RES — TFW-56: Remove the Review Mode Axis — Iteration 1

Date: 2026-08-13 Author: Researcher (Claude Code) Status: 🔬 RES — iteration 1 complete Parent HL: HL-TFW-56 — 🔒 FROZEN 2026-08-13 Mode: Pipeline · focused (loops_per_stage: 1) Stage files: 1_briefing.md · 2_gather.md · 3_extract.md · 4_challenge.md

---

Research Context

HL TFW-56 proposes deleting the code / docs / spec review axis on one empirical claim: across 18 REVIEW files in this repository, 38 mode-specific checklist rows produced 33 ✅, 4 N/A, 1 ⚠️ and 0 ❌ — "the axis has never produced a finding". The HL then applies TFW-53's base-rate rule (a check that cannot fail is ceremony) and promotes three survivor rows into the universal checklist.

The HL itself named the blind spot: the claim was never tested outside this repository. This iteration measured it against two external TFW installs — ai-first-devices (AFD) and helpdesk — totalling 203 mode-carrying reviews and 637 mode-specific rows, audited every consumer of the config key and header, built the coverage matrix, and attacked H6 (priming vs rows) head-on.

The finding is that the HL's central empirical claim does not hold outside this repository, and the frozen §3 coverage table that depends on it is wrong in two specific places. The rest of the HL — the gate, the three duplicate files, the wrong-by-default config key, the stale pointers, docs/spec being synonyms — survives intact and is well-supported.

Briefing

See 1_briefing.md. Owner directions at the gate: measurement gated on a review-surface drift check (Q1b); helpdesk included as a second replication (Q3); H6 approach delegated to the researcher (Q2) — resolved as both a pre-declared observational test and the decision H6 forces. Autonomous execution authorised for stages 2-4.

Decisions

#	Decision	Rationale
D1	The drift gate passes. AFD's entire review surface is byte-identical to this repo's 1.0.0 despite VERSION 0.9.0; helpdesk's drift is Evidence-Layer only and orthogonal to the mode axis. All 203 reviews are comparable; none excluded	Per-file diff across 9 review-surface files × 2 installs (2_gather G1)
D2	H3 is refuted. Mode rows fire at 10.2% raw / 7.7% hard across 637 rows, against a universal-row baseline of 8.4% raw / 8.3% hard in the same reviews. All eight rows fired somewhere. The "never fires" property belongs to this repository's 39-row sample, not to the mechanism	2_gather G2/G3; stress-tested in 4_challenge C1
D3	The two HL claims separate: one is true, one is false. "Mode rows never changed a verdict" is true (0 of 203 reviews had a mode row as the sole non-✅). "Mode rows produce no findings" is false (49 hard non-✅, incl. 20 ❌; 62 of 65 do not restate a failing universal row)	2_gather G4; 3_extract E4
D4	H1 is refuted as stated. The HL's disposition table misclassifies two rows: Test coverage (23.4% — the highest-firing row of all eight) and Code quality (6 hard ❌) are marked "already covered" and are not	3_extract E1
D5	Four mode rows are one check in three genres. Test coverage, Analytical quality, Source verification and Source attribution all produce findings of a single shape: the artifact carries a green signal that does not establish the claim. Combined firing 16.1% (28/174) — the most productive check in TFW review. The HL's promoted wording ("claims traceable to sources") captures roughly a third of it	3_extract E2
D6	Corrected survivor set is four rows — S1 Evidence bears on the claim · S2 Backward compatibility · S3 Design soundness · S4 Safety — collapsed by residue, not by name. Content quality is a genuine duplicate and is correctly dropped	3_extract E2; stress-tested 4_challenge C5
D7	H2 has a hole. Three docs/spec verify actions (spot-check claims · citations traceable to real artifacts · data claims against primary sources) have no unconditional home. They are Verify-stage actions, so promoting judge rows does not rescue them. The HL asserts H2 from code's two actions only	3_extract E1
D8	H5 is confirmed and closed. workflows/review/{code,docs,spec}.md are byte-identical across three installs, two framework versions and two product domains. Never used as an extension point	2_gather G5
D9	H4 splits. No unknown consumer exists — gen_docs.py and editions/ are clean; every hit is already in the HL's file list. But update.md categorises files, never keys, so a removed config key is invisible to its 🟢/🟡/🔴 triage. Failure mode is silent orphaning, not corruption	2_gather G5; 3_extract E5
D10	H6 is unresolved and no longer decision-critical. The observational test is unavailable, not null: the 77 unlabelled REVIEW files are structurally different documents (pre-TFW-38 or bespoke layouts, median 0 parseable Judge rows), not labelled reviews minus the label. Since the rows demonstrably carry signal, deletion loses coverage whether or not priming also existed	4_challenge C4
D11	C1 (the HL as frozen) is eliminated; C3 is the strongest survivor — delete the axis, promote the corrected four rows, and migrate the three orphaned verify actions into verify.md. C4 (project-optional) survives on the HL's own pre-registered filter and is the honest runner-up	4_challenge C2/C3
D12	The promotion design needs structure, not just appended rows. 10-11 flat equal-weight rows exceed the 5-9 checklist band and invite the composite dilution effect; always-present rows raise habituation exposure for low-firing checks. F21 explicit-N/A becomes load-bearing rather than decorative	4_challenge C3; 2_gather G7
D13	Self-correction, recorded. Gather's first headline said mode rows are more productive than universal rows. On the hard measure (discounting ⚠️ cells carrying "acceptable / not blocking / TS did not require"), the ranking reverses: 7.7% vs 8.3%. The defensible claim is equal productivity, ~8% each — which still refutes "never fires"	4_challenge C1; Gather headline amended in place

