RF — TFW-53 / Phase A: Contract in Artifacts¶

Date: 2026-08-13 Author: Executor (Claude Code) Status: 🟢 RF — Complete Parent HL: HL-TFW-53 — 🔒 FROZEN TS: TS Phase A ONB: ONB Phase A — 2 blocking questions, both answered Post-review pass: 2026-08-13 — REVIEW findings 1–4 corrected. Verdict was ✅ APPROVE and none was blocking; corrections are recorded in §2 Decision 9 rather than applied silently, because a change to a reviewed artifact should be visible as a change. One template defect (TD-137) fixed; three were inaccuracies in this report, now corrected in place. Third pass: 2026-08-13 — REVIEW §8 re-review. AC-13 and AC-14 added to the TS as a corrective pass, not new scope. AC-13 re-anchors the recovery command (TD-139 / R3); AC-14 closes the last occurrence of the "two fields" error and re-measures every count in §1 (R1, R2). See §2 Decision 10. AC-15, added during the same pass and then revised again after the third review pass (TD-143): rule 15 now ships a subject-only recovery form, not a compressed --grep one — testing showed no --grep form can be subject-only. 162 → 56 words, the measured behaviours moved to §7 for /tfw-knowledge, and — on an owner challenge mid-pass — every operating-system and shell name removed from the shipped core. See §2 Decisions 11–14.

1. What Was Done¶

New Files¶

File	Description
`tasks/TFW-53__hl_contract_and_goal_defence/phase-a/evidence/EV__phase-a__contract_in_artifacts.md`	Structured evidence: environment, per-AC table, verdict, three detailed exhibits (E2, E9, E11)
`tasks/TFW-53__hl_contract_and_goal_defence/phase-a/evidence/baseline_recovery.txt`	AC-6 — the documented recovery command run on live history under both shells, plus the rejected form and the pre-rule baseline
`tasks/TFW-53__hl_contract_and_goal_defence/phase-a/evidence/classification_exercise.md`	AC-8 — five discriminating RES iter1 rows classified from the shipped rule text alone, circularity limit stated first

Zero new framework files. The phase adds nothing to .tfw/ and nothing to a project's root — HL §3.1's "0 new artifacts in a project's root" claim holds.

Modified Files¶

File	Changes
`.tfw/templates/HL.md`	Contract header block (5 fields + a field-usage note, left open for Phase C), with the `Baseline` field carrying the subject-only recovery form; three-state marker on every section heading (`🔒 FROZEN` / `🟢 FREE` / `🟢 APPEND-ONLY`) plus a subsection-inheritance rule; new `## 12. Amendment Log` with 10-column grammar, `Type` and `Verdict` vocabularies, append-only/no-edit-before-verdict/remark-is-not-a-verdict instruction and the `No amendments.` default; §3.1 gate clause block appended (4 properties + checkpoint statement). +79 / −13
`.tfw/templates/RES.md`	`HL Update Recommendations` rewritten as two classed tables — `Refinements` (free sections, coordinator applies) and `Amendment Proposals` (frozen sections, owner verdict required, AC-2's column grammar minus the three fields a researcher cannot fill — `Date`, `Proposer`, `Verdict`); classify-never-apply instruction; explicit-N/A for both classes; line 32 `<!-- … Coordinator applies these. -->` deleted. +33 / −3
`.tfw/conventions.md`	§3 → new `#### HL Contract` subsection: state table + 21 numbered rules across four groups (contract, Contract Baseline, delegated authority, Phase HL), rule 15 carrying the subject-only recovery form; §5 → REJECT branch (a) redefined as filing an amendment, verdict vocabulary untouched; §14 → 7 anti-patterns appended. +55 / −0

(corrected in the third pass — REVIEW §8.3 R1 and R2, TS AC-14. The RES.md cell said "two fields" where Decision 7 says three: the third and last occurrence of that error, one section above its own correction. All per-file counts were carried from before the second pass and are re-measured here, after the third.)

Also modified — executor artifacts, not framework files¶

File	Changes
`README.md`	Task Board row for TFW-53: status and the ONB / RF column links
`tasks/TFW-53__…/phase-a/ONB__phase-a__contract_in_artifacts.md`	§2 Entry Points — three markdown links to `.tfw/…` replaced with backticked paths. Those paths sit outside the docs output tree, so mkdocs emitted an unresolvable-link WARNING for each; backticks match the style the TS itself uses for the same three files. No content changed (disclosed in the post-review pass — commit `e37a8dc` made this change without recording it. REVIEW finding 3)

Total: 167 insertions, 16 deletions across the three framework files; 0 framework files created.

