Skip to content

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

Date: 2026-08-13 Author: Executor (Claude Code) Status: 🟠 ONB — answered 2026-08-13, cleared to execute. Q1 → (a) · Q2 → (a) + owner ruling on this project's north star · Q3 → (a). TS amended: review.md:85 added to §4, AC-1 and AC-11 reworded, AC-2 gains the deferral confession, AC-6/AC-7 gain the two-file case, AC-12 added for DoD-23 Parent HL: HL-TFW-53 — 🔒 FROZEN, re-frozen after A1–A14, baseline e8ee76e TS: TS Phase C — approved by the owner 2026-08-13 Predecessors read: RF Phase B, REVIEW Phase B ✅ APPROVE, RF TFW-56, RES iter2

1. Understanding

Phases A and B stopped the goals from moving. This phase catches the case where the goals held and the work went elsewhere anyway. The Judge's mapping-integrity check — structurally unable to 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 and never a Phase HL. The check cannot be satisfied by assertion: one fused field carrying a quoted clause and a named material harm, an excess-and-adjacency test, an override clause that rules out "the TS scoped it this way" and "tests are green", and a third outcome when the reference set contradicts itself. Around it: PV Index gains priority 0, conventions.md defines the north star with admission criteria, templates/HL.md reaches the anchor structurally, the citation namespace stops colliding, and the whole thing is replayed against the six TFW-48/49 phase REVIEWs and three sound ones so the check is shown to discriminate rather than to fire on everything.

Nine of the eleven ACs are textual. The two that carry risk are AC-5 (review.md word budget) and AC-11 (the replay, which can return a finding against the check itself).

2. Entry Points

File Where What is there today
.tfw/templates/review/judge.md line 16 Row 2, two clauses: (a) Mapping integrity · (b) Design soundness (4.5%). 10 rows total, rows 8–10 carry measured rates inside the row (TFW-56 grammar). 639 words
.tfw/workflows/review.md line 28 "Master HL for the task — understand vision, design philosophy, architecture decisions" — no revision named
.tfw/workflows/review.md line 35 Reviewer Identity: "Quality guardian. Your job is to protect the project from unverified claims and incomplete work." Two quality failure modes, no purpose
.tfw/workflows/review.md line 85 Step 3's HL §7 Principles check paragraph — the mapping-integrity check, stated as a workflow instruction. Not listed in TS §4 → Q1
.tfw/workflows/review.md lines 89–98 Step 4 Decide — verdict set APPROVE / REVISE / REJECT, no routing rule for a goal failure
.tfw/workflows/review.md whole file 1,065 words (wc -w). F2: ≤1,200 hard, 700–900 working
.tfw/templates/REVIEW.md line 32, §4 §3 row 2 label "Philosophy aligned — mapping integrity + design soundness"; §4 Verdict has no purpose finding
.tfw/glossary.md lines 221–237 PV Index priorities 1–7, priority 1 = "README Values"; "Who scans PV" block directly below
.tfw/conventions.md §3, §14 §3 = artifact types + HL Contract rules 1–21 (Phase A's); §14 = anti-pattern list, last entry is TFW-56's checklist-row rule
.tfw/templates/HL.md lines 3–16 Header block; the contract block ends with "Add further header fields below this block, not inside it"
.tfw/compilable_contract.md line 59 P{N} → "KNOWLEDGE.md §0 Philosophy row" — §0 was removed by D37. Line 65 carries a second stale §0 reference → §6.5
Replay corpus git show 721ca15:<path> 6 phase REVIEWs (TFW-48 a/b/c, TFW-49 a/b/c) + review/judge.md per phase, all present. Approved TFW-49 HL at 9e19a4f present
Sound corpus working tree [REVIEW TFW-50](../../TFW-50__minimal_agent_commit_attribution/REVIEW__TFW-50__minimal_agent_commit_attribution.md), REVIEW TFW-42/A, REVIEW TFW-47/B — all three present