Open Questions

#	Question	Status	Answer
Q1	Are the external corpora valid replications?	✅ closed	Yes — D1. AFD is byte-identical on the whole review surface; helpdesk differs only in Evidence-Layer rows
Q2	Do mode rows produce findings outside this repo?	✅ closed	Yes — D2. 49 hard non-✅ incl. 20 ❌
Q3	Did a mode row ever flip a verdict?	✅ closed	No — D3. 0 of 203 as sole driver
Q4	Which rows are genuinely absent from the universal set?	✅ closed	Four by residue — D6
Q5	Does anything unknown consume the key or header?	✅ closed	No — D9
Q6	Does /tfw-update handle a removed framework key?	✅ closed	No rule exists — D9. Silent orphaning
Q7	Is a mode file ever used as a project extension point?	✅ closed	Never — D8
Q8	Did the label prime the reviewer?	🔴 unresolved	No test available in this corpus — D10. Two weak probes point away from strong priming; external evidence says role labels do affect LLM output
Q9	Does S3 (design soundness) become a row, or a sharpening of U2 Philosophy aligned?	🟡 open	TS/HL decision. 4 of 6 ❌ are arguably principle violations. Silence loses 6 ❌
Q10	Does the axis hold in non-software, non-markdown domains?	🔴 untestable	No corpus exists in reach. Named blind spot, carried forward

Hypotheses (from HL §10)

#	Hypothesis	HL Status	RES Status	Evidence
H1	The eight mode rows contain exactly three checks absent from the universal set; the other five are synonyms or already mandated	open	❌ REFUTED as stated	Five residues, collapsing to four distinct checks. Test coverage (23.4%) and Code quality (6 ❌) are misclassified as "already covered". 3_extract E1/E2
H2	No verify action is lost: code's two distinctive actions are already unconditional	open	🟡 PARTIALLY CONFIRMED	True for all four code actions. False for three docs/spec actions, which the HL never enumerated. 3_extract E1
H3	The finding replicates in AFD: mode rows produce ~0 findings across ~149 reviews	open	❌ REFUTED	408 rows, 20 ❌ + 18 ⚠️, 9.3% raw / 8.3% hard. Helpdesk independently: 190 rows, 14.2% raw. 2_gather G2
H4	No consumer breaks; update.md's CONFIG merge handles a removed key	open	🟡 SPLIT	Consumer audit ✅ clean (incl. gen_docs.py, editions/). Removed-key handling ❌ — no rule exists. 2_gather G5, 3_extract E5
H5	No project uses the mode files as an extension point	open	✅ CONFIRMED	Byte-identical across 3 installs / 2 versions / 2 domains. 2_gather G5
H6	The axis's value was in its rows, not in priming	open	🔴 UNRESOLVED — test unavailable	The unlabelled corpus is not a control group. Reported as unavailable, not as a null. Decision-relevance collapsed by D2/D3. 4_challenge C4