Measured after the third pass with git diff --numstat ffe6c6a -- .tfw/conventions.md .tfw/templates/HL.md .tfw/templates/RES.md. --numstat is quoted deliberately: --stat prints changed lines per file (insertions plus deletions, so HL reads 92 and RES reads 36), and reading that column as insertions is what produced the mismatch REVIEW R2 caught. The --shortstat total for the same three files is 3 files changed, 167 insertions(+), 16 deletions(-). A plain git diff --stat ffe6c6a -- .tfw/ also lists knowledge_state.yaml, changed by the /tfw-knowledge run in commit 8140a85 — not this phase; scoping the path to the three files excludes it.

Across the three commits this phase produced (e37a8dc, 267bd06, and the third pass), two further files were modified outside .tfw/ — the README board row and the executor's own ONB, both executor-writable and both listed below. Budget (30 files / 15 new / 3000 LOC / 30 modified): within all four by an order of magnitude.

2. Key Decisions¶

🚫 WITHDRAWN added to the §12 verdict vocabulary — the single substantive addition beyond the TS. AC-2's Evidence clause asked for a diff of the shipped grammar against the live rows "and record any field the template cannot hold". Row A11 of HL-TFW-53 §12 is 🚫 WITHDRAWN by the coordinator, 2026-08-10, and the drafted four-value vocabulary had no home for it. Recording the gap without closing it would have shipped a template that cannot carry the artifact AC-2 names as its test corpus. Constrained to retraction by the proposer, before a ruling — deleting the row would break append-only, and ❌ REJECTED would credit the owner with a decision they never made. Full exhibit: EV §E2.
The tripwire is time-scoped: "under §5 and §6 as they stand at the moment of classification." Six words added during the AC-8 exercise, which surfaced the ambiguity: row A1 classifies as an amendment against the contract of 2026-08-08 and as a refinement against the contract of today, because A1's own approval added the DoD items that now accept it. Both readings were defensible from the draft text. Full exhibit: classification_exercise.md §Finding.
Subsection state is inherited, not repeated. §3.1 and §3.2 carry no marker and are frozen with §3 — the rule in the template states the inheritance once instead of repeating it at every subsection, which would put three markers on §3 alone. Two subsections do carry a marker, for two different reasons: §7.2 because its state genuinely differs from its parent (free inside a frozen §7), and §7.1 because it is an ## heading — a markup sibling of §7, not a child of it — so inheritance from §7 is not visually obvious to a reader or to a parser. The second reason is the heading-level defect recorded as observation 1 / TD-131; the explicit marker is a mitigation for it, not a redundancy. (corrected in the post-review pass — the original wording claimed §7.2 was the only marked subsection, which the shipped template contradicts. REVIEW finding 4)
The §3.1 gate is appended, never a rewrite. The pre-existing Working Backwards paragraph, the four format options and the §3.2 contrast survive verbatim; the new block reframes the options as choices of which rendering rather than permission to skip one. DoF and AC-11's last bullet both required this.
No budget, slot, size, count or cut-order language anywhere — amendment A12 removed the property from the contract on 2026-08-10. Verified by grep, recorded in RF §4.
conventions.md rules are numbered continuously across the four groups (1–21) rather than restarting per group. A reviewer or a later phase can cite "rule 15" unambiguously; restarting the count would create three different rule 1s inside one subsection.
RES Amendment Proposals carries 7 of §12's 10 columns. The three a researcher cannot fill are Date and Proposer, which the coordinator adds on transcription, and Verdict, which opens as PROPOSED. # is present in both tables: the researcher numbers rows locally, and the coordinator re-assigns them into the HL's continuing sequence, because §12 is append-only and never renumbers. The mapping is stated in the template so the transcription is mechanical rather than interpretive. (corrected in the post-review pass — the original said "two fields" and then listed four, and this decision repeated the error as 7 + 4 = 11. REVIEW finding 1 / TD-137; the template sentence is fixed in the same pass)
Build gate substituted, with the substitution recorded. project_config.yaml build.lint/test/verify are unconfigured starter placeholders that verify nothing. pytest docs/scripts/ + mkdocs build is the pipeline that actually consumes conventions.md and .tfw/templates/** (Source Manifest rows 4 and 13). Approved in ONB Recommendation 6.
