EV — TFW-56: Remove the Review Mode Axis¶

Date: 2026-08-13 Author: Executor (Claude Code) Task: TFW-56 TS: TS TFW-56

Environment¶

Field	Value
OS	Windows 11 Pro 10.0.26200
Language / Runtime	Python 3.13.5
Shells	Git Bash (MSYS2), git 2.42.0.windows.1
Build	`python -m pytest docs/scripts/` (68 tests) — the docs pipeline is the only build that consumes the changed files
CI / Pipeline	`.github/workflows/docs.yml` — run locally here

project_config.yaml build.lint/test/verify are unconfigured starter placeholders (echo "configure your … command"). conventions.md, glossary.md, .tfw/workflows/** and .tfw/templates/** are Source Manifest rows 4, 5, 12 and 13, so the docs build is the substitution — the same substitution used and approved in EV TFW-53/A.

Evidence¶

#	AC	What was verified	Environment	Result	Artifact
E1	AC-1	The mode-file folder is gone and `review.md`'s steps are contiguous with Step 0 = Session Naming	Git Bash	VERIFIED	§E1 — `ls` fails; 8 headings, 0-7, no gaps
E2	AC-2	Judge checklist is exactly ten universal rows, no `Mode:` field, no Mode-Specific section	Git Bash	VERIFIED	§E2 — row count 10, `grep -c "Mode"` → 0
E3	AC-2	S1-vs-U7 dry-run against an archived RF/EV/REVIEW triple — do rows 7 and 8 produce different answers?	Reading, against TFW-53 Phase A	VERIFIED	§E3 — row 7 ✅, row 8 ❌ with three findings. Different answers, different reasoning
E4	AC-3	Eight mode checklist rows each accounted for	Reading	N/A	TS Evidence field: `N/A — this is a documentation-completeness check, verifiable by reading.` Table in RF §3 / §2
E5	AC-4	Eight distinct verify actions each accounted for; `map.md` and `verify.md` carry no `Mode:`	Reading + grep	N/A	TS Evidence field: `N/A — file content, verifiable by reading.` Table in RF §2; grep result in §E6 below
E6	AC-5	`REVIEW.md` §3 matches `judge.md` row-for-row, same order; no header field, no placeholder	Reading	N/A	TS Evidence field: `N/A — template content.` §E4 carries the row-list diff
E7	AC-6	`default_mode` absent from both config files; `min_verify_ratio: 0.42` and its comment intact	Git Bash	VERIFIED	§E5 — both file excerpts, before and after
E8	AC-7	`config.md` no longer routes `review.default_mode`; the `min_verify_ratio` row names the correct step	Git Bash	N/A	TS Evidence field: `N/A — cross-reference check.` Resolution recorded in §E6
E9	AC-8	No mode vocabulary in `conventions.md` / `glossary.md`; anti-pattern present	Git Bash	N/A	TS Evidence field: `N/A — text content.` `grep -rn "review mode" .tfw/` → 0, recorded in §E6
E10	AC-9	Six adapter and entry-point copies carry no mode reference and match their sources	Git Bash	VERIFIED	§E6 — grep → 0 matches across all six; five `diff` runs, all empty
E11	AC-10	`VERSION` bumped; CHANGELOG `### Removed` names the key	Git Bash	N/A	TS Evidence field: `N/A — changelog content.` `1.1.0`; key named at CHANGELOG L21
E12	AC-11	TD-106 closed with the reason	Reading	N/A	TS Evidence field: `N/A`
E13	AC-12	The grep gate, run verbatim, with its output	Git Bash	VERIFIED	§E7 — zero matches, exit 1. Plus a supplementary sweep, because one quarter of the mandated pattern is dead
E14	AC-12	History intact: no existing task REVIEW file modified, no past CHANGELOG entry edited	Git Bash	VERIFIED	§E8 — 41 files under `tasks/` still carry `Review Mode`; nothing under `tasks/` in the diff except this task's own artifacts
E15	AC-12	Diffstat: 3 deletions, no framework file created	Git Bash	VERIFIED — with an anomaly that must be read	§E9 — the substance holds; the three deletions were swept into a concurrent session's commit, so they do not appear in this task's own diffstat
E16	Build gate	Docs pipeline tests, since the changed files are Source Manifest rows	Python 3.13.5	VERIFIED	§E10 — 68 passed

Verdict¶

Evidence verdict: 9/16 VERIFIED, 0 DEFERRED, 0 BLOCKED, 7 N/A

The 7 N/A are the TS's own Evidence: fields, quoted verbatim above — not executor judgement. E15 is VERIFIED on substance with a trace anomaly recorded in full; it is not marked N/A or DEFERRED, because the fact being evidenced is true and checkable — it is the attribution of the change that is wrong, and hiding that behind a softer status would be the failure this layer exists to prevent.