The HL's own pre-registered filter, applied honestly. HL §10 states: "H3 false → the axis works in other projects; make it project-optional instead of removing it." H3 is false. By the HL's written rule, the indicated response is C4, not C1. Research does not overrule that — it reports that the condition the HL wrote has fired, and that C3 is a stronger option than either, for the reasons in 4_challenge C3.

HL Update Recommendations

Refinements — free sections, coordinator applies

#	§	What to update	Source
R1	§2	Replace "The axis has never produced a finding" with the measured position: 637 rows across 203 reviews in 3 repositories, 10.2% raw / 7.7% hard non-✅ against an 8.4% / 8.3% universal baseline; 0 of 203 verdict flips. Keep the local 38-row table as the local sub-sample it is	D2, D3 · 2_gather G2
R2	§2	Correct "A REVIEW header reads docs + code — multi-select invented in the field": 19 of 203 labels (9%) deviate from the enum — 6 multi-value, 13 with free-text qualifiers, of which 8 encode verification depth, not genre	2_gather G6
R3	§2	Add: docs and spec being synonyms is confirmed — their four rows collapse into a single residue that fires hardest in code reviews	D5 · 3_extract E2
R4	§7.2	Add citations: Gawande Do-Confirm selection-by-consequence (supports the Safety row at 4.0%); the 5-9 checklist band; LLM-judge composite dilution and redundant-criteria degradation	2_gather G7
R5	§8	/tfw-update removed-key semantics: ⬜ unverified → 🔴 verified gap. No rule exists; failure mode is silent orphaning of default_mode: code in existing projects, not corruption	D9 · 3_extract E5
R6	§9	Update three risk rows: "a mode row carried value the 38-fill sample cannot show" — materialised, probability Medium → Confirmed; "an external project depends on the mode files as an extension point" — closed, H5 confirmed; "the axis's real function was priming" — unresolved, test unavailable	D2, D8, D10
R7	§9	Add new risk: promoted rows exceed the 5-9 checklist band and dilute the checklist (10-11 flat equal-weight rows). Mitigation: structure + explicit-N/A grammar, not appended rows	D12 · 4_challenge C3
R8	§10	Mark H1 ❌ refuted-as-stated · H2 🟡 partial · H3 ❌ refuted · H4 🟡 split · H5 ✅ confirmed · H6 🔴 unresolved. Record that the §10 filter condition for H3 has fired	All
R9	§10	Close the "External base rate" and "Consumer audit" blind spots; keep "Priming vs rows" open with the reason it could not be closed; keep "Non-code projects" open — no corpus in reach	D10, Q10
R10	§2	Add the trace-integrity observation: AFD's .tfw/VERSION reads 0.9.0 while its review surface is byte-identical to 1.0.0 — VERSION does not track file drift. Affects any future cross-project replication	2_gather G1

Amendment Proposals — frozen sections, owner verdict required

The frozen unit is the declarative claim (conventions.md §3.5). Each row below targets a specific claim, not section prose.