The post-review pass corrects in place and says so, rather than editing quietly. REVIEW Phase A returned ✅ APPROVE with five findings, none blocking; four fall inside the executor's Role Lock. One was a real defect in a shipped artifact — templates/RES.md said "minus the two fields" and then enumerated four, in the template that defines a column grammar (TD-137). Three were inaccuracies in this report: a rationale contradicted by its own deliverable (finding 4), an arithmetic repeat of the template's error (finding 1), and a build-gate claim scoped so that the phase's own new warning fell outside it (finding 2) — plus one undisclosed modification, the ONB link change (finding 3). All are corrected above, each carrying an inline note naming the finding it answers. Silent correction was the available alternative and was rejected: this phase's entire thesis is that a change to a reviewed artifact must be visible as a change, and an RF that quietly matches its review is the RF-side version of the drift the contract exists to stop. Not done here: TD-138 (the mkdocs/non-.md gap) is a framework decision, and closing the TD-137 row in TECH_DEBT.md is a coordinator action at /tfw-docs — the fix is shipped, the registry entry is theirs to retire.
The anchor is shipped with its own limit stated, not with implied selectivity. AC-13's replacement command was handed over pre-tested, so the cheap path was to paste it. Running it instead produced a third measurement: git log --grep's ^ anchors to the start of any line, not to the subject — probed with git log -E --grep="^[TD-137](../../../reference/tech-debt.md)", which returns 267bd06, a commit where [TD-137](../../../reference/tech-debt.md) appears only as the first token of a body line. That means the anchor removes the failure actually observed (a pattern quoted mid-sentence, as in f379c5e) but would not remove a body line beginning with a conforming […/freeze/…] prefix. Rule 15 ships that limit as its third bullet, with the practical instruction that follows from it — indent or inline example subjects when quoting them in a message body. Shipping the command without the limit would have made the same class of claim the first form made: correct in its wording, over-promising in its purpose. Probe recorded as §10 of baseline_recovery.txt.
Rule 15 keeps the instruction and gives up the narrative. It had grown one paragraph per corrective cycle — 162 words against a block median of 37, longer than rules 17–21 combined — because each fix documented the failure that produced it. That is the right content in the wrong file: conventions.md is read at execution time by an agent that needs to know what to do, and a rule four times the length of its neighbours is the F2 attention problem in miniature. The two things an editor must not undo — the ^ anchor and the absent leading slash — survive as instructions with one clause of reason each, plus "Both were learned from live failures" so that a future editor knows the constraints are empirical rather than stylistic. The measurements move to §7 as fact candidates, which is the channel /tfw-knowledge consumes.
No pointer from conventions.md to the knowledge that left it — deliberately. The obvious finish is a trailing "why, measured: knowledge/environment.md". knowledge/environment.md exists but does not yet carry these facts; the /tfw-knowledge pass writes them. Shipping the pointer now creates a citation that resolves to a file without the content — the exact shape of S32, the defect this project spent four months not noticing because "a citation that resolves is not a citation that is relevant". The pointer belongs in the same pass that writes the target.
No operating system or shell is named in the three shipped files. Raised by the owner mid-pass on reading rule 15: "is it ok that there is something about windows in conventions?" — it is not. conventions.md is tool-agnostic core copied into every project, and HL §7.1 bars environment- and vendor-specific text there (F13). The compressed rule initially read "because Git Bash on Windows rewrites one as a path"; shipped as "because some shells rewrite a leading slash as a path". The constraint on the editor is identical, and a project on any platform reads a reason that applies to it. The specific platform, shell and measured counts live in §7 FC1 — an environment fact, which is per-project by definition and is precisely why it does not belong in the shared core. Verified: grep -niE "windows|macos|linux|git bash|msys|powershell" over conventions.md, templates/HL.md and templates/RES.md → 0 matches.
The recovery mechanism changed from --grep to subject filtering, and the reason is a measurement, not a preference. AC-13 shipped an anchored --grep form; AC-15's revision argued that the 132 words defending it were defending a weaker mechanism, and that a subject-only form makes the documented limit disappear rather than shorten. Testing that claim produced the decisive result: no --grep form can be subject-only. Git greps a commit message line by line, so ^ (-E) and \A (-P) both anchor to a line start, never to the subject — both candidates returned the constructed fixture whose body quotes a conforming prefix. git log --format="%h %s" filtered on the subject excludes it in both shells, because %s is the subject. Deviation from the frozen HL §4 parenthetical "recoverable via git log --grep", classified under this phase's own rule 6: the declarative claim of Phase A deliverable 3 is a diffable, recoverable baseline, DoD-5 states diffability and names no command, so the command is a specification detail inside an approved deliverable and clears the tripwire — a refinement, not an amendment. Flagged for the coordinator to confirm, since the phrase does appear in a frozen section.