E1 — The selection is absent (AC-1)¶

$ ls .tfw/workflows/review/
ls: cannot access '.tfw/workflows/review/': No such file or directory

$ grep -n "^## Step" .tfw/workflows/review.md
16:## Step 0: Name This Session
52:## Step 1: Map
60:## Step 2: Verify
78:## Step 3: Judge
89:## Step 4: Decide (Synthesize → REVIEW)
100:## Step 5: Tech Debt Collection
109:## Step 6: Update Traces
116:## Step 7: Knowledge Capture (KNW)

Contiguous 0-7, and Step 0 is Session Naming — the TFW-standard opening this file has never had (knowledge/process.md F19, TD-106).

The only surviving occurrence of "mode" or "WAIT" in the whole file:

$ grep -ni "mode\|WAIT" .tfw/workflows/review.md
7:> **Role:** Reviewer (coordinator in review-locked mode)

"review-locked mode" is the Role Lock, not a genre. It is invisible to the AC-8 gate (grep -rn "review mode") because the words are not adjacent, and it means one thing.

Internal step references, read end to end. Three references pointed at step numbers. All three were wrong before this task and are correct after it, with no edit:

Reference	Says	Verify was	Verify is	Before	After
`review.md` L143 anti-pattern	"Step 2 (Verify)"	Step 3	Step 2	❌ stale	✅
`conventions.md` L495 anti-pattern	"Step 2 (Verify)"	Step 3	Step 2	❌ stale	✅
`config.md` L92 propagation row	"Step 2: Verify"	Step 3	Step 2	❌ stale	✅

One reference needed an actual edit, and it pointed at Judge rather than Verify:

Reference	Was	Now	Why
`glossary.md` Principles Check	"`review.md` Step 4"	"`review.md` Step 3"	Judge moved 4 → 3

glossary.md L157 (Session Naming → review.md Step 0) needed nothing and becomes true for the first time.

E2 — Ten universal rows (AC-2)¶

$ grep -c "^| [0-9]* |" .tfw/templates/review/judge.md
10

$ grep -c "Mode" .tfw/templates/review/judge.md
0

#	Row	Rate carried in the row
1	DoD met?	—
2	Philosophy aligned — (a) mapping integrity · (b) design soundness	4.5% on (b)
3	Tech debt documented	—
4	Style & standards	—
5	Observations collected	—
6	RF completeness (§7-9)	—
7	Evidence completeness — does the evidence exist?	—
8	Evidence sufficiency — does the evidence establish the claim?	16.1%
9	Backward compatibility	8.5%
10	Safety — kept on consequence, not on rate	4.0%

Order is HL §3.1's frozen after-diagram. Content quality is absent — the one true duplicate of row 4.

Explicit-N/A is structural, not advisory. Status vocabulary on every row is ✅/❌/⚪; the section header states that ⚪ N/A requires a stated reason and that a row skipped as a bare ✅ leaves the stage incomplete; and the Checkpoint carries - [ ] Every ⚪ N/A carries a stated reason — no row skipped as a bare ✅?. Three sites, so deleting any one leaves the rule visible.

E3 — The S1-vs-U7 dry-run (AC-2)¶

Subject: TFW-53 Phase A — RF · EV · REVIEW (✅ APPROVE, three passes, Review Mode: spec).

Chosen because it is this repository's most recent completed review, it has a real 15-row EV file with resolving attachments, and it was scored under an owner mode override — so it is the hardest available test of whether the new row adds anything the old checklist missed.

This is a dry-run of a checklist, not a re-review. I am TFW-56's executor, not TFW-53's reviewer. Nothing here revises TFW-53/A's APPROVE verdict, and a row-8 ❌ in a dry-run is not a claim that the verdict should have been REVISE — that verdict rested on the whole picture, and row 8 did not exist when it was made.