#	§	Type	Proposed change	Evidence	Cost	Alternatives considered
A1	§3 (coverage table)	SUPERSEDE	Replace the 8-row disposition table. Promote four rows by residue: S1 Evidence bears on the claim (absorbs Test coverage · Analytical quality · Source verification · Source attribution), S2 Backward compatibility, S3 Design soundness, S4 Safety. Drop Content quality as a true duplicate of U4. The current table's "Test coverage → already covered" and "Code quality → already covered" are contradicted by measurement	Test coverage 23.4% non-✅ (highest of eight); Code quality 6 hard ❌ with contract-violation findings, not style; the four-genre convergence at 16.1% (3_extract E1/E2)	Rewrites the §3 table and the §3.1 before/after diagram; universal checklist becomes 11 rows, not 10; DoD-3 and DoD-4 must be reworded to match	(a) keep the HL's three rows — loses ~65% of the S1 signal, violates the HL's own DoF-1; (b) keep the axis (C4) — the HL's own §10 filter indicates this, but it retains a gate with 0 verdict flips
A2	§3 / §4 deliverables	EXTEND	Add a deliverable: migrate the three orphaned docs/spec verify actions into verify.md as unconditional actions, or decline each with a stated reason	H2 is asserted from code's two actions only; the three docs/spec actions have no unconditional home (3_extract E1)	One added deliverable and ~3 lines in verify.md. Without it, DoF-1 is violated by construction	(a) treat as covered by promoted judge rows — false: these are Verify-stage actions; (b) decline explicitly — permitted by DoD-4's own grammar, but must be written, not implied
A3	§5 DoD-3	SUPERSEDE	Name the four corrected promoted rows instead of the three current ones, and require the explicit-N/A grammar to be structural (a skipped row visibly marked, not silently ✅)	D6, D12; F21; composite-dilution and 5-9-band evidence (2_gather G7)	Reworded acceptance criterion; a template-grammar requirement the executor must satisfy	Leave DoD-3 as is — it would then accept a promotion set the measurement contradicts
A4	§5 DoD-4	EXTEND	Extend the "every removed row accounted for" requirement to the verify actions as well as the checklist rows	A2's evidence	One clause	Rely on DoF-1 alone — DoF-1 is a failure condition, not an acceptance test; DoD-4 is where the accounting is actually enforced
A5	§5	EXTEND	Add a DoD item for the removed-key gap: CHANGELOG ### Removed must name the key (not only files), and update.md Step 3 must extend 🔴 Breaking to removed config keys	update.md categorises files only; a removed key falls through its triage (3_extract E5)	~4 lines in update.md, one CHANGELOG convention. Generalises past this task	Task-local migration note only — cheaper, leaves the framework gap for the next key removal
A6	§6 DoF-2	RESTRICT	Sharpen DoF-2 from "a row that cannot produce a finding" to "a promoted row whose firing rate is not evidenced, or a promoted set that pushes the universal checklist past the point where rows are read rather than used"	The corpus now provides per-row firing rates, so "cannot produce a finding" is testable rather than rhetorical; the 5-9 band and dilution evidence make set size a real failure mode	None — narrowing. Per conventions.md §3.10 a RESTRICT applies on filing	Leave DoF-2 as prose — it would then be satisfiable by assertion, which is what let §3's table through

A1 is the load-bearing proposal. A2-A6 are consequential on it. If the owner declines A1 and keeps §3 as frozen, the task remains executable — but it ships a promotion set that this measurement contradicts, and DoF-1 ("a check disappears without a recorded home") is triggered on the day it lands rather than discovered later.

