Skip to content

TS — TFW-53 / Phase C: Goal Defence in Review

Date: 2026-08-13 Author: Coordinator (Claude Code) Status: 🟡 TS_DRAFT — approved by the owner 2026-08-13, cleared for execution · amended after ONB (Q1 → review.md:85 in scope · Q2 → north-star locus ruled · Q3 → deferral confession ships · AC-12 added for the uncovered DoD-23) · amended after REVIEW 🔄 REVISE — AC-13, AC-14, AC-15 added for the second execution pass, same executor (§5 has no TS_APPROVED token and this task does not invent one — the approval is recorded, the status vocabulary is unchanged.) Parent HL: HL-TFW-53 — 🔒 FROZEN, re-frozen after A1–A14 Covers: frozen DoD 18–29 Predecessor read (Pre-TS gate): RF Phase B, REVIEW Phase B — ✅ APPROVE Sibling read: RF TFW-56 — landed first; it rewrote every file this phase touches

1. Objective

Phases A and B stopped the goals from moving. This phase catches the other case: the goals held, and the work went somewhere else anyway. The Judge's mapping-integrity check — which cannot detect a principle violated by the mapping itself — is replaced by a Purpose Check whose reference set is the committed contract baseline plus a Project North Star, never the TS. A reviewer gains the authority to reject work that is verified, complete, tested and beside the point, and loses the ability to assert alignment without citing the clause it serves.

2. Scope

In Scope

  • templates/review/judge.md — Purpose Check replacing row 2 clause (a); reference set, fused citation-and-harm, excess-and-adjacency test, third outcome, override clause
  • workflows/review.md — context-loading line 28, Reviewer Identity, not fit for purpose verdict routing
  • templates/REVIEW.md — row 2 realignment + Purpose Check finding in the synthesis
  • glossary.md — PV Index priority 0 and the priority 1 relabel
  • conventions.md — Project North Star definition, two §14 anti-patterns
  • templates/HL.md — north-star header field with its fallback chain
  • compilable_contract.md — the NS{n} / PP{n} / P{n} namespace rows
  • Replay validation against TFW-48/49 and three sound reviews

Out of Scope

  • Adapter and entry-point sync — Phase D deliverable 3 owns it. Do not touch .claude/, .agent/, .agents/ or .tfw/adapters/
  • Glossary articles for Project North Star, Purpose Check, not fit for purpose, deferral confession — Phase D deliverable 1. This phase adds the PV Index rows, not the term articles
  • Writing this repository's north-star designation into either README. The owner has ruled what it is — the philosophy in the root README.md and the founder essay .tfw/README.md (HL §11 S38) — and AC-6/AC-7 build for it. But a TFW-55 session is restructuring both files right now, so a section-level pointer landed today is stale on arrival. Phase C ships the slot and the grammar; the pointer lands with TFW-55 or with the next HL written here. The fallback (master HL §1 at the frozen baseline) must therefore still work on day one
  • Version bump, CHANGELOG — Phase D
  • Any part of the AT execution mode (HL DoF-4)

3. Principles Check

# Principle (HL §7) Enforced by Gate
P3 Structural enforcement over guidelines AC-1, AC-2 The check lives in the template that is filled every review, not in review.md prose
P8 Tool-agnostic by behavior AC-1, AC-6 No vendor mechanism named; the anchor is a file and a PV row, not a memory layer
P13 Purpose is a distinct question, judged where verdicts are formed AC-1 One row inside Judge, no fifth stage, no new document
P14 Every gate needs a materiality bar AC-2 The harm clause is mandatory; a phrasing objection cannot satisfy it
P15 Alignment must be cited, not asserted AC-2 A ✅ with no quoted clause fails the row
P16 Judge against the baseline, never the spec AC-3, AC-4 TS and Phase HL named as invalid references in both files
P9 Naming creates behavior AC-4, AC-9 not fit for purpose is the finding name; NS{n} cannot collide with P{n}
P7 Token density AC-5 review.md word ledger
P2, P4, P5, P6, P10, P11, P12 contract/amendment mechanics N/A Discharged in Phases A and B; no review-side surface
P1, P17 contract earns autonomy; failed trace N/A P1 is the task-level thesis, not a phase deliverable; P17 is Phase E