Row 7 — Evidence completeness. Does the evidence exist? → ✅

All 15 TS ACs appear in the EV table. Only the 4-status vocabulary is used (7 VERIFIED, 0 DEFERRED, 0 BLOCKED, 8 N/A). Both attachments resolve — baseline_recovery.txt and classification_exercise.md are present in phase-a/evidence/. The 8 N/A quote the TS's own Evidence: fields verbatim rather than asserting executor judgement. Same answer the archived review reached on its row 7.

Row 8 — Evidence sufficiency. Does the evidence establish the claim? → ❌, three findings

#	The green signal	What it does not establish
1	E2 marked VERIFIED, concluding "the amended template can carry HL-TFW-53's own §12 unchanged"	E2's own table records `Type` = `RESTRICT` as "not exercised by this corpus… recorded as a coverage gap". One of the shipped enum's values has zero live coverage. The exhibit establishes the narrow claim; the verdict line `7/15 VERIFIED` reads as the broad one
2	E13 marked VERIFIED, shipping an anchored `--grep` recovery form	Its conclusion was superseded inside the same pass by E15: "no `--grep` form can be subject-only". The measurements stand; the conclusion does not. A VERIFIED exhibit whose conclusion has been retracted is the exact shape row 8 exists to catch
3	E11 scored 4/4 — pass	The exhibit states the earlier budget and cut-order property is "absent from both the rule and this check" — removed by amendment A12 and deliberately not implemented. 4/4 is against a rule with a property removed, and the score does not carry that

Result: rows 7 and 8 produced different answers, from different reasoning. AC-2 is met.

Two properties of this outcome matter more than the ❌ itself:

Row 8's findings are not absences, so row 7 structurally cannot see them. Every one is a present, VERIFIED artifact whose stated scope is narrower than the claim it is offered for. Row 7 asks whether the row is filled; all three rows were filled, correctly, in the 4-status vocabulary.
All three findings were written by TFW-53/A's own executor, inside the EV file, and none is surfaced by its verdict line. That is the case for the row: the honesty was already in the trace, and no checklist row was asking the question that would promote it into the verdict. Under the old checklist this was reachable only through spec mode — and TFW-53/A was reviewed as spec, scored both mode rows ✅, and did not surface them.

E4 — REVIEW.md §3 against judge.md (AC-5)¶

Row lists compared in order. REVIEW.md §3 carries condensed labels by design — it is a synthesis document, not a second copy of the stage file — so the check is one-for-one on identity and order:

#	`judge.md`	`REVIEW.md` §3	Match
1	DoD met?	DoD met? (all TS acceptance criteria)	✅
2	Philosophy aligned — mapping integrity + design soundness	Philosophy aligned — mapping integrity + design soundness	✅
3	Tech debt documented	Tech debt documented	✅
4	Style & standards	Style & standards	✅
5	Observations collected	Observations collected	✅
6	RF completeness (§7-9)	RF completeness (§7-9 present)	✅
7	Evidence completeness — does it exist?	Evidence completeness — does it exist?	✅
8	Evidence sufficiency — does it establish the claim?	Evidence sufficiency — does it establish the claim?	✅
9	Backward compatibility	Backward compatibility	✅
10	Safety	Safety	✅

10/10, same order. The  comment and the > **Review Mode**: header field are both gone.

Repaired in passing: REVIEW.md §3 had 6 rows against judge.md's 7 — the Evidence completeness row added in 0.8.8 never reached it. Closing that gap is required by AC-5 and is not scope creep; it is recorded in RF §2 as a decision.

E5 — The config key, and its sibling (AC-6)¶

Before — .tfw/project_config.yaml:

  review:
    default_mode: code        # code / docs / spec — determines review checklist
    min_verify_ratio: 0.42    # minimum fraction of changed files to verify (escalate to 1.0 on discrepancy)

After — grep -n "review:" -A4 .tfw/project_config.yaml:

59:  review:
60-    min_verify_ratio: 0.42    # minimum fraction of changed files to verify (escalate to 1.0 on discrepancy)

After — grep -n "review:" -A4 .tfw/templates/project_config.yaml:

63:  review:                             # ← FRAMEWORK: updated by tfw-update
64-    min_verify_ratio: 0.42

min_verify_ratio keeps its value, its inline comment byte-for-byte, and its # ← FRAMEWORK annotation in the template. Its behavioural home — review.md Step 2's parameter table with Default 0.42 | Hard | min_verify_ratio and the ⌈5 × 0.42⌉ = 3 worked example — is unchanged. This is the one place where collateral damage would have been silent, which is why AC-6 asked for both excerpts rather than an assertion.

