ADVERSARIAL-REVIEW.md:565-566

The review names the good parts explicitly (verification/ADVERSARIAL-REVIEW.md:551-563):

• The workflow _READMEs
• app/web/README.md's "what is real, and what is mocked" section
• 03-SUBMISSION.md:133-145 — the explicit honesty ledger, including "38 cinematography rules" ⛔ false and "Beats Higgsfield's Cinema Studio" ⛔ unfalsifiable
• framing.py:102-103 — documenting that in_frame checks the face, not the feet

7.2 The ten claims that a runnable check contradicts or does not back

— verification/ADVERSARIAL-REVIEW.md:570-581

7.3 The pattern in the ten

They are not ten unrelated mistakes. They cluster:

The third cluster is the dangerous one, because it is the failure mode the entry was built to avoid. H4 and H7 are both cases of something sitting inside the verified set that no check reaches.

7.4 The spec-vs-code divergence

A separate finding, not in the H-series:

"01-SPEC.md:229 says yellow is 45° < |azimuth| ≤ 60°, **or |elevation| > 30°**, and 01-SPEC.md:230 says red is 'beyond both'. app/cinematographer/reliability.py:75-102 requires both axes to be out for red only when azimuth exceeds 60; an elevation-only violation falls through to red" — verification/ADVERSARIAL-REVIEW.md:585-588

— verification/ADVERSARIAL-REVIEW.md:590-595

The review's judgement is fair to both sides:

"The code's behaviour is arguably the more useful one (a 40° elevation is outside the trained envelope), but the spec and the code disagree, and app/web/README.md:118 describes the constants as 'taken from the spec'." — verification/ADVERSARIAL-REVIEW.md:597-599

8. The defect register

The review's full register, 17 defects (verification/ADVERSARIAL-REVIEW.md:608-626):

1 CRITICAL · 5 MAJOR · 11 MINOR.

8.1 The severity distribution is the story

The engine's arithmetic is not in the register. Not one defect is a wrong measurement. The defects are in the claims about the measurements, the packaging, and the coverage.

That is the honest summary of this build: the code is better than the documentation around it.

9. What survived

Worth stating explicitly, because a defect register alone misrepresents the result.

"After that, this is a strong entry with an unusually honest verification story. The differential check, the parity check and the workflow validator are all real and all re-runnable; I attacked each of them and they held up far better than I expected." — verification/ADVERSARIAL-REVIEW.md:722-725

10. Remediation status

Observed directly from the source tree at 2026-10-29 17:38 PT (2026-10-30T00:38Z). The tree was being edited concurrently by another lane while this file was written, so this section is a snapshot, not a final state.

10.1 Fixed

10.2 Partially fixed

The CRITICAL defect is not yet closed. The shim is necessary but not sufficient: either the README's clone target must change, or a repo-root __init__.py must be added, or cinematographer/ must move to the repo root. The review listed all three options (verification/ADVERSARIAL-REVIEW.md:696-700).

10.3 Not fixed

10.4 The verification harnesses still pass after the edits

Re-run at 17:38 PT:

$ PYTHONDONTWRITEBYTECODE=1 python3 verification/differential_check.py verification/go-reference-dump.json
differential checks run: 289
ALL FIXTURES AGREE with the original Go implementation.

$ PYTHONDONTWRITEBYTECODE=1 python3 app/workflows/validate_graphs.py
PASS  reshoot_ltx_crossview.json  (35 nodes, 49 links)
PASS  reshoot_wan22_fun_control.json  (25 nodes, 33 links)
All structural checks passed. This does NOT mean the graphs run: they have never been executed.

The three edits did not break anything. But the CRITICAL defect is invisible to all three harnesses, which is exactly the point of §3.5.

11. What the review could not verify

The review's own list (verification/ADVERSARIAL-REVIEW.md:668-690), which is the most important section for anyone relying on this package:

The second row is the one to carry forward. The entry's strongest claim — 289 checks against the original Go implementation — reduces to agreement with a JSON file whose provenance is asserted, not audited. The harness source is preserved, so the claim is auditable, but nobody audited it.

12. The three claims the reviewer would least like a judge to find

"The three claims I would least like a judge to find first, in order:

1. verification/README.md:122 — a verification README that tells you to run a script that does not exist.
2. validate_graphs.py:72 — a 'read from the node sources, not guessed' table with one guess in it.
3. app/README.md:41 — 'Restart ComfyUI. The nodes appear under the Cinematographer category.' They do not."

— verification/ADVERSARIAL-REVIEW.md:734-740

Two of the three are now fixed (§10.1). The third is the CRITICAL one, and it is only half fixed (§10.2).

The ordering is itself informative. The reviewer ranked a dead command in a verification README above a broken install, because the dead command is the one that most directly contradicts the document's own stated rule:

"The rule for this directory: a claim lives here only if a command in this file reproduces it. Anything else is marked UNVERIFIED." — verification/README.md:5-6

A verification README that violates its own rule is a worse look than a packaging bug, because the README's rule is the entry's whole thesis in miniature.

13. Decision register — the adversarial phase

A9 is worth noting: the review lane's write scope was this file only (verification/ADVERSARIAL-REVIEW.md:5). The fixes were a separate lane's work. That separation is why the defect register and the remediation status can be read independently.

14. Evidence index for this file