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

Date: 2026-08-10 Author: Coordinator (Claude Code) Status: 🟡 TS_DRAFT — Awaiting approval Parent HL: HL-TFW-53 — 🔒 FROZEN, baseline git log --grep="TFW-53/freeze" ⚠️ Phase B/C/D/E executors: copy this slash-free form. The '/TFW-53/freeze/' form this header carried until 2026-08-13 returns zero rows under Git Bash on Windows — the exact bug AC-6 exists to fix, reproduced in the TS that fixes it. Closed as TD-136.

1. Objective¶

Make the HL contract exist as artifact state. Today an approved HL is indistinguishable from a draft: there is no approved marker, no record of what may not move, and no channel for a change that must not be made silently. This phase gives templates/HL.md a contract header, frozen/free marking and an append-only §12 Amendment Log; gives templates/RES.md a two-class recommendation split so a researcher classifies instead of proposing edits wholesale; and writes the governing rules into conventions.md — what freezes, at what granularity, how the baseline is recovered, and what a Phase HL may no longer do. Nothing here changes a workflow. Phase B makes the workflows obey what this phase defines.

2. Scope¶

In Scope¶

templates/HL.md — contract header block, frozen/free section marking, §12 Amendment Log, and the §3.1 Result Visualization gate (four properties).
templates/RES.md — HL Update Recommendations split into Refinements and Amendment Proposals; removal of the line that instructs the coordinator to apply them.
conventions.md — HL Contract definition (§3), granularity rule, baseline reference, Phase HL constraint, REJECT branch (a) redefinition (§5), and seven anti-patterns (§14).

Out of Scope¶

Any change to plan.md, review.md or research/base.md — Phase B and C own those. A rule defined here that is not yet enforced anywhere is the expected intermediate state.
The Project North Star, PV Index priority 0/1, and the north-star header field — Phase C. templates/HL.md is touched again there; see §9.
Glossary articles and adapter sync — Phase D.
The ❌ REJECTED status row and trace restoration — Phase E, which also edits conventions.md §5.
The salami git diff pre-TS check — open thread, decided in Phase B TS or recorded in TECH_DEBT.md.

3. Principles Check¶

#	Principle (from HL §7)	Enforced by	Gate
P2	Classify, never edit	AC-3	`templates/RES.md` offers two classes and states the researcher never applies; the "Coordinator applies these" line is gone
P3	Structural enforcement over guidelines	AC-1, AC-2, AC-6	The contract is artifact state (header + §12) and a git-recoverable baseline, not advice
P5	Evidence, cost, alternative	AC-2	§12's column grammar makes all three mandatory fields; a row cannot be filled without them
P6	Narrow D19, don't revoke it	AC-3	RES still produces HL feedback; only the frozen channel changes from write to propose
P7	Token density	AC-8	Additions are numbered rules and table rows; no prose blocks. `conventions.md` growth measured in AC-8
P9	Naming creates behavior	AC-1, AC-2	The shipped words are `Contract`, `FROZEN`, `Amendment`, `EXTEND`/`SUPERSEDE` — not "update recommendation"
P10	Authority cannot self-extend	AC-7	`conventions.md` states a delegated mandate is a ceiling and may not be cited to accept an overrun
P11	A remark is not a verdict	AC-5	`conventions.md` states research-thread input is never approval, and an owner-initiated change is itself an amendment
P12	A frozen baseline must be diffable	AC-6	Commit-before-research rule plus the reserved `freeze` scope word
P1	The contract earns the autonomy	N/A	Task-level claim realised across all phases; no single Phase A artifact enforces it
P4	Batch, don't interrupt	N/A	Escalation behaviour lives in `plan.md` — Phase B
P8	Tool-agnostic by behavior	N/A	Adapter parity is Phase D; nothing here is tool-specific
P13–P16	Purpose check, materiality bar, citation, judge against baseline	N/A	Review-side; Phase C
P17	A failed trace is the most valuable trace	N/A	Phase E

4. Affected Files¶

File	Action	Description
`.tfw/templates/HL.md`	MODIFY	Contract header block; frozen/free marking; new `§12 Amendment Log`; §3.1 gate clauses
`.tfw/templates/RES.md`	MODIFY	Two-class recommendation split; delete line 32 comment
`.tfw/conventions.md`	MODIFY	§3 HL Contract + granularity + baseline + Phase HL rule; §5 REJECT branch (a); §14 seven anti-patterns

Budget: 0 new files, 3 modifications, est. ~180 added lines. Defaults: max 14 files, max 8 new, max 1200 LOC, max 12 modified. Well within all four.