4. Affected Files

File Action Description
.tfw/templates/review/judge.md MODIFY Row 2 clause (a) replaced by the Purpose Check; the three tests and the override clause added below the table; one Checkpoint item
.tfw/workflows/review.md MODIFY Line 28 context row, Reviewer Identity block, verdict routing for not fit for purpose, and the HL §7 Principles check paragraph in Step 3 — that paragraph and nothing else in that step (added 2026-08-13 on ONB Q1: it is the second live statement of the mapping-integrity check, and leaving it ships DoF-11)
.tfw/templates/REVIEW.md MODIFY §3 row 2 wording realigned to judge.md; Purpose Check finding surfaced in §4 Verdict
.tfw/glossary.md MODIFY PV Index gains priority 0; priority 1 relabelled
.tfw/conventions.md MODIFY Project North Star defined (§3); two §14 anti-patterns
.tfw/templates/HL.md MODIFY Header block gains the north-star field and its fallback chain
.tfw/compilable_contract.md MODIFY One table, reference patterns only. P{N} row corrected — it still points at KNOWLEDGE.md §0, removed by D37; NS{N} and PP{N} rows added
phase-c/evidence/EV__phase-c__goal_defence_in_review.md CREATE Structured evidence, per-AC table, verdict
phase-c/evidence/purpose_check_replay.md CREATE AC-11 — the replay corpus, per-review outcome, and the derived base rate

Budget: 2 new, 7 modified. Limits: 30 files, 15 new, 3000 LOC, 30 modified. Net LOC small and additive.

5. Acceptance Criteria

AC-1: The Purpose Check replaces clause (a) of row 2, and clause (b) survives intact

judge.md row 2 currently carries two separately answered clauses: (a) mapping integrity and (b) design soundness (4.5%). DoD-20 replaces the mapping-integrity check. Clause (b) arrived four hours earlier from TFW-56 carrying six hard ❌ and its own measured rate; replacing the whole row deletes it silently.

  • [ ] Row 2 clause (a) is the Purpose Check: is this what we set out to do? — no mapping-integrity check remains anywhere in judge.md
  • [ ] Clause (b) Design soundness (4.5%) is present, still separately answered, its meaning unchanged
  • [ ] The checklist is still 10 rows. The Purpose Check does not become row 11
  • [ ] The row carries its base rate and the reason it is kept on consequence rather than frequency, in the grammar conventions.md §14 now requires: roughly 4 goal-based blocks in 149 reviews, against a miss that cost six days of work rejected wholesale. No project is named inside the template — judge.md ships to every TFW install and its grammar for provenance is a bare corpus figure (F13) (corrected 2026-08-13, ONB R1)
  • [ ] The mapping-integrity check is also gone from review.md Step 3 — see AC-4. Removing it from the template and leaving it in the workflow ships DoF-11

Gate: read row 2; grep -rc "mapping integrity" .tfw/templates/review/judge.md .tfw/workflows/review.md → 0 in both; row count = 10; clause (b) diffs clean against git show HEAD~:.tfw/templates/review/judge.md Evidence: the before/after of row 2, both clauses quoted in full

AC-2: The check cannot be passed by assertion [depends: AC-1]

Four properties, all in judge.md, all failing conditions rather than advice:

  • [ ] Fused citation-and-harm — one field, one sentence: quote the clause served and name the concrete harm at stake. A citation that resolves but is irrelevant fails. A harm asserted with no citation fails. ✅ with an empty field fails
  • [ ] Excess-and-adjacency — does the result deliver something the cited clause does not ask for, or that a baseline non-goal, a DoF item or a phase boundary excludes?
  • [ ] Deferral confession — does the spec or the result itself name a different home for this work and ship it here anyway? One sentence. (added 2026-08-13, ONB Q3: frozen DoD-30 makes Phase D define the term, and a glossary article with no referent in the framework is a label that resolves and means nothing — S32 one layer up.)
  • [ ] Override clause — "the TS scoped it this way" and "tests are green" are stated as insufficient grounds to ✅
  • [ ] Materiality bar — the harm must be material impact on the value, never phrasing. A wording objection does not satisfy the harm clause (P14; AFD's first firing was a false positive on prose-only rationale)