E6 — Propagation, vocabulary and adapter parity (AC-7, AC-8, AC-9)¶

$ grep -n "review\." .tfw/workflows/config.md
92:| `review.min_verify_ratio` | [`.tfw/workflows/review.md`](../../../reference/workflows/review.md) | Step 2: Verify | Min verify ratio |

One row, and its pointer resolves: review.md:60 is ## Step 2: Verify. The review.default_mode row is gone with its key.

$ grep -rn "review mode" .tfw/ --exclude=CHANGELOG.md
(no matches — exit 1)

$ grep -rniE "review mode|review-mode|default_mode: code|code / docs / spec|mode-aware|mode file" \
       .claude/commands/tfw-review.md .claude/commands/tfw-config.md \
       .agent/workflows/tfw-review.md .agent/workflows/tfw-config.md \
       .tfw/adapters/codex/skills/tfw-review/SKILL.md .agents/skills/tfw-review/SKILL.md
(no matches — exit 1)

Behavioural parity, five diff runs, all empty:

Copy	Source	Result
`.claude/commands/tfw-review.md`	`.tfw/workflows/review.md`	identical
`.agent/workflows/tfw-review.md`	`.tfw/workflows/review.md`	identical
`.claude/commands/tfw-config.md`	`.tfw/workflows/config.md`	identical
`.agent/workflows/tfw-config.md`	`.tfw/workflows/config.md`	identical
`.agents/skills/tfw-review/SKILL.md`	`.tfw/adapters/codex/skills/tfw-review/SKILL.md`	identical

The four full copies were byte-identical to their sources before this task as well (checked at onboarding), so "nothing project-specific was clobbered" is established against a clean baseline rather than assumed. The Codex skill line "Follow the review-mode WAIT gate…" now reads "Follow every gate in the workflow exactly as it requires, including each stage self-check gate before advancing" — the four self-check gates are what actually remains to be followed.

map.md and verify.md carry no Mode: field:

$ grep -n "Mode" .tfw/templates/review/map.md .tfw/templates/review/verify.md
(no matches)

E7 — The grep gate, verbatim (AC-12)¶

$ grep -rn "code / docs / spec\|default_mode: code\|Review Mode\|review/{code" \
       .tfw/ .claude/ .agent/ .agents/ --exclude=CHANGELOG.md
$ echo $?
1

Zero matches. Recorded with its exit status because a silent pass and an empty output look the same on the page.

A finding about the gate itself, not about the sweep. One of the four alternatives, review/{code, matched nothing in this repository before any change was made — the real string was always .tfw/workflows/review/{mode}.md. A quarter of the acceptance test could never fail, which is precisely the anti-pattern this task adds to conventions.md §14. Two narrower holes: the pattern is case-sensitive, so glossary.md's lowercase "coordinator in review mode" was invisible to it, and config.md's review.default_mode registry row does not match default_mode: code. AC-8 and AC-9 cover all three places by other means, so the ACs together are sound — the single command is not. Supplementary sweep, wider and case-insensitive:

$ grep -rniE "review mode|review-mode|mode-aware|mode-specific|default_mode: code|review/\{(code|mode)|Mode: \{code" \
       .tfw/ .claude/ .agent/ .agents/ --exclude=CHANGELOG.md
.tfw/workflows/research/base.md:79:  Mode-specific: ☐ {from mode file}
.tfw/workflows/research/deep.md:21:Mode-specific:
.claude/commands/tfw-research.md:79:  Mode-specific: ☐ {from mode file}
.agent/workflows/tfw-research.md:79:  Mode-specific: ☐ {from mode file}

All four hits are the research mode axis (focused / deep), which is out of scope and untouched. No review-mode residue at any case or spelling.

E8 — History is intact (AC-12)¶

$ grep -rl "Review Mode" tasks/ | wc -l
41

41 files under tasks/ still carry the string, including tasks/TFW-53__hl_contract_and_goal_defence/phase-a/REVIEW__phase-a__contract_in_artifacts.md, whose header still reads > **Review Mode**: spec with its owner-override note — the file used as this task's own dry-run subject. Nothing under tasks/ appears in this task's diff except its own artifacts (ONB, this EV file, RF). No past CHANGELOG entry was edited; ## [1.1.0] is inserted above ## [1.0.0], and the ## [Unreleased] heading is left in place.