3. Questions (blocking — cannot proceed without answers)

# Question Answer
1 review.md:85 still instructs the mapping-integrity check, and it is not in TS §4. The paragraph reads "For each mapped principle: verify the linked AC was met in RF §3. If a principle was mapped to an AC but that AC failed — flag as a principle violation." AC-1 deletes that check from judge.md; nothing in the TS touches this paragraph, so as specified the phase ships a workflow that instructs the check its own template no longer contains. This is the third appearance in this task of the surviving-channel defect (templates/RES.md:32 → :133 were the first two). (a) Replace it with the Purpose Check instruction + reference set — ~45 words out, ~45 in, so AC-5 pays nothing; (b) delete it — −45 words of AC-5 headroom, but the "mapped to a failed AC = principle violation" rule loses its only statement; (c) leave it, report in RF §6 — costs nothing now, ships a live contradiction and exposes HL DoF-11 ("the goal check … reverts to mapping integrity"). My recommendation: (a). (a) — replace. Your reading is right and the cost analysis is right. Two additions: (i) the mapping-integrity rule is not "losing its only statement", it is being retired on purpose — DoD-20 says the check is replaced, and DoF-11 names "reverts to mapping integrity" as a failure condition of this phase. Option (c) does not merely ship a contradiction, it ships DoF-11. (ii) review.md:85 is now in TS §4 with its limit — "the HL §7 Principles check paragraph in Step 3, and nothing else in that step". Recorded before you touch it, same discipline as templates/RES.md:133 in Phase B. An unrecorded scope extension is still a scope extension, and I am the one who has to write it down
2 Q9 — is the nominated-HL north-star locus supported, excluded, or left silent? HL §8 records Q9 as "decide at Phase C TS"; the TS restates both branches ("supported only as nominated-and-frozen, or excluded explicitly") without choosing, so the decision is still open and AC-6 cannot be closed without it. (a) Exclude explicitly — one sentence in conventions.md: the locus is a designated section of the root README.md, and a task HL may not be nominated. Cost: AFD's live practice (HL-AFD-2) becomes non-conformant and would have to move its 45 load-bearing lines into its README. (b) Support as nominated-and-frozen — requires a project-level freeze mechanism this task has not scoped (HL §9 says so in as many words); shipping the permission without the mechanism is a rule with no enforcement site (F4, HL DoF-1). (c) Say nothing — silence reads as permission and reproduces the measured drift (AFD's anchor grew 10 → 14 principles after approval, six unlogged markers). My recommendation: (a) — it is the only branch whose mechanism exists on the day it ships. (a) — exclude explicitly. Owner ruling 2026-08-13, and the practice now agrees with the rule: this project's north star is its philosophy, and it lives in the root README.md and the founder essay .tfw/README.md. A README locus, never a task HL. Three consequences you must build for, all recorded as HL §11 S38. (i) The anchor is a list, not a path — AC-7's header field takes more than one location, with explicit N/A when a project has none. (ii) Priority 0 and priority 1 may point at the same file, and the rule must permit it. DoD-18 relabels priority 1 because .tfw/README.md § Values and Principles is byte-identical across projects and cannot carry project information — true everywhere except here, where the methodology is the product. One clause in conventions.md: the two priorities are distinguished by what the section says, not by which file holds it; a project whose product is its method may designate sections of the same file to both. Without that clause this repository can never conform to a rule it wrote. (iii) Do not write the designation into either README. A TFW-55 session is restructuring both files right now (§5 risk 5) — a section-level pointer landed today is the stale-pointer defect this task keeps finding (S32, TD-163). Phase C ships the slot and the grammar; the owner's designation is recorded in HL §11 S38 and lands as a header field when the next HL here is written
3 Does the deferral confession test ship in this phase? Frozen §4 Phase C lists it in the Phase C vocabulary and frozen DoD-30 makes Phase D write its glossary article — but no DoD item and no AC puts it anywhere in judge.md, and TS §2 lists only three tests. Research defined it (D16/E7): the spec or the result says the right home for this is elsewhere, and ships it here anyway. (a) Ship it as a third test below the table — ~20 words in a template, no review.md cost, and Phase D's article then has a referent; (b) do not ship it — Phase D defines a term that appears nowhere in the framework, which is the S32 defect (a label with nothing behind it) one layer up. My recommendation: (a), and if it is (b) I will say so in the RF rather than let Phase D discover it. (a) — ship it. Not a scope extension: frozen DoD-30 obliges Phase D to define the term, and a glossary article whose referent exists nowhere in the framework is S32's defect one layer up — a label that resolves and means nothing. Rule 6's tripwire clears, because shipping it is what makes DoD-30 non-vacuous under §5 as it stands. AC-2 now carries it as a fourth test, worded as research defined it: the spec or the result names a different home for this work and it ships here anyway. Keep it to one sentence below the table — TS §8's "row 2 grows into a paragraph" risk applies to the block too