Fact Candidates

#	Category	Candidate	Source	Confidence
FC1	environment	Three TFW installs are reachable from this machine and usable as measurement corpora: ai-first-devices (149 REVIEWs, TFW 0.9.0), helpdesk (70 REVIEWs, TFW 0.8.7), this repo (61, 1.0.0). The owner authorised read-only use of both externals for cross-project base-rate measurement	User, briefing gate 2026-08-13	★★★
FC2	process	A .tfw/VERSION value does not track actual framework-file drift: AFD reads 0.9.0 while its entire review surface is byte-identical to 1.0.0. Cross-project replication must diff files, not compare version stamps	2_gather G1	★★★
FC3	convention	Reviewers repurpose the single free-text slot in the REVIEW header to record verification depth (full mode — §6 guardrail, abbreviated, Round 3, 89,6% LOC-budget) — 8 of 13 qualifier instances. The template gives them nowhere else to declare how hard they looked	2_gather G6	★★☆
FC4	process	.tfw/workflows/update.md triages at file granularity only; a removed config key is invisible to it. Applies to every future framework key removal, not just this one	3_extract E5	★★★
FC5	risk	tfw.review.min_verify_ratio sits inside the tfw.review block that update.md marks framework → update, so a project that tuned 0.42 loses the tuning on any upgrade. Pre-existing; not caused by this task	2_gather G5	★★☆
FC6	process	Review verdicts in the corpora are never driven by a single checklist row: in 203 mode-carrying reviews, no REVISE had a mode row as its sole non-✅, and every such REVISE carried a median of 4 failing universal rows	2_gather G4	★★★

Strategic Insights (Research)

#	Category	Insight	Source	Confidence
SS1	process	Owner overrode min_iterations: 2 down to 1 on the argument that four of six hypotheses were settled by measurement rather than judgement. The argument held — but the measurement inverted the HL's premise instead of confirming it, which is the outcome an iteration override is least equipped to absorb. The lesson is not "don't override"; it is that an override is safe when measurement is cheap, and the cost of being wrong lands on the contract, not on the research	User, iterations.yaml 2026-08-13	★★★
SS2	philosophy	Owner delegated the H6 method to the researcher ("dont know you decide") — and the right answer turned out to be reporting that the test could not be run. A delegated methodological choice whose honest output is "unavailable" is only possible where a null and an absence are distinguished. TFW's F21 (explicit N/A over silent skip) is the same principle applied one level up, to research method rather than to checklist rows	User, briefing gate 2026-08-13	★★★
SS3	process	Owner volunteered the second corpus ("also you could use helpdesk project") after seeing only the reconnaissance counts. That instinct is what produced the independent replication: helpdesk's 14.2% is measured on a 6-row universal checklist (no Evidence-completeness row), so it tests the mode rows against a weaker universal set and still finds them productive. A single external corpus would have left AFD's numbers attributable to one project's reviewing culture	User, 2026-08-13	★★★
SS4	philosophy	The owner's framing in HL S1 — «что проверять задается рамкой задачи» — is a design argument and survives the measurement untouched. What the measurement refutes is the empirical argument stacked on top of it ("and besides, the rows never fire"). The strategic lesson: when a design argument is already sufficient, adding a weak empirical prop makes the whole case falsifiable at the weakest joint. The HL would be in a stronger position today with S1 alone	User, HL §11 S1 · 2026-08-13	★★☆

Findings Map

Root-cause chain — why the HL's premise failed

HL §2 claim: "the axis has never produced a finding"
  │
  ├── measured on 38 rows / 18 reviews ── ALL from this repository
  │     │
  │     ├── this repo is markdown-only ......... no code to break, no tests to miss
  │     ├── default_mode: code is WRONG here ... HL says so itself, §2
  │     └── so 8 code-genre rows were scored ... 3 N/A + 5 ✅  ← the "0 findings"
  │
  └── generalised to "the mechanism cannot fire"
        │
        └── TESTED on 598 more rows in 2 software repos
              │
              ├── AFD      408 rows → 20 ❌ + 18 ⚠️  (9.3% raw / 8.3% hard)
              ├── helpdesk 190 rows →  0 ❌ + 27 ⚠️  (14.2% raw / 7.9% hard)
              └── universal-row baseline, same reviews: 8.4% raw / 8.3% hard
                    │
                    └── VERDICT: mode rows ≈ universal rows ≈ 8%
                         "never fires" was a property of the SAMPLE