3. Acceptance Criteria¶

AC-1 — The HL declares its own contract state ✅ - [x] Header carries **Contract** with two states (📝 DRAFT — not yet approved / 🔒 FROZEN — approved by {owner} YYYY-MM-DD), approval date in the frozen form - [x] Three states marked, matching HL §3 exactly: 🔒 FROZEN §1/§3/§4/§5/§6/§7 · 🟢 FREE §2/§7.2/§8/§9/§10/§11 · 🟢 APPEND-ONLY §12 - [x] Header points at §12 as the amendment channel - [x] Every section heading carries its marker inline (## 1. Vision 🔒 FROZEN); subsection inheritance stated as a rule - [x] Gate: section lists cross-checked against HL-TFW-53 §3's table — agree exactly, 13 sections in 3 states

AC-2 — §12 Amendment Log with an enforceable column grammar ✅ - [x] Section exists with all 10 columns: # · Date · § · Type · Proposer · Proposed change · Evidence · Cost · Alternatives considered · Verdict - [x] Type values enumerated: EXTEND / SUPERSEDE / RESTRICT (A10's token — the pre-A10 APPLIED — restrictive appears nowhere) - [x] Restrictive rule shipped: applies on filing, verdict ✅ APPLIED — no owner verdict required, restrictive-free prohibited. Stated in conventions.md §3 rule 10 and restated in the template - [x] Empty-verdict value is PROPOSED, with the D28 reasoning line - [x] Explicit-N/A default: If nothing was ever proposed, write: **No amendments.** - [x] Instruction text: append-only, no edit before verdict, a research-thread remark is not a verdict - [x] Evidence: grammar diffed against 12 live rows; one gap found (🚫 WITHDRAWN) and closed → EV §E2

AC-3 — RES classifies instead of instructing ✅ - [x] Two distinct tables: Refinements and Amendment Proposals - [x] Both carry a § column naming the target HL section - [x] Amendment Proposals carries Type, Evidence, Cost, Alternatives — consistent with AC-2 - [x] Instruction states the researcher classifies but never applies; proposals go to HL §12 as PROPOSED - [x] Gate: grep -n "Coordinator applies these" .tfw/templates/RES.md → 0 matches

AC-4 — conventions.md defines the HL Contract ✅ - [x] §3 #### HL Contract: six frozen sections, moment of freezing (owner approval, rule 1), §12 append-only (rule 4) - [x] Free sections named; rule 2 states research updates them directly - [x] Split matches templates/HL.md and HL-TFW-53 §3 — one decision, restated twice, re-decided nowhere

AC-5 — An amendment verdict is a distinct recorded act ✅ - [x] Rule 8: a verdict is a distinct recorded act - [x] Rule 8: input inside a research thread, review or chat is evidence, never approval - [x] Rule 9: an owner-initiated change is an amendment, owner as Proposer, verdict on the same row

AC-6 — The baseline is diffable ✅ - [x] Rule 13: approved HL committed before the first research iteration - [x] Rule 14: reserved freeze scope word per the D55 grammar, applying to the first freeze and every re-freeze - [x] Rule 15: recovery form is slash-free — git log --grep="{TASK-ID}/freeze" — with the MSYS reason stated - [x] Rule 16: no header field can name its own commit - [x] Gate: run under both shells → 5 freeze commits in each; the rejected /freeze/ form returns 0 rows in Git Bash → baseline_recovery.txt

AC-7 — Delegated authority is a ceiling ✅ - [x] Rule 17: a delegated mandate is a ceiling, never a source of new permission - [x] Rule 18: no agent may widen its own grant - [x] Rule 19: delegation is never authority to accept a scope or budget overrun

AC-8 — Granularity rule and carve-out ✅ - [x] Rule 5: the frozen unit is the declarative claim, not the section text - [x] Rule 5: claim-level list — phase set and each phase's declared outcome, §3 to-be claims, each §5/§6 item, each §7 principle, §1 - [x] Rule 6: deliverable-list changes inside an approved phase are refinements, with the §5/§6 tripwire - [x] Rule 7: non-substantive edits are not amendments - [x] Word-count delta measured and reported (§4) - [x] Gate: five discriminating rows (2 tripwire, 1 inverse, 2 control) classified from the shipped text → 5/5 agreement, circularity limit stated first → classification_exercise.md

AC-9 — The Phase HL may no longer author a contract ✅ - [x] Rule 20: derivation-only — may restate master content and add execution context - [x] Rule 21: may not carry its own §1, §5, §6 or §7 - [x] §14 carries the matching anti-pattern - [x] Gate: git show 721ca15:…TFW-48/phase-a/HL__phase-a__method_kernel.md → violation on four independent counts → EV §E9

AC-10 — ❌ REJECT branch (a) no longer thaws the contract ✅ - [x] §5 redefines branch (a) as filing an amendment against the frozen sections - [x] States that re-entry to 📝 HL_DRAFT reopens free sections only and does not thaw §1/§3/§4/§5/§6/§7 - [x] APPROVE / REVISE / REJECT unchanged; the §5 status table untouched (Phase E's territory)

AC-11 — Working Backwards and visualization mandatory in §3.1 ✅ - [x] Working Backwards explicitly required — a description of the plan does not satisfy §3.1 - [x] Visual rendering mandatory, not a format option; prose alone insufficient - [x] The value is shown, not only the artifact - [x] Complete enough to hold at once — phase labels on every change, one line per phase - [x] States §3.1 is the owner's checkpoint before the spend of tokens and time - [x] Clause block, not a rewrite — the four format options and the §3.2 contrast survive verbatim - [x] Gate: no budget, slot, size, count or cut-order language anywhere (grep, §4) - [x] Evidence: shipped rule applied to HL-TFW-53's own §3.1 → 4/4 pass → EV §E11