5. Acceptance Criteria¶

AC-1: The HL declares its own contract state¶

templates/HL.md carries a header block that makes an approved HL structurally distinguishable from a draft, and marks every section as frozen or free.

[ ] Header block carries a contract state field with at least the states "draft" and "frozen", plus the approval date
[ ] Three states are marked, matching HL §3 exactly: 🔒 FROZEN (§1, §3, §4, §5, §6, §7), 🟢 FREE (§2, §7.2, §8, §9, §10, §11), 🟢 APPEND-ONLY (§12) (corrected 2026-08-10 — the original two-state wording contradicted its own gate)
[ ] A pointer to the amendment channel (§12) is present in the header
[ ] Each section heading carries its state marker inline (## 1. Vision 🔒 FROZEN), so the state is visible at the section and not only in the header. Verified safe for the docs build: gen_docs.py derives no anchors from headings

Gate: open .tfw/templates/HL.md, confirm all five items by reading. Cross-check the section lists against HL-TFW-53 §3's frozen/free table — the two must agree exactly. Evidence: N/A — template content; correctness is established by reading, and there is no runtime behaviour to observe.

AC-2: `§12 Amendment Log` exists with an enforceable column grammar¶

The template carries an append-only amendment log whose columns make evidence, cost and a considered alternative impossible to omit.

[ ] §12 Amendment Log section exists with columns: number, date, section, type, proposer, proposed change, evidence, cost, alternatives considered, verdict (proposer added 2026-08-10 per ONB Q1 — DoD-6 requires a logged proposer and had no field for one)
[ ] Type values are enumerated. Third value is RESTRICT if amendment A10 is approved, APPLIED — restrictive verbatim if rejected. Ship this bullet last; everything else in AC-2 is unblocked
[ ] A restrictive change (adds a DoF, narrows scope, drops a deliverable) applies on filing with the verdict ✅ APPLIED — no owner verdict required; restrictive-free is prohibited. Rule stated in conventions.md §3 beside the granularity rule (iter1 D8, previously stated in no DoD item — ONB Q2)
[ ] The empty-verdict value is PROPOSED, with one line of D28 reasoning: it describes the state of the request, which is what the log tracks
[ ] The section carries an explicit-N/A default ("No amendments.") per F21
[ ] Instruction text states the log is append-only, that a frozen section may not be edited before its verdict, and that a research-thread remark is not a verdict

Gate: open the template; confirm the column set and the three type values; confirm the explicit-N/A line is present. Evidence: The amended template must be able to carry HL-TFW-53's own §12 unchanged. Diff the shipped column grammar against all live rows in HL-TFW-53 §12 — twelve as of 2026-08-13, and the count keeps moving — and record any field the template cannot hold. This is a real artifact with real rows, not a constructed example.

AC-3: RES classifies instead of instructing¶

templates/RES.md splits its recommendations into two classes and no longer tells the coordinator to apply them.

[ ] HL Update Recommendations contains two distinct tables: Refinements and Amendment Proposals
[ ] Both tables carry a column naming the target HL section
[ ] The Amendment Proposals table carries evidence, cost and alternatives columns, consistent with AC-2's grammar
[ ] Instruction text states the researcher classifies but never applies, and that amendment proposals go to HL §12 as proposed
[ ] The comment  is removed

Gate: grep -n "Coordinator applies these" .tfw/templates/RES.md returns nothing; both tables present on reading. Evidence: N/A — template content.

AC-4: `conventions.md` defines the HL Contract [depends: AC-1]¶

The governing definition exists in one place: what freezes, when it freezes, and what append-only means.

[ ] §3 defines the HL Contract: the six frozen sections, the moment of freezing (owner approval), and §12 as append-only
[ ] The definition names the free sections and states that research updates them directly
[ ] The frozen/free split matches templates/HL.md exactly — one source of truth, restated not re-decided

Gate: read conventions.md §3; diff the two section lists against templates/HL.md and against HL-TFW-53 §3. Evidence: N/A — convention text.

AC-5: An amendment verdict is a distinct recorded act [depends: AC-4]¶

The rule that closes the TFW-49 S6 failure mode, in both directions.

[ ] conventions.md states that a verdict on an amendment is a distinct, recorded act
[ ] It states that input given inside a research thread is never approval
[ ] It states that an owner-initiated change to a frozen section is also an amendment, logged with the owner as proposer and the verdict on the same row

Gate: read conventions.md; confirm all three statements present and unambiguous. Evidence: N/A — convention text.

AC-6: The baseline is diffable [depends: AC-4]¶

An approved contract that cannot be diffed is not frozen. This closes the TFW-48 failure mode, where drift is documented and permanently unverifiable.