The convergence nobody proposed — four rows, one check

        genre-specific name              what it actually found
  ┌───────────────────────────┬──────────────────────────────────────────┐
  │ code:  Test coverage      │ "suites green, but the acceptance         │
  │        23.4% ── 141 rows  │  contract passes with a forbidden         │
  │                           │  production collector"                    │
  ├───────────────────────────┼──────────────────────────────────────────┤
  │ spec:  Analytical quality │ "собственные completeness gates           │
  │        25.0% ──   8 rows  │  отмечены зелёными при невыполнении"      │
  ├───────────────────────────┼──────────────────────────────────────────┤
  │ spec:  Source attribution │ "один primary-source claim неверен,       │
  │        22.2% ──   9 rows  │  восемь source bindings отсутствуют"      │
  ├───────────────────────────┼──────────────────────────────────────────┤
  │ docs:  Source verification│ "migration and changeset source           │
  │        12.5% ──  16 rows  │  checks fail"                             │
  └───────────────────────────┴──────────────────────────────────────────┘
                    │
                    ▼
     ONE CHECK:  "the artifact carries a green signal,
                  and the green signal does not establish the claim"
                    │
     combined:   28 non-✅ / 174 rows = 16.1%
                 ── the highest-firing check in TFW review ──
                    │
     HL promotes: "Claims traceable to sources"  ← docs/spec fragments only
                   captures ~35% of the signal, drops the code instance (141 of 174 rows)

Configuration survival

                        [D1](../../../../knowledge-index.md#architecture-decisions)      [D2](../../../../knowledge-index.md#architecture-decisions)      [D3](../../../../knowledge-index.md#architecture-decisions)      [D4](../../../../knowledge-index.md#architecture-decisions)      [D5](../../../../knowledge-index.md#architecture-decisions)
                     substance firing priming consumers ext
  C1 HL as frozen ......  ✗       ✗      ?       ✓       ✓   ELIMINATED
  C2 delete + 4 rows ...  ✓       ✓      ?       ✓       ✓   survives
  C3 C2 + verify actions  ✓       ✓      ?       ✓       ✓   SURVIVES — strongest
  C4 project-optional ...  =       =      ✓       ~       ✓   survives (HL's own §10 filter)
  C5 non-gated descriptor  —       —      ✓       ✓       ✓   rider on C2/C3 only
  C6 extend enum .......  ✗       ?      ✓       ✓       ✗   ELIMINATED (F13)
  C7 genre → rigour ....   —       —      ✓       ~       ✓   DEFERRED (right signal, wrong task)

Priority matrix for the coordinator

	Low cost to fix	High cost to fix
High impact if ignored	A2 orphaned verify actions · A5 removed-key rule	A1 corrected promotion set (rewrites frozen §3)
Low impact if ignored	A6 DoF-2 sharpening · R10 VERSION observation	Q9 S3-vs-U2 placement · C7 rigour axis (sibling task)

Iteration Status

• Iteration: 1 of 1 (min, coordinator override) / 3 (max)
• Hypotheses tested: H1 ❌ refuted-as-stated · H2 🟡 partial · H3 ❌ refuted · H4 🟡 split · H5 ✅ confirmed · H6 🔴 unresolved
• Hypotheses deferred: H6 — the observational test is unavailable in this corpus (no control group exists); closing it requires an experiment, not a measurement. Its decision-relevance for the C1-vs-C3 choice has collapsed, so deferring it does not block the task
• Gaps discovered:
• Three docs/spec verify actions have no unconditional home — invisible to the HL's judge-row-only analysis
• update.md has no removed-key rule at any granularity
• Q9 — S3 Design soundness is unresolved between "new row" and "sharpen U2 Philosophy aligned"; 6 hard ❌ depend on the answer
• The promotion target (10-11 flat rows) exceeds the 5-9 checklist band and invites composite dilution — a design problem the HL does not address
• Q10 non-software domains remain untestable — no analytics/curriculum/business-process corpus is in reach. This was a named HL blind spot at the start and it is still open
• Superseded decisions: D13 supersedes Gather's first headline (mode rows more productive than universal rows → equally productive, ~8%). No prior-iteration decisions exist to supersede