4. Recommendations (suggestions, not blocking)

  1. Do not name AFD inside the shipped template. AC-1 asks the row to carry "roughly 4 goal-based blocks in 149 AFD reviews". judge.md ships to every TFW project, and its established grammar for provenance is a bare corpus reference ("measured non-✅ rates from a 637-row corpus", rows 8–10 carry only the number). Default I will apply unless told otherwise: the rate and the consequence-not-frequency reason ship in full; the project's name does not. Recorded as a deviation in the RF either way.
  2. Put the three tests and the override clause below the table, not in the row. TS §4 already says this; naming it here because TS §8 lists "row 2 grows into a paragraph" as a risk and the row has to stay two quotable sentences to keep AC-1's clause-(b) property true.
  3. Reconcile, do not compute, the base rate in AC-11. Nine reviews cannot produce a rate. What the replay can establish is discrimination — fires on the rejected corpus, silent on the sound one — plus consistency with the ~4-in-149 figure. I will state that boundary in the replay file rather than present 9 reviews as a measurement.
  4. TD-155 stays open and this is the phase that was told to close it. Its action reads "→ next phase that may edit conventions.md §3 (not B — out of scope)" — Phase C does edit §3. But the fix is a rewording of Phase A's rules 13–15, and both HL §7.1 and TS §9 forbid rewriting another phase's entries. Reporting it so the reviewer does not read the omission as an oversight.
  5. review.md will be reported at ~1,155 words, not resolved into 700–900. Same posture Phase B took at 1,195 and the reviewer accepted, and AC-5 bullet 3 instructs it explicitly.

Coordinator response to §4 and §6, 2026-08-13.

R1 — no project name in a shipped template: agreed, and AC-1 is corrected to say so. It asked for "4 goal-based blocks in 149 AFD reviews"; judge.md ships to every TFW project and its own grammar is a bare corpus figure. F13. The rate and the consequence-not-frequency reason ship; the provenance stays in this task's trace, where a reader who wants it can find it.

R2 — tests and override clause below the table: confirmed, it is what TS §4 specifies and what keeps AC-1's clause-(b) property true.

R3 — reconcile, do not compute: agreed, and this is the better science. Nine reviews cannot produce a rate. AC-11's word "reconciled" means exactly what you propose — state the boundary, show discrimination, do not dress nine samples as a measurement. AC-11 is reworded to say discrimination, not rate.

R4 — TD-155 stays open: correct. Both §7.1 and TS §9 forbid rewriting Phase A's rules 13–15, and that prohibition outranks a debt row's routing note. The note was written before the section-ownership rule existed. Report it; I will re-route it to Phase D.

R5 — report ~1,155, do not resolve: confirmed. AC-5 bullet 3 instructs it and the Phase B precedent stands.

§6.1 — DoD-23 had no acceptance criterion. That is my gap, not yours, and it is now AC-12. You were right to flag it rather than implement it silently: an executor implementing a frozen DoD item with no AC is indistinguishable, from the outside, from an executor exceeding scope. AC-12 gives it a gate so the RF has something to close against.

§6.3 — compilable_contract.md:65 stays out of scope, deliberately. AC-9's "one table, nothing else" exists because that file is the reference registry for the whole framework and a namespace pass is not licence to edit prose in it. Report the second stale §0 reference in RF §6; it goes to Phase D with TD-155.

§6.4 — the duplicate deliverable number is fixed. Phase C had two items numbered 4; renumbered to 5–8 in the HL, marked as a rule 7 non-substantive correction. Your deliverable accounting can now read against a clean list.

§5 risks — all seven accepted as written. Two get instructions. Risk 2: record post-drift replays per review, never averaged; the divergence from AC-11's "the approved HL" is real and is the harder test, so state it as a divergence rather than a footnote. Risk 5: if README.md is still held by the other session, leave the board row uncommitted and say so in the RF — I will land it. Do not stage a hunk you did not write.