AC-12 — The anti-pattern set is complete ✅ - [x] All seven anti-patterns present: frozen edit without a logged verdict · unclassified recommendations · amendment applied before its verdict · research on an uncommitted HL · remark treated as a verdict · delegation cited to accept an overrun · Phase HL authoring its own AC or principles - [x] Gate: reproducible §14-block count, 28 → 35 = exactly 7 additions, 0 removals (§4)

AC-13 — Rule 15's recovery form is anchored to the commit subject ✅ (added to the TS 2026-08-13; TD-139 / REVIEW §8.3 R3 — its form was then superseded by AC-15 within the same pass; the objective it was written for is met by the stronger mechanism) - [x] conventions.md rule 15 no longer matches message text at all — it filters %s, which is the subject. Anchoring was the best available fix while the mechanism stayed --grep; AC-15's revision removed the need for it - [x] templates/HL.md header Baseline field carries the same form as rule 15 - [x] Rule 15 states why: --grep also returns commits that merely quote a freeze subject - [x] The no-leading-/ constraint survives with its reason. The ^-anchor clause is gone with the --grep mechanism it defended - [x] Gate: run under both shells → 5 commits each, f379c5e absent from both; both superseded forms run alongside → 6 in both - [x] Evidence appended to baseline_recovery.txt under dated headings; the first pass is intact

AC-14 — RF §1 is internally consistent after the corrective passes ✅ (added to the TS 2026-08-13; REVIEW §8.3 R1 and R2) - [x] RF §1's Modified Files table no longer says "two fields" — it says three and names them, agreeing with Decision 7 - [x] Per-file counts and the total re-measured after this pass, not carried from before it: +79/−13, +33/−3, +55/−0, total 167 / 16 (re-measured once more after AC-15's rewrite, which is why conventions.md reads 55 and not the 58 measured mid-pass — compression removed three lines) - [x] Both corrections carry the inline post-review marker the previous pass used - [x] Gate: the diffstat reproduces every number in RF §1; §1's field count agrees with Decision 7

AC-15 — Rule 15 carries the rule, not its history ✅ (added to the TS 2026-08-13, revised the same day after the third review pass — TD-143) - [x] Rule 15 ships a subject-only recovery form: git log --format="%h %s", filtered on ^\S+ \[[^]]*/{TASK-ID}/freeze/. %s is the subject, so the body cannot match by construction. Candidates tested and rejected are recorded in EV §E15 — including the finding that no --grep form can be subject-only - [x] The "Known limit" bullet and the human-facing indent instruction are deleted, not shortened — with a subject-only form there is nothing to state. grep -niE "known limit|indent it" → 0 - [x] The no-leading-slash constraint survives as an instruction, with its one-clause reason - [x] Rule 15: 162 → 56 words (ceiling 60), against the block's 37-word median - [x] The --grep-matches-bodies behaviour and the two-shell asymmetry moved to §7 FC1, FC4 and FC5 for the /tfw-knowledge pass, with their measurements intact - [x] No other rule in the block grew: all 20 remaining per-rule counts verified byte-identical - [x] Gate: block total 958 → 852, −106 = the rule's own delta exactly. Shipped form run in both shells → 5 commits each, f379c5e absent, and the constructed body-quote fixture excluded (EV §E15) - [x] Additional, on owner challenge mid-pass: no OS, shell or vendor name survives in any of the three shipped files (grep -niE "windows|macos|linux|git bash|msys|powershell" → 0). See Decision 13

15 of 15 acceptance criteria met.

4. Verification¶

Build gate — project_config.yaml build.* are unconfigured starter placeholders (echo "configure your … command") and verify nothing; substituted per ONB Recommendation 6:

python -m pytest docs/scripts/ -q → 68 passed in 60.30s
python -m mkdocs build --config-file docs/mkdocs.yml → built in 35.55s, exit 0. strict is not enabled in docs/mkdocs.yml. The three framework files this phase changed produced no warnings. The phase did add one new warning, from a file it created rather than changed: the EV file links baseline_recovery.txt, and mkdocs cannot resolve a link to a non-.md artifact. The .txt is TS-mandated and the link is correct on disk, so this is a framework gap (TD-138), not a broken reference — but it is a warning this phase introduced and the original wording was scoped so that it fell outside the claim. Left in place deliberately: TD-138 offers three fixes (exclude evidence/ from the docs tree, register non-.md artifacts as extra_files, or have the EV template stop linking attachments), all of which are framework decisions outside Phase A's scope. (rescoped in the post-review pass — REVIEW finding 2)