Open Threads (for next iteration)

#	Thread	Why it matters	Suggested focus
1	H6 priming — unresolved, no observational test available	Only decides whether C5 (a non-gated descriptor) rides along as insurance. Cheap to adopt, cheap to skip	Not a research question any more — it is an experiment (re-review one RF with and without the label) or an owner judgement call. Do not spend an iteration on measurement that cannot exist
2	Q9 — S3 placement (own row vs sharpened U2)	6 hard ❌ lose their home if neither is chosen. Silence is the failure mode DoF-1 names	Resolvable at TS time by reading the 6 findings against U2's current wording. ~30 minutes, no new corpus needed
3	Q10 — non-software domains	F13 claims domain-agnosticism; all evidence is from software + one markdown framework repo	No corpus exists. Recommend the coordinator record it as a standing limitation rather than schedule an iteration that cannot gather data
4	C7 — genre→rigour axis	8 of 13 field qualifiers encode verification depth; the field is being repurposed. Real signal about what reviewers need	Sibling task or HL amendment after TFW-56 lands. Explicitly out of scope here — it touches min_verify_ratio (DoF-4)

Recommendation

• [x] SUFFICIENT — proceed to /tfw-plan, but the coordinator cannot proceed straight to TS. A1-A6 target frozen sections. Per conventions.md §3.3/§3.8 they must be transcribed into HL §12 as PROPOSED and ruled by an explicit owner verdict before any TS is written against §3/§5/§6.
• [ ] MORE NEEDED
• [ ] BLOCKED

A second iteration would add little: of the four open threads, one is an experiment rather than a measurement (H6), one is a 30-minute reading task at TS time (Q9), one has no reachable data (Q10), and one is out of scope by design (C7). The binding constraint now is an owner ruling on A1, not more evidence.

⚠️ Coordinator decides whether to continue or proceed. Researcher recommends but does NOT decide.

Conclusion

This iteration set out to replicate a base-rate finding outside its home repository and instead inverted it. Across 203 mode-carrying reviews and 637 mode-specific rows in three TFW installs — a corpus 16× the one the HL reasons from — the code/docs/spec checklist rows fire at roughly 8%, statistically indistinguishable from the universal rows they were to be folded into, and every one of the eight fired somewhere. The HL's "never produced a finding" turns out to be a property of a 39-row markdown-only sample scored under a default_mode: code the HL itself calls wrong for this project. What survived the measurement is substantial and I want to be precise that the task is not dead: the 🛑 gate is genuinely unjustified (0 verdict flips in 203 reviews), the three mode files are byte-identical across three installs and duplicate verify.md in their first action, docs and spec really are synonyms, the key is wrong-by-default and routed by three stale pointers, and H5 is closed outright. The deletion is defensible on cost and design; it is the empirical prop — and the frozen §3 coverage table resting on it — that does not hold. Research's specific contribution beyond the numbers is E2: read by what they found rather than by what they are called, four of the eight rows are one check in three genre costumes — does the evidence offered actually establish the claim — firing at 16.1%, the most productive check TFW review has, and the HL's promoted wording would have carried about a third of it while the rest disappeared into a deletion diff that looked like simplification. Self-critique: the duplication figure (62 of 65 findings with no other home) is lexical, not semantic, and overstates uniqueness — it is an upper bound and I have marked it as one; the ⚠️ classification proved softer than ❌ in a way I did not anticipate and had to correct mid-research (D13), which is recorded rather than smoothed over; H6 was the hypothesis the HL called most likely to be refuted and I could not test it at all, because the control group I expected turned out to be a different kind of document rather than the same document minus a label. Reporting that as unavailable rather than as a null is the single most important thing in this file that is not a number.

---

RES — TFW-56: Remove the Review Mode Axis — Iteration 1 | 2026-08-13