[ ] conventions.md requires the approved HL to be committed before the first research iteration
[ ] It defines the baseline reference as a reserved freeze scope word in the commit subject, consistent with D55's [agent/task/scope/role] grammar, and states that it applies to the first freeze as well as every re-freeze
[ ] The documented recovery form is slash-free — git log --grep="{TASK-ID}/freeze". The /freeze/ form returns silently empty under Git Bash on Windows because MSYS rewrites the leading slash as a path (measured; ONB Recommendation 1)
[ ] It states that no header field can name its own commit, so the baseline lives in the commit subject and not in the file

Gate: read conventions.md; then run the documented recovery command against this repository under both Git Bash and PowerShell and confirm both return the TFW-53 freeze commits. Evidence: Run the documented command in this repository and capture its output. It must return the freeze commits created during TFW-53 planning. This is the mechanism working on live history, not a described intention.

AC-7: Delegated authority is a ceiling [depends: AC-4]¶

The clause that prevents TFW-49's self-authorised overruns, landed here so TFW-54 inherits it rather than inventing it.

[ ] conventions.md states that a delegated mandate is a ceiling and never a source of new permission
[ ] It states that no agent may widen its own grant
[ ] It states that delegation is never valid authority to accept a scope or budget overrun

Gate: read conventions.md; confirm all three statements. Evidence: N/A — convention text.

AC-8: The granularity rule and the carve-out are stated [depends: AC-4]¶

The decision that moves escalation from 4.6 to ~2.3 proposals per iteration. Without it the contract is correct and unusable.

[ ] §3 states that the frozen unit is the declarative claim, not the section text
[ ] It states what is frozen at claim level: the phase set and each phase's declared outcome, §3's to-be claims, §5/§6 items, §7 principles, §1
[ ] It states that a deliverable-list change inside an already-approved phase is a refinement unless it cannot be accepted under the existing §5/§6 — the tripwire
[ ] It states that non-substantive edits (typos, broken links, formatting) are not amendments
[ ] conventions.md total word count is measured before and after; the delta is reported in the RF

Gate: read §3; then classify five discriminating recommendation rows from research/iter1/RES.md using only the shipped text — including at least two D4-tripwire cases where the naive reading and the RES classification diverge — and compare against the RES assignment. Evidence: Record the exercise in the EV file with the shipped rule text quoted and each row's outcome, and state the circularity limit explicitly: the RES classifications were produced by the same researcher who wrote the rule, so agreement demonstrates the text is readable, not that the rule is correct. A 5/5 score presented as validation is worse than a 3/5 presented honestly.

AC-9: The Phase HL may no longer author a contract [depends: AC-4]¶

Closes the drift channel one level below the master HL.

[ ] §3 defines the Phase HL as derivation-only: it may restate master content and add execution context
[ ] It states the Phase HL may not carry its own §1, §5, §6 or §7
[ ] §14 carries an anti-pattern for a Phase HL that authors acceptance criteria or principles

Gate: read §3 and §14. Then read git show 721ca15:tasks/TFW-48__*/phase-a/HL__phase-a__method_kernel.md and confirm the shipped rule would classify it as a violation. Evidence: Record in the EV file which specific parts of the historical TFW-48 Phase A HL the shipped rule prohibits — its own DoD, DoF and principle list. The rule is validated against the artifact that motivated it, not against a hypothetical.

AC-10: `❌ REJECT` branch (a) no longer thaws the contract [depends: AC-4]¶

The one documented path that reopened frozen sections with no proposal, no evidence and no log.

[ ] conventions.md §5 redefines branch (a) "rework HL" as filing an amendment against the frozen sections
[ ] It states that re-entry to 📝 HL_DRAFT does not thaw frozen sections
[ ] The existing three-verdict vocabulary (APPROVE / REVISE / REJECT) is unchanged

Gate: read conventions.md §5; confirm the branch text changed and the verdict set did not. Evidence: N/A — convention text.

AC-11: Working Backwards and visualization become mandatory in §3.1 [depends: AC-1]¶

Four properties, in short clauses, appended to the existing §3.1 instruction block.

⚠️ Revised 2026-08-10 by amendment A12. The earlier wording carried a budget and cut-order property. It is removed from the contract entirely — do not implement it, in §3.1 or anywhere else. If any draft of this clause already mentions budget, slot, size, count or cut order, delete it.