Measurements

Metric	Before	After	Delta	Command
`conventions.md` words	3,952	5,094	+1,142 (+28.9%)	`$c = Get-Content .tfw/conventions.md -Raw; ($c -split '\s+' \\| Where-Object {$_ -ne ''}).Count`
`conventions.md` lines	513	568	+55	`(Get-Content .tfw/conventions.md).Count`
§14 block items	28	35	+7, −0	`awk '/^## 14\) Anti-patterns/,/^### 14\.1/' .tfw/conventions.md \\| grep -c '^- '`
Rule 15 words	162	56	−106 (AC-15 ceiling 60)	parse the block between the `15.` and `16.` markers
`#### HL Contract` block words	958	852	−106 = rule 15's own delta	parse between `#### HL Contract` and the next `###`

(figures re-measured after AC-15's rewrite. The file peaked at 5,200 words mid-pass when AC-13 added ~132 words to rule 15; AC-15 took them back out and then some. §14 is unchanged by both — AC-13 and AC-15 touch §3 only, verified per-rule.)

On the word delta: TS §6 warns conventions.md is near its attention budget, and +28% in one phase is a real number that Phases B, C and E inherit. It is not compressed below usability here (ONB Risk 5 ruling), but it is the measurement the next three phases should watch. F2's 700–900/1200-word budget governs workflow documents and does not apply to conventions.md — no workflow file was touched.

Targeted gate commands

Gate	Command	Result
AC-3	`grep -n "Coordinator applies these" .tfw/templates/RES.md`	0 matches ✅
AC-11	`grep -niE "cut order\\|budget\\|slot" .tfw/templates/HL.md`	0 matches ✅
AC-2	`grep -c "APPLIED — restrictive" .tfw/templates/HL.md`	0 — the pre-A10 token ships nowhere ✅
AC-6 / AC-13 / AC-15	shipped form: `git log --format="%h %s"` filtered on `^\S+ \[[^]]*/TFW-53/freeze/` (both shells)	5 commits each; `f379c5e` absent; constructed body-quote fixture absent ✅
superseded #1	`git log --grep="TFW-53/freeze"` (unanchored, both shells)	6 — returns the commit that merely quotes a freeze subject ✅ correctly rejected
superseded #2	`git log -E --grep="^\[[^]]*/TFW-53/freeze/"` and `git log -P --grep="\A…"`	6 with the fixture present — no `--grep` form is subject-only ✅ correctly rejected
AC-14	`git diff --numstat ffe6c6a -- .tfw/conventions.md .tfw/templates/HL.md .tfw/templates/RES.md`	`58/0`, `79/13`, `33/3` — reproduces RF §1 ✅
AC-15	rule 15 word count, before / after	162 → 57 (ceiling 60) ✅
AC-15	`#### HL Contract` block word count	958 → 853, delta −105 = the rule's own delta; 20 other rules unchanged ✅
AC-15	`grep -niE "windows\\|macos\\|linux\\|git bash\\|msys\\|powershell"` over the three shipped files	0 matches ✅

DoF check — no workflow file (plan.md, review.md, research/base.md) was opened for writing; no Phase C, D or E deliverable was delivered early; §14 was appended to, never restructured or renumbered; git diff --stat confirms exactly three modified files.

5. Evidence¶

See EV file for evidence details.

Evidence verdict: 6/14 VERIFIED, 0 DEFERRED, 0 BLOCKED, 8 N/A

The eight N/A are the TS's own Evidence: field values, quoted verbatim in the EV table. Six are VERIFIED: the five ACs TS §6 identified as observable against a live artifact or live history (AC-2, AC-6, AC-8, AC-9, AC-11), plus AC-13, whose gate is a command run in two shells and whose evidence is appended to the same transcript file as AC-6's.

6. Observations (out-of-scope, not modified)¶

#	File	Line(s)	Type	Description
1	`.tfw/templates/HL.md`	150, 155	style	§7.1 is an `##` heading while §7.2 is `###`, so §7.1 escapes the §7 nesting §7.2 observes. Pre-existing; ruled out of scope in ONB §6.6. Fixing it changes heading depth in every HL and is a Phase D terminology-pass candidate
2	`docs/scripts/gen_docs.py`	442–465	todo	`_replace_phase` globs `tasks/{id}/Phase{X}/{TYPE}__Phase{X}.md`, but D50 renamed phase folders to `phase-a/`. Every phase reference (`RF TFW-53/A`) silently falls back to the task root and mis-resolves. Live bug; coordinator flagged it for `TECH_DEBT.md` triage at review (ONB §3.1 inconsistency 7)
3	`.tfw/compilable_contract.md`	§2 Reference Format	todo	`P{N}` resolves to `KNOWLEDGE.md §0`, which D37 removed. HL §7 principle references have no resolution target. Phase C owns the namespace work (`NS{n}` / `PP{n}`)
4	`.tfw/project_config.yaml`	1–4, 110–113	todo	`project.name: my-project` and three `echo` placeholders for `build.lint/test/verify`. A consequence of this repository's dual identity (`constraint.md` F6 — simultaneously the upstream template and a live project), but it means no TFW task in this repo has ever had a working `build` gate from config
5	`.tfw/conventions.md`	§3 HL Contract	naming	The subsection is 21 numbered rules in one block. Readable now; if Phases B/C/E each append their own §3 subsections it will need sub-headings. Flagged so the next phase notices before it is a rewrite
6	`tasks/TFW-53__…/phase-a/TS__phase-a__…md`	6, 83	todo	TS header still carries the broken `git log --grep='/TFW-53/freeze/'` form that AC-6 exists to replace, and AC-2's Evidence says "nine live rows" where §12 now has twelve. Both are stale-at-write, not defects in the shipped work; noted so the reviewer does not read them as the contract