5. Risks Found (edge cases, potential issues not in TS)

  1. The replay can return a finding against the check itself. AC-11 forbids tuning the corpus if the check fires on a sound review — it must be reported and execution stopped. Concretely likely on REVIEW TFW-42/A: it is the one sound-corpus review the research iteration did not use (research ran TFW-46/A, TFW-47/B, TFW-50), so it is genuinely untested. That independence is worth having, and it is also where a false positive would appear first.
  2. TFW-48's reference set is not recoverable at its frozen baseline. Its pre-amendment HL was never committed — that is the documented TFW-48 failure and the reason DoD-5 exists. So three of the six rejected-corpus replays must run against the HL as it stood at 721ca15, i.e. post-drift. That is a strictly harder test for the check (it has to fire using master DoF-12 and P7 while reading the drifted contract), not an easier one, but it is a divergence from AC-11's "the approved HL" wording and will be recorded per review rather than averaged away.
  3. AC-11 may land on the third outcome rather than on a fire. TFW-49's approved contract is internally contradictory (§1 "readable without special tooling" vs DoD-3's "versioned structural validator"). A reviewer running the shipped check on 49/A can legitimately return reference set inconsistent → owner. That satisfies "non-approve" but not "fires"; the replay will report which of the two it is per row instead of collapsing them.
  4. Two of three adapter surfaces will instruct the old review flow from the moment this phase lands. .claude/commands/tfw-review.md and .agent/workflows/tfw-review.md are out of scope (Phase D owns the sync, TS §2). TD-157 already records this window for tfw-plan; after this phase it is open for tfw-review too, and the D54 parity promise is live-broken until Phase D runs.
  5. A concurrent session holds README.md. A TFW-55 session has an uncommitted Task Board row change at line 302 right now. Phase B's fbdf443 swept a sibling task's work into this task's commit for exactly this reason (TD-144). I will stage by explicit path, and for README.md stage only my own hunk after checking the diff — or leave the board row uncommitted and say so in the RF.
  6. The third outcome costs review.md words that the TS ledger did not price. TS §6 budgets +90 for line 28, the invalid-reference statement, the identity clause and the verdict routing. DoD-23's contract-defect routing is a fourth routing branch. Mitigation: the mechanism lands in judge.md (a template, off the budget) and review.md gets one clause inside the routing block it already gains.
  7. The check's own first firing could be a wording objection. AFD's was, and it was demoted after owner challenge. The materiality bar ships in the same pass (AC-2, P14) — the risk is not that it is missing but that it is stated as advice instead of as a failing condition. It will be worded as the latter.

6. Inconsistencies with Code (spec vs reality)

  1. Frozen DoD-23 has no acceptance criterion. The third outcome (reference set internally inconsistent → owner as a contract defect) is a frozen DoD item added by approved amendment A6. It appears in TS §2 In Scope and TS §4's description of judge.md, but no AC and no gate names it, and TS §7 does not list its absence as a failure. I will implement it — it is inside the declared scope, so this is an AC-coverage gap, not a scope extension — and record it in the RF. Flagging it because a reviewer checking DoD 18–29 against the AC set will find the same hole.
  2. review.md:85 — see Q1. Spec says the mapping-integrity check is replaced; the workflow still instructs it.
  3. compilable_contract.md line 59 is the defect AC-9 names; line 65 is a second one it forbids me to touch. Where references appear opens with "KNOWLEDGE.md §0 Source column" — the same section D37 removed. AC-9 says "One table, nothing else in that file", so the stale prose reference survives the pass that corrects the table. Reported, not fixed.
  4. HL §4 Phase C has two deliverables numbered 4 (the Reviewer Identity amendment and the verdict semantics, HL lines 518–519). Non-substantive under rule 7 and inside a frozen section I have no channel to; recorded so the RF's deliverable accounting reads against a known defect rather than looking like a miscount.
  5. glossary.md:213 places the Knowledge Gate in "Phase 0 of plan.md" and plan.md has Steps 0–7 with no Phase 0. Already TD-163, already assigned to Phase D. Not touched here; noted because Phase C does edit glossary.md and the reviewer will see the file in the diff.
  6. process.md F19's first half is already historical — it describes review.md's non-standard Step 0, deleted by TFW-56. The fact carries its own correction dated 2026-08-13, so nothing is required; confirming I read it against the shipped file rather than trusting it.

7. Knowledge Citations

HL §7.2 carries 26 items. Every one was opened and checked against the shipped file, not accepted from the table. Six new items are added below.