Gate: fill the row against RF Phase B once, in the evidence file, and show a bare citation with no harm being rejected by the row's own wording Evidence: the dry-run, with the failing variant shown beside the passing one

AC-3: The reference set is stated, and the invalid references are named [depends: AC-1]

  • [ ] judge.md states the reference set: the master HL at its committed frozen baseline plus the Project North Star
  • [ ] The TS and any Phase HL are named as invalid references, with the one-line reason: the TS is downstream of any drift (P16), and a Phase HL holds nothing approved (conventions.md §3, Phase HL is derivation-only)
  • [ ] The fallback chain is stated where the reviewer reads it: project north star → master HL §1 at the frozen baseline. The absence of a north star never blocks a review
  • [ ] Recovering the baseline points at conventions.md §3 rule 15 — the recovery form is not restated (rule 15 exists so it can be corrected in one place; A13 is what happens when it is copied)

Gate: read the block; confirm no second copy of the recovery command exists in .tfw/ Evidence: N/A — textual, verified by reading

AC-4: review.md carries the identity, the reference and the routing

  • [ ] Line 28 reads master HL at its frozen baseline, not "Master HL for the task"
  • [ ] The HL §7 Principles check paragraph in Step 3 is replaced by the Purpose Check instruction naming the reference set. Not deleted and not left: deleting costs the step its only content, leaving it makes the workflow instruct a check its own template no longer contains — which is DoF-11 verbatim. Budget-neutral, roughly 45 words out and 45 in (added 2026-08-13, ONB Q1)
  • [ ] Reviewer Identity names the third defended object — goals, values and north star — alongside unverified claims and incomplete work, with authority to block work that is verified, complete and beside the point
  • [ ] A goal failure is stated as sufficient grounds for ❌ REJECT with every quality check passing; the finding is named not fit for purpose; the verdict routes to the owner, not back to the executor
  • [ ] No new verdict token. The vocabulary stays ✅ APPROVE / 🔄 REVISE / ❌ REJECT
  • [ ] Identity text is not load-bearing: every property above is also enforced by AC-1–AC-3 in judge.md. D46 recorded "Quality guardian, not rubber stamp" and only the first half ever shipped — identity text in this repository has a measured survival rate of ½

Gate: grep -n "frozen baseline\|not fit for purpose" .tfw/workflows/review.md; confirm the verdict set is unchanged Evidence: N/A — textual

AC-5: review.md stays inside the attention budget [depends: AC-4]

F2: working range 700–900 words, hard degradation above 1,200. review.md is 1,065 words today — TFW-56 shortened it by deleting the mode step.

  • [ ] Final count is below 1,200, measured with wc -w and recorded with the before figure
  • [ ] Every word removed to pay for an addition is a genuine restatement, paired in a ledger with the text it restated. Phase B's precedent: a removal that deletes the sole statement of a mechanism is forbidden, and manufacturing "duplication" to buy headroom is the trim this rule exists to prevent
  • [ ] If the working range 700–900 is not reachable without such a trim, report it, do not resolve it — that is what Phase B did at 1,195 and the reviewer accepted it

Gate: wc -w .tfw/workflows/review.md before and after Evidence: the word ledger, one row per removal

AC-6: PV Index gains a source that answers "what are we building"

  • [ ] glossary.md PV Index gains priority 0 — Project North Star: what we are building, why, and what we are deliberately not building. It is stated as distinct in kind from the seven existing sources, which are all how we build
  • [ ] Priority 1 is relabelled .tfw/README.md § Values and Principles — methodology values. Reason to record: that section is byte-identical across projects and cannot carry project information; the current label "README Values" points at a section a real project does not have
  • [ ] conventions.md defines the Project North Star with its admission criteria — a clause states what the product is for or must never become; if a single task's implementation choice could satisfy or violate it, it is a principle (HL §7), not a north star. This is a criterion, not a size cap
  • [ ] The locus is a designated section of a README, and a task HL may never be nominated — excluded explicitly, one sentence (owner ruling 2026-08-13, ONB Q2; the nominated-HL branch needs a project-level freeze mechanism this task has not scoped, and AFD's anchor grew 10 → 14 principles after approval with no log)
  • [ ] The anchor may be more than one location. This project's own north star is its philosophy across the root README.md and the founder essay .tfw/README.md (HL §11 S38)
  • [ ] Priority 0 and priority 1 may point at the same file, and the rule says so. They are distinguished by what the section says — what we are building versus how we build — not by which file holds it. A project whose product is its own method may designate sections of one file to both. Without this clause the framework's own repository can never conform to the rule it ships
  • [ ] The "Who scans PV" block below the index still resolves after the renumbering