7. Fact Candidates¶

fact-candidates: processed 2026-08-18 (/tfw-knowledge, TFW-53 A–E)

#	Category	Candidate	Source	Confidence
> For the `/tfw-knowledge` pass: FC1, FC4 and FC5 are the three behaviours AC-15 removed from `conventions.md` rule 15. They are recorded here with their measurements so the compression loses no knowledge — the rule keeps the instruction, `knowledge/environment.md` should receive the why. `conventions.md` deliberately carries no pointer to them yet: a citation that does not resolve is the S32 failure this project has already documented once, and the pointer belongs in the same pass that writes the target.

| FC1 | environment | The owner's environment runs two shells against the same repository, and they disagree: Git Bash (MSYS2) rewrites a leading / in a command argument into a filesystem path before the program sees it. Measured on this repository: git log --grep="/TFW-53/freeze/" returns 0 rows under Git Bash and 5 under PowerShell 5.1; the slash-free form returns 5 under both. Any shell command written into TFW framework text must be verified in both, or it ships broken for half the sessions on this machine | Executor, measured 2026-08-13; owner accepted as ONB Recommendation 1 | High | | FC4 | environment | git log --grep matches the whole commit message, not the subject line. Measured: the unanchored git log --grep="TFW-53/freeze" returns 6 commits where 5 are real freeze commits — f379c5e matches only because its body quotes the broken pattern it was fixing. Any convention that identifies commits by a subject prefix must anchor, or it degrades as soon as the mechanism is discussed in a commit message — which the mechanism's own corrective commits guarantee | Executor, measured 2026-08-13; REVIEW §8.3 R3 | High | | FC5 | environment | git log --grep cannot be made subject-only. Git matches a commit message line by line, so ^ under -E and \A under -P both anchor to a line start, not to the subject. Probes: git log -E --grep="^[TD-137](../../../reference/tech-debt.md)" and git log -P --grep="\ATD-137" both return 267bd06, where [TD-137](../../../reference/tech-debt.md) occurs only as the first token of a body line; and a constructed fixture whose body quotes a conforming freeze prefix was returned by every --grep candidate and excluded by git log --format="%h %s" filtered on the subject. Consequence for any convention identifying commits by a subject prefix: filter %s, not the message | Executor, measured 2026-08-13 with a constructed negative fixture | High | | FC2 | process | The owner treats a defect found in their own artifact as a result, not a cost: both ONB blocking questions and four of five ONB risks were answered by changing the TS or the contract, including one owner amendment (A12) that removed a property the owner themselves had asked for two days earlier. The gate is used to find errors, not to confirm plans | Coordinator ONB rulings 2026-08-10 ("it is a defect in what I wrote", "your mitigation is better than my AC", A11 withdrawn) | High | | FC3 | process | An executor is an accepted source of amendment proposals in this project: A10 is logged with Proposer = Executor (Phase A ONB Q2) and was approved. The amendment channel is not research-only, and the ONB is a live intake for it | HL-TFW-53 §12 row A10 | High |