[ ] Working Backwards, explicitly required — §3.1 is written from the finished state as if it already exists, never as a description of the plan that will produce it
[ ] Visual rendering is mandatory, not a format option — ASCII diagrams, flows, file and folder trees, before/after tables, mockups, sample output. Prose alone does not satisfy §3.1
[ ] The value is shown, not only the artifact — what the result is worth is visible in the same picture
[ ] Complete enough to hold at once — for multi-phase tasks every change carries its phase label and each phase gets one line saying what it is for
[ ] The clause states that §3.1 is the owner's checkpoint before the spend of tokens and time
[ ] The addition is a clause block, not a rewrite: the existing format options and the §3.2 contrast survive intact

Gate: read templates/HL.md §3.1; confirm the four properties and the checkpoint statement are present, the pre-existing instruction text is retained, and no budget or cut-order language appears anywhere. Evidence: Apply the shipped §3.1 rule to HL-TFW-53's own §3.1 and record whether it passes all four properties. This HL now carries a file-level change map, a per-phase purpose map, an end-to-end flow and a six-months-later view, so a pass is expected — a failure would be a finding worth reporting.

AC-12: The anti-pattern set is complete [depends: AC-4, AC-5, AC-7, AC-9]¶

conventions.md §14 gains the seven anti-patterns this phase's rules imply.

[ ] Editing a frozen section without a logged owner verdict
[ ] Submitting recommendations without classification
[ ] Applying an amendment before its verdict
[ ] Starting research on an uncommitted HL
[ ] Treating a research-thread remark as a verdict
[ ] Citing one's own delegation as authority to accept an overrun
[ ] A Phase HL that authors its own acceptance criteria or principles

Gate: a reproducible count of the §14 block specifically — not a loose file-wide pattern, since conventions.md carries bullet lists in nine other sections. Record both counts and the exact command used, so the reviewer can rerun it. Confirm exactly seven additions and no removals. Evidence: N/A — convention text.

AC-13: Rule 15's recovery form is anchored to the commit subject¶

(added 2026-08-13 after the second review pass — TD-139 / REVIEW §8.3 R3. Corrective pass, not new scope.)

The shipped form git log --grep="{TASK-ID}/freeze" is unanchored, and --grep searches the whole commit message. Commit f379c5e matches because its body quotes the broken pattern it was fixing, so the documented command now returns six commits where five are real freeze commits. No false negatives — AC-6's gate as worded still passes — but a later reader is handed a non-freeze commit as a baseline candidate, and the noise grows with every commit that discusses the mechanism.

[ ] conventions.md §3 rule 15 carries the anchored form: git log -E --grep="^\[[^]]*/{TASK-ID}/freeze/"
[ ] templates/HL.md header Baseline field carries the same form
[ ] Rule 15 states why it is anchored: --grep matches message bodies, so an unanchored pattern matches any commit that merely discusses freezing
[ ] The leading character is ^, not / — this is what keeps it MSYS-safe. State that too, or the next editor will "simplify" it back

Gate: run the new form under both Git Bash and PowerShell; both must return exactly the real freeze commits and must not return f379c5e. Run the old form alongside to show the difference. Evidence: append both transcripts to evidence/baseline_recovery.txt under a dated second-pass heading. Do not overwrite the first pass — it is the record of the first failure this rule survived.

AC-14: RF §1 is internally consistent after the corrective passes¶

(added 2026-08-13 — REVIEW §8.3 R1 and R2. Folded in because the RF is being reopened for AC-13 anyway, exactly as the reviewer recommended.)

[ ] RF §1's Modified Files table no longer says "minus the two fields a researcher cannot fill" — it is three, and Decision 7 already says three. This was the third occurrence of the same numeric error, one section above its own correction
[ ] RF §1's per-file counts and totals are re-measured after this pass, not carried from before it
[ ] Both corrections carry the same inline _(corrected in the post-review pass — REVIEW finding N)_ marker the previous pass used

Gate: git diff --stat against the freeze baseline reproduces every number in RF §1, and RF §1's field count agrees with Decision 7. Evidence: N/A — RF-internal accuracy, verifiable by reading against a diffstat.

AC-15: Rule 15 carries the rule, not its history [depends: AC-13]¶

(added 2026-08-13 — owner review of shipped density. Same rule AC-13 already reopens, so no extra context.)

Measured: rule 15 is 162 words against a median of 37 across the block's 21 rules, and longer than rules 17–21 combined. It grew one paragraph per corrective cycle — MSYS after the first, anchoring plus a "Known limit" after the second — until it documented its own scar tissue. conventions.md carries rules; KNOWLEDGE.md and knowledge/ carry why.

Revised 2026-08-13 after the third review pass (TD-143). Do not simply compress the current wording. The reviewer's finding is stronger than the owner's and supersedes it: the 132 added words defend a weaker mechanism. --grep searches the whole message, so ^ matches the start of any line, so the rule needs a documented limit and an instruction to humans about indenting example subjects when quoting them. A subject-only form has no such limit by construction — the limit disappears instead of being documented.

[ ] Rule 15 