Gate: read the index; confirm priorities 1–7 kept their content and only their labels moved Evidence: N/A — textual

AC-7: The HL template reaches the anchor structurally

  • [ ] templates/HL.md header block gains a north-star field, placed below the contract block per that block's own instruction ("Add further header fields below this block, not inside it")
  • [ ] The field takes one or more locations, not a single path — this project's designation is two files
  • [ ] The field states the fallback: project north star → master HL §1 at the frozen baseline
  • [ ] A task with no project north star renders a valid header — explicit N/A grammar (F21), not an absent field

Gate: read the header; confirm the contract block is not modified Evidence: N/A

AC-8: The REVIEW template matches the checklist and surfaces the finding [depends: AC-1]

  • [ ] REVIEW.md §3 row 2 wording matches judge.md row 2 one-for-one — TFW-56 DoD-6 made row-for-row alignment an invariant of this file
  • [ ] The Purpose Check finding is surfaced in the synthesis, not only in the table: a not fit for purpose finding appears in §4 Verdict with its citation and its harm
  • [ ] The template is still ten rows and no new section is added (F22)

Gate: diff REVIEW.md §3 against judge.md's table row by row Evidence: the ten-row alignment check

AC-9: The citation namespace cannot collide

  • [ ] NS{n} is declared for north-star clauses, PP{n} for a project principle registry (KNOWLEDGE.md §0 where a project has one), P{n} unchanged for HL §7
  • [ ] compilable_contract.md's P{N} row is corrected: it currently resolves P{N} to KNOWLEDGE.md §0 Philosophy row, a section D37 removed. knowledge/constraint.md F4 records the double semantics this created. Correct the row to HL §7 and add the two new rows. One table, nothing else in that file
  • [ ] PP{n} is declared even though this repository has no §0 — it is a reserved namespace for projects that keep one, and reserving it is what stops the collision AFD has (three live P8s). If it is declared and unused here, say so in the RF rather than inventing a §0

Gate: grep -n "P{N}\|NS{N}\|PP{N}" .tfw/compilable_contract.md Evidence: the corrected table

AC-10: The two review-side anti-patterns are registered

  • [ ] conventions.md §14: a reviewer approves work that satisfies the TS but not the approved contract or the north star
  • [ ] conventions.md §14: a reviewer asserts alignment without citing the clause it serves
  • [ ] Section-level coordination held (HL §7.1): Phase C appends to §14 and owns the review-flow description. Phase A's HL-contract entries and TFW-56's checklist-row entry are not edited

Gate: git diff .tfw/conventions.md shows additions only, in §14 and the north-star definition

AC-11: Replay — the check fires on the failures and stays quiet on the sound work

DoD-29. A check that fires on everything is as useless as one that fires on nothing. This is also what conventions.md §14 now demands of any checklist row: an evidenced rate, or a written reason for keeping it on consequence.

  • [ ] Run the Purpose Check as shipped against the TFW-48/49 phase REVIEWs, recovered with git show 721ca15:<path> — the reference set is the approved HL at 9e19a4f, not the drifted one
  • [ ] Run it against three TFW reviews that were genuinely sound. Recommended: [TFW-50](../../TFW-50__minimal_agent_commit_attribution/HL-TFW-50__minimal_agent_commit_attribution.md), TFW-42/A, TFW-47/B. Substituting one is allowed with the reason written; do not use TFW-53's own phase reviews — a check cannot be validated on the task that authored it
  • [ ] At least one non-approve on the former set, none on the latter. If the check fires on a sound review, that is a finding about the check, not about the review — report it and stop, do not tune the corpus
  • [ ] Every outcome carries the filled citation-and-harm field, so a reader can see why it fired or did not
  • [ ] What the replay establishes is discrimination, not a rate — nine reviews cannot produce one. State that boundary in the replay file and check the result for consistency with the ~4-in-149 figure the row cites, rather than presenting nine samples as a measurement (corrected 2026-08-13, ONB R3)
  • [ ] Three of the six rejected-corpus replays run post-drift, because TFW-48's pre-amendment HL was never committed — that is the documented TFW-48 failure and the reason DoD-5 exists. Record it per review as a divergence from "the approved HL", never averaged away. It is the harder test, not the easier one (ONB §5 risk 2)
  • [ ] A row that returns the third outcome (reference set inconsistent → owner) is recorded as such, not collapsed into "fired". TFW-49's approved contract is genuinely self-contradictory, so this is a likely and legitimate result (ONB §5 risk 3)