E9 — Diffstat, and a trace anomaly that has to be reported (AC-12)¶

The substance holds. Three files deleted, zero framework files created, net LOC negative. The only new file is this task's own evidence artifact, which is not a framework file.

The anomaly. The three deletions are not in this task's commit. A concurrent session working on TFW-53 Phase B committed fbdf443 [claude-code/TFW-53/phase-b/executor] enforce the contract in the workflows at 17:44, and that commit swept up my already-staged git rm of the mode files:

$ git show --stat fbdf443
 .tfw/templates/RES.md           |  2 +-      ← TFW-53/B's own work
 .tfw/workflows/plan.md          | 82 +++---   ← TFW-53/B's own work
 .tfw/workflows/research/base.md |  6 ++-      ← TFW-53/B's own work
 .tfw/workflows/review/code.md   | 15 ------   ← [TFW-56](../HL-TFW-56__review_mode_removal.md)'s deliverable
 .tfw/workflows/review/docs.md   | 12 ------   ← [TFW-56](../HL-TFW-56__review_mode_removal.md)'s deliverable
 .tfw/workflows/review/spec.md   | 12 ------   ← [TFW-56](../HL-TFW-56__review_mode_removal.md)'s deliverable

Scope of the leak, bounded by timing: at 17:44 the only TFW-56 change staged in the index was that git rm. Every other file in this task was still an unstaged working-tree edit and was not captured. So the leak is exactly three deletions, and they are the three that were meant to be deleted.

Consequence, stated plainly. The outcome AC-1 and AC-12 require is real and verifiable — the folder is gone from the working tree and from HEAD. What is wrong is the attribution: recovering "when did TFW-56 delete the mode files" by reading TFW-56's commits will find nothing, and the deletion is filed under a task whose own TS says review.md and its neighbourhood are "Phase C only". No attempt was made to repair this by rewriting history — fbdf443 is another session's commit and rewriting it is not an executor's call. It is recorded here, in RF §2 and in RF §6.

This task's own diffstat, staged by explicit path only — never git add -A, which is how the leak became possible in the first place:

 .agent/workflows/tfw-config.md                  |  1 -
 .agent/workflows/tfw-review.md                  | 27 +++-------
 .agents/skills/tfw-review/SKILL.md              |  2 +-
 .claude/commands/tfw-config.md                  |  1 -
 .claude/commands/tfw-review.md                  | 27 +++-------
 .tfw/CHANGELOG.md                               | 19 +++++++
 .tfw/VERSION                                    |  2 +-
 .tfw/adapters/codex/skills/tfw-review/SKILL.md  |  2 +-
 .tfw/conventions.md                             |  1 +
 .tfw/glossary.md                                |  6 ++--
 .tfw/project_config.yaml                        |  3 +-
 .tfw/templates/REVIEW.md                        | 21 +++++---
 .tfw/templates/project_config.yaml              |  1 -
 .tfw/templates/review/judge.md                  | 35 ++++++++-----
 .tfw/templates/review/map.md                    |  1 -
 .tfw/templates/review/verify.md                 | 16 +++++-
 .tfw/workflows/config.md                         |  1 -
 .tfw/workflows/review.md                        | 27 +++-------
 README.md                                       |  2 +-
 TECH_DEBT.md                                    |  2 +-
 20 files changed, 105 insertions(+), 94 deletions(-)

20 modified + 3 deleted (in fbdf443) = 23 framework and root files touched, one more than the TS's 22 because .tfw/project_config.yaml carries two changes rather than one — the key removal and the version field, which tracks VERSION on every release (RF §2). Against the project budget of 30 files / 15 new / 3000 LOC / 30 modified: within budget on every axis, net LOC −94 + 105 = +11 in the modified set and −39 once the three deleted files are counted.

E10 — Build gate (Step 10)¶

$ python -m pytest docs/scripts/ -q
....................................................................    [100%]
68 passed in 33.47s

The docs pipeline is the consumer of every changed .tfw/ file — conventions.md, glossary.md, workflows/** and templates/** are Source Manifest rows 4, 5, 12 and 13. Deleting three workflows/** files removes three generated pages; the suite passes, and gen_docs.py carries no reference to review modes (grep -rn "review" docs/scripts/*.py → no matches), confirming RES D9's consumer audit against the shipped state rather than against the pre-change state.

Attachments¶

No binary artifacts. Every exhibit above is a command with its output, or a reading against a named file and line.

EV — TFW-56: Remove the Review Mode Axis | 2026-08-13