8. Strategic Insights (Execution)¶

#	Insight	Category	Source
S1	The owner removed a requirement rather than satisfying it, and named why: it was never asked for. A9's budget and cut-order property was folded in by the coordinator from reference material the owner had supplied as an example, then shipped into the DoD as if it had been requested. The owner's A12 verdict — remove it from §3.1, from `plan.md` and from the DoD — is the amendment protocol running against the coordinator's own inflation, two days after the coordinator wrote it. Implication: the drift this task exists to prevent does not only enter through research findings; it enters through the coordinator's reading of owner input. A9's own §12 row records the owner as `Proposer`, which is exactly what made the inflation invisible — an owner-proposed amendment carries the owner's authority even where its content is the coordinator's	philosophy	Owner ruling 2026-08-10; ONB §3.1 Risk 1; HL §12 A9/A11/A12
S2	The gate paid for itself on its first use, and the payment was in contradictions, not in questions. Both blocking questions turned out to be internal contradictions inside the frozen contract — DoD-2's column enumeration against DoD-6's proposer requirement, and a `Type` value that its own research defined as a `Verdict`. Neither was reachable by reading the TS; both required cross-reading HL, TS and RES iter1 against each other. The coordinator's own assessment: "That is what this gate is for, and it worked on its first use." Implication for Phase C: this is the same defect class as A6's "reference set is internally inconsistent" third outcome — and it was found by an executor at ONB time, not a reviewer at Judge time. The contract-defect route may need an intake earlier in the pipeline than review	philosophy	Coordinator ONB §3.1 Assessment, 2026-08-10
S3	A rule's first live instance is where you learn whether it is enforceable. Three of this phase's shipped clauses exist only because the rule was run against real history rather than reasoned about: the slash-free recovery form (the documented command returned nothing), the `freeze`-word-applies-to-the-first-freeze clause (TFW-53's own baseline commit is non-conforming), and `🚫 WITHDRAWN` (the corpus contained a disposition the vocabulary had no name for). Implication: the evidence requirement is not a reporting overhead on this kind of work — it is the design mechanism. Every one of the three would have shipped broken under a reading-only gate	philosophy	Executor, AC-6 / AC-2 evidence collection 2026-08-13

9. Diagrams¶

Where each rule lives, and what it governs. Phase A ships three files; the arrows are what a later phase must not break.

.tfw/templates/HL.md ─────────────────────────────── the artifact that carries the contract
  header   Contract: 📝 DRAFT | 🔒 FROZEN + date        ← AC-1   (Phase C appends north-star field here)
  §1 §3 §4 §5 §6 §7   🔒 FROZEN  (inline markers)      ← AC-1
  §2 §7.2 §8 §9 §10 §11  🟢 FREE                       ← AC-1
  §3.1     4-property gate clause block                ← AC-11  (A9, narrowed by A12)
  §12      🟢 APPEND-ONLY — 10-column grammar          ← AC-2
              │
              │ transcription: coordinator adds # · Date · Proposer, Verdict opens PROPOSED
              │
.tfw/templates/RES.md ────────────────────────────── where a finding is classified
  Refinements          → free sections   → coordinator applies, no ceremony
  Amendment Proposals  → frozen sections → HL §12, PROPOSED, owner rules   ← AC-3
  (line 32 "Coordinator applies these" — deleted)                          ← AC-3 / DoF-3

.tfw/conventions.md ──────────────────────────────── the governing definition
  §3  #### HL Contract
        rules 1–4    what freezes, when, append-only          ← AC-4
        rules 5–7    granularity: declarative claim, tripwire ← AC-8
        rules 8–12   verdict as an act, RESTRICT, Type        ← AC-5 / AC-2
        rules 13–16  Contract Baseline, git recovery          ← AC-6
        rules 17–19  delegated authority is a ceiling         ← AC-7
        rules 20–21  Phase HL is derivation-only              ← AC-9
  §5  REJECT branch (a) → file an amendment                   ← AC-10
  §14 +7 anti-patterns                                        ← AC-12

The classification decision, as the shipped rules execute it.

finding
  │
  ├─ typo · broken link · formatting? ──────────────────────────► apply, no log        [rule 7]
  │
  ├─ targets §2 §7.2 §8 §9 §10 §11? ───────────────────────────► refinement, apply     [rule 2]
  │
  ├─ narrows — adds a DoF, tightens scope, drops a deliverable?
  │      └──────────────► apply ON FILING + §12 row
  │                       Type RESTRICT · Verdict ✅ APPLIED — no owner verdict required   [rule 10]
  │
  ├─ changes only a deliverable list inside an approved phase?
  │      └─ acceptable under §5/§6 as they stand NOW? ── yes ───► refinement, apply     [rule 6]
  │                                                    └─ no ──► amendment          [rule 6 tripwire]
  │
  └─ changes a phase's set or declared outcome · a §3 claim · §5 · §6 · §7 · §1
         └───────────────────────────────────────────────────► §12 row, Verdict PROPOSED  [rule 5]
                                                                 Type EXTEND | SUPERSEDE  [rule 11]
                                                                        │
                                                                owner verdict — a distinct act  [rule 8]
                                                                ├─ ✅ APPROVED → apply → re-freeze commit
                                                                ├─ ❌ REJECTED → row stays; original holds
                                                                └─ 🚫 WITHDRAWN → proposer retracts, row stays

Phase A defines this machine. Nothing executes it yet — plan.md and research/base.md still carry the old instructions, and Phase B is what makes them obey. A rule defined here and enforced nowhere is the expected intermediate state (TS §2).

RF — TFW-53 / Phase A: Contract in Artifacts | 2026-08-13