Gate: purpose_check_replay.md — one row per review, outcome, citation, harm Evidence: the replay file, mandatory. This AC cannot be marked N/A

AC-12: The third outcome exists and routes to the owner [depends: AC-1]

Frozen DoD-23, added by approved amendment A6. It had no acceptance criterion until the executor found the gap at ONB; this AC closes it.

  • [ ] judge.md carries a third outcome distinct from pass and fail: the reference set is internally inconsistent — the baseline and the north star, or two clauses of the baseline, cannot both be satisfied
  • [ ] It is recorded as a finding and routed to the owner as a contract defect, not to the executor as a work defect. The executor has no channel to a frozen section, so routing it down is routing it nowhere
  • [ ] review.md's routing block gains one clause for it — the mechanism lives in judge.md, which is off the F2 budget
  • [ ] The precedent is stated in one line so the outcome is not read as theoretical: TFW-49's approved §1 promised "readable without special tooling" while its approved DoD-3 required a versioned structural validator. Part of the scope the owner later rejected was a faithful reading of the DoD the owner approved

Gate: read the block; confirm the routing target is the owner and that a reviewer can reach the outcome without leaving judge.md Evidence: covered by AC-11 — at least one 48/49 replay row is expected to land here

Second execution pass — corrective, added 2026-08-13

Origin: REVIEW Phase C 🔄 REVISE, discrepancy D1, verified independently by the coordinator against 9e19a4f before being accepted.

What is not reopened, stated first so the scope of this pass is unambiguous: no re-run of the replay corpus, no change to the Purpose Check mechanism, no change to the reference-set rule, the three tests, the override clause or the outcome table, no new evidence artifact, and no change to any of the seven framework files beyond AC-14's one blockquote. AC-1 through AC-12 keep their passing status. AC-11's pass condition holds at 4 of 6 and the 0-of-3 sound-corpus result is untouched. This is the correction of one reading, not a re-validation.

What the coordinator has already done, so you do not look for it and do not touch it: HL §2's evidence row and §9's risk row are corrected, and a correction of record is appended below the §12 table. A6's row itself is not rewritten — the log is append-only. The HL is closed to you under the Role Lock; if you find another site, report it in RF §6 and I will land it.

AC-13: Replay row 49/A is re-scored against the full sentence

The row ruled TFW-49 Phase A a contract defect on the reading that §1's "readable without special tooling" cannot be jointly satisfied with DoD-3's versioned structural validator. At 9e19a4f the sentence continues — "…without special tooling, while structural validation prevents quiet drift between Coordinator, Researcher, Executor, Reviewer, adapters, and repositories" — and approved DoF-8 makes prose-only or unversioned-only enforcement a failure condition. §1 asks for both properties; DoD-3 discharges the second.

  • [ ] The row quotes §1 to the end of the sentence and states that §1 pairs readability with structural validation
  • [ ] The recorded outcome is what the shipped check actually returns — ✅ aligned, or a fire on excess grounds if the 1,708-line delivery is argued that way, but not the third outcome
  • [ ] The row's outside confirmation is corrected or dropped: TFW-50's approved "one precise Markdown rule … without enforcement software" ran under a different HL with no DoF-8, so it evidences a change in the owner's preference between two contracts, not an inconsistency inside one
  • [ ] Every dependent site is propagated, none missed: purpose_check_replay.md §1, §3, the §0 research-divergence note and the §3 AC-11 condition table · EV E11 and E12 · RF §3 AC-11 and AC-12 (drop "Exercised for real: replay row 49/A") · RF §2 decision 6 · RF §5's 13/13
  • [ ] The re-score is recorded as a correction with its cause named — a quotation ended early — not silently rewritten. This phase's own thesis is that a change to a shipped artifact must be visible as a change

Gate: git show 9e19a4f:tasks/TFW-49__agent_commit_identity_and_attribution/HL-TFW-49__agent_commit_identity_and_attribution.md — read §1 whole and DoF-8, then diff every site named above Evidence: the corrected replay row, with the full sentence quoted

AC-14: The shipped precedent line no longer teaches the error [depends: AC-13]

judge.md's Purpose Check block closes with a blockquote offering this same reading as the canonical illustration of a contract defect. It ships to every install, three paragraphs below the rule that a citation which resolves but is irrelevant fails the row. This is the item that outlives the task.