# HL §7.2 ref Read? Applied / N/A Notes
1 .tfw/README.md § Structural Enforcement ✅ Applied The check lives in judge.md, the template filled every review — not in review.md prose. Heading confirmed at line 100
2 .tfw/README.md § Naming Creates Behavior ✅ Applied Purpose Check, not fit for purpose, Project North Star, NS{n} ship as the named vocabulary. Heading at line 104
3 .tfw/README.md § Candor Over Flattery ✅ Applied Q1–Q3 are the uncomfortable questions; each carries a cost and a refusal consequence. Heading at line 88
4 KNOWLEDGE.md §1 D19 ✅ N/A here Research→HL channel; discharged in Phases A/B. No review-side surface
5 KNOWLEDGE.md §1 D20 ✅ Applied indirectly Implicit approval is why the reference set must name the frozen baseline rather than "the HL"
6 KNOWLEDGE.md §1 D23 ✅ Applied Workflow compression — AC-5's ledger; review.md additions are clauses inside existing blocks, not new prose sections
7 KNOWLEDGE.md §1 D24 ✅ Applied Pattern A: the reference set, the tests and the override clause are inline in judge.md; only conventions.md rule 15 is referenced, and AC-3 requires exactly that
8 KNOWLEDGE.md §1 D31 ✅ N/A here Filesystem state; the north star is a README section, not a marker file
9 KNOWLEDGE.md §1 D49 ✅ Applied Gates > guidelines — the fused field is a failing condition, not advice
10 KNOWLEDGE.md §1 D54 ✅ Partially N/A Adapter parity is Phase D; §5 risk 4 records the window this phase opens
11 knowledge/philosophy.md F4 ✅ Applied Structural over exhortation — the reason AC-4 says the identity text must not be load-bearing
12 knowledge/philosophy.md F13 ✅ Applied Domain-agnostic — drives Recommendation 1 (no project name in a shipped template) and the wording of the harm clause
13 knowledge/philosophy.md F21 ✅ Applied Explicit N/A — AC-7's header field must render for a project with no north star
14 knowledge/philosophy.md F22 ✅ Applied Template minimalism — no new REVIEW section (AC-8), no row 11 (AC-1)
15 knowledge/philosophy.md F25 ✅ Applied Framework proposes, human decides — Q1–Q3 present options with costs, not decisions
16 knowledge/process.md F4 ✅ Applied Numbered steps + gates; the review.md routing addition is a labelled branch, not a paragraph
17 knowledge/process.md F6 ✅ Applied The recorded, unfixed instance — the reason the Purpose Check exists at all
18 knowledge/process.md F14 ✅ Applied Agents route around non-structural rules; hence the check in the template, not in identity text
19 knowledge/process.md F20 ✅ Applied HL authoritative on WHAT; the executor reports the DoD-23 gap (§6.1) rather than resolving it silently
20 knowledge/constraint.md F2 ✅ Applied ≤1,200 / 700–900 — AC-5, measured before and after
21 .tfw/conventions.md §7 ✅ N/A here Execution modes; AT is TFW-54
22 .tfw/conventions.md §15 ✅ Applied Role Lock — I write ONB, RF, evidence and the framework files in TS §4. No HL, no TS, no REVIEW
23 KNOWLEDGE.md §1 D55 ✅ Applied Commit subjects [claude-code/TFW-53/phase-c/executor]; local commits only, push on explicit approval
24 knowledge/process.md F11 ✅ Applied Organic emergence → formalisation: AFD's memory-only reviewer rule becomes repo text
25 KNOWLEDGE.md §1 D43 ✅ Applied Knowledge Citations as the anti-hallucination device — the fused field is the same mechanism, with the harm half added because a resolving citation is not a relevant one (SS2)
26 KNOWLEDGE.md §1 D46 ✅ Applied Reviewer Identity, and its measured ½ survival rate — AC-4's non-load-bearing requirement

New items the coordinator's table does not carry:

# Source Item Why it belongs here
N1 KNOWLEDGE.md §1 D61 TFW-56: universal 10-row Judge checklist, measured rates carried inside each promoted row, structural explicit-N/A at three sites, D42 revoked The grammar AC-1 must match. The HL predates TFW-56, so §7.2 cannot cite it — and it is the single most binding constraint on how row 2 may be written
N2 knowledge/constraint.md F4 "P{N} has irreconcilable double semantics: KNOWLEDGE.md §0 vs HL §7" — ✅ verified AC-9's subject, cited in TS §5 but absent from §7.2. It is also the evidence that reserving PP{n} is a fix rather than a precaution
N3 KNOWLEDGE.md §1 D53 evidence/ mandatory in every task directory; RF §5 is a pointer, not a table Governs phase-c/evidence/ and the EV file's shape
N4 knowledge/philosophy.md F24 Instructions produce compliance, heuristics produce competence; cross-stage structural dependency enforces better than a mandate Why the citation and the harm are one field: a reviewer who cannot fill it has failed the row by construction, with no separate rule to obey
N5 knowledge/process.md F19 Corrected 2026-08-13: review.md Step 0 is Session Naming; Judge is Step 3 Confirms the frozen HL Phase C pointer to "Step 4 Judge" is stale, non-substantive (rule 6)
N6 knowledge/process.md F25 A fabricated citation traversed an entire pipeline undetected and was caught only by the owner The concrete precedent for why an unverifiable citation must fail the row rather than pass it

Cross-references: use Reference Format (e.g. [RF TFW-18](../../TFW-18__knowledge_consolidation/RF__PhaseB__knowledge_quality.md), D24, [TD-72](../../../reference/tech-debt.md)). See compilable_contract.md §2. Build script resolves to hyperlinks.

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