  • [ ] The blockquote no longer presents surface tension between two clauses as joint unsatisfiability
  • [ ] Either a pair of clauses that genuinely cannot both be satisfied replaces it, or — since the corpus now yields zero verified instances — the outcome's shape is described abstractly and the absence of an observed case is stated plainly. Saying "not observed" is better than shipping a weak example; HL §9's risk row now reads Unmeasured for exactly this reason
  • [ ] The surviving true half of the old line — part of the scope the owner later rejected was a faithful reading of the DoD the owner approved — is not silently repurposed as a contract-defect example. It is a purpose or excess illustration, or it goes
  • [ ] No net word growth in judge.md. A corrective pass may not grow the artifact it corrects; if the replacement is longer, pay for it inside the same block
  • [ ] review.md is untouched by this AC and stays at its measured count

Gate: read the block; wc -w .tfw/templates/review/judge.md before and after Evidence: the before/after blockquote, quoted in full

AC-15: The two documentary defects are repaired [depends: AC-13]

  • [ ] RF §8 S2 is corrected or withdrawn. Its premise — "the check found its own predecessor's defect while being validated on it" — rests on the ruling AC-13 removes. It is High confidence and bound for KNOWLEDGE.md at KNW, so approving it as written writes a false fact into project memory that later work will cite. What survives is smaller and still real: a replay is worth re-running whenever the check changes, and here the re-run confirmed research's original verdict rather than overturning it
  • [ ] The 48/A quotation marks are fixed. The note reads "preserves all ten mapped principles" inside quotation marks; the source reads "preserves Phase HL P1–P10". A paraphrase inside quotation marks is the same defect class as the truncation — in this phase of all phases, the form is part of the deliverable
  • [ ] No other RF §7 or §8 item is touched — the reviewer checked all three RF §7 candidates and they hold

Gate: re-read both sources; confirm every remaining quotation in the replay and the RF is verbatim to its full sentence Evidence: N/A — textual, verified against the sources named

Evidence Artifacts

File Description
evidence/EV__phase-c__goal_defence_in_review.md Environment header, per-AC table, verdict (required)
evidence/purpose_check_replay.md AC-11 corpus and outcomes (required)

6. Technical Guidance

Reference material, not instructions. Deviate with justification in the RF.

What TFW-56 changed under this phase, four hours before it started. Read RF TFW-56 before touching anything:

  • review.md steps renumbered — Judge is Step 3, not Step 4. The frozen HL Phase C context block still says "Step 4 Judge"; that pointer is stale, not a contract change
  • The Judge checklist is 10 rows, not 7. Rows 8/9/10 (evidence sufficiency, backward compatibility, safety) are new and carry measured rates inside the row — match that grammar
  • Row 2 was deliberately split into two quotable clauses so this phase could replace one and leave the other. RF TFW-56 §6 observation 2: "nothing in either task's frozen DoD requires it to. One line in Phase C's TS closes this." AC-1 is that line
  • conventions.md §14 gained: a review checklist row is added without an evidenced firing rate. AC-11 is how this phase satisfies it
  • verify.md gained a Claim & Source Checks section — unrelated, do not disturb
  • VERSION is 1.1.0. Phase D bumps again; do not touch it here

No amendment to TFW-53 is required by TFW-56. Every frozen DoD 18–29 was checked against the shipped state: none names a mode file, the mapping-integrity check still exists as clause 2(a) so DoD-20 has its target, and review.md got shorter, which makes DoD-28 easier. TFW-56's own DoF-8 — "the change collides with TFW-53 Phase C and forces an amendment against its frozen DoD" — did not trigger.

If forced to choose, the order is frozen (HL §4 Phase C, deliverable weighting): reference-set rule > forcing function > judge.md row > identity text. The anchor degrades gracefully; the reference-set rule does not.

Word ledger for review.md. Line 28 revision ≈ +8, invalid-reference statement ≈ +22, Reviewer Identity third clause ≈ +32, verdict routing ≈ +28, third-outcome clause ≈ +14. The Step 3 paragraph replacement is budget-neutral — roughly 45 words out, 45 in. Total ≈ +104, landing near 1,170 of 1,200. Stated up front so it is not discovered at ONB the way AC-6 was in Phase B; revised 2026-08-13 after the ONB priced the two items the first ledger missed.

Negative controls worth reading before writing the row: AFD-38/phase-b/review/judge.md — a Judge that scored ✅ on the very AC containing the violation, then had its APPROVE retracted by its own author. And AFD-48/phase-b/ — the false-positive precedent that produced the materiality bar.

7. Definition of Failure

  • ❌ Clause (b) Design soundness is lost, weakened, or fused back into a single unquotable sentence — a 4.5% check with six hard failures deleted as collateral
  • ❌ The Purpose Check ships as row 11. The tail positions are the weakest ones for an LLM judge (HL TFW-56 §7.2 #25), and the strongest new check must not sit there
  • ❌ The check can be satisfied by a citation alone, or by a harm alone — the two halves must be one field
  • ❌ A wording objection satisfies the harm clause. That is P14's failure and AFD's first firing
  • ❌ The TS or a Phase HL is left usable as a reference point
  • ❌ The mapping-integrity check survives anywhere — in judge.md or in review.md Step 3. HL DoF-11 names "reverts to mapping integrity" as a failure of this phase, and the check currently has two live statements
  • ❌ The third outcome ships without a route to the owner, or routes to the executor
  • ❌ A project name appears inside a shipped template (F13)
  • ❌ The corrective pass reopens the mechanism, the corpus or an already-passing AC — the scope of the second pass is three documents and one blockquote
  • ❌ A weak contract-defect example is substituted for the withdrawn one rather than admitting the corpus has none
  • ❌ judge.md grows on a corrective pass, or a corrected quotation is fixed without naming why it was wrong
  • ❌ A new verdict token is introduced, or a fifth review stage appears in any form
  • ❌ review.md crosses 1,200 words, or headroom is bought by deleting the sole statement of a mechanism
  • ❌ The replay is skipped, marked N/A, or run against TFW-53's own reviews
  • ❌ Adapter copies, glossary term articles, or a north star for this repository are written here — all three are later phases
  • ❌ A vendor mechanism is named in conventions.md or glossary.md (HL §7.1)

8. Phase Risks

Risk Mitigation
The Purpose Check becomes a rubber stamp — ✅ aligned 145 times in 149 The fused citation-and-harm field makes ✅ cost work; AC-11 measures whether it discriminates at all
It becomes the opposite — reviews blocked on phrasing The materiality bar ships in the same pass, not after. AFD retro-fitted it and paid for a false positive first
Row 2 grows into a paragraph Two clauses, each one quotable sentence. TFW-56 set the length precedent for rows 8–10
The north star does not exist yet in this repository, so the mechanism is untested where it matters most The fallback chain is a first-class AC, not a footnote. TFW-55 fills the anchor; this phase must work before it does
review.md budget squeezed between AC-4's additions and F2 Ledger up front (§6); report rather than trim, per Phase B
Phase D finds three terms shipped without glossary articles Deliberate — Phase D deliverable 1 owns them. Named in §2 so it does not read as an omission

9. Cross-Phase Modifications

File Also modified in Coordination note
.tfw/conventions.md A (§3 + HL-contract §14 entries), B (§14), E (§5, §13, §14) C appends only. Never rewrite another phase's entries
.tfw/glossary.md D (term articles) C adds PV Index rows; D adds the articles. Do not pre-empt
.tfw/templates/HL.md A (contract header, §12), B (§3.1 corrective) C appends one field below the contract block
.tfw/templates/review/judge.md, review.md, REVIEW.md TFW-56 (landed) Read RF TFW-56 first. Row 2 clause (b) is the trap
adapter copies D Out of scope here by design

TS — TFW-53 / Phase C: Goal Defence in Review | 2026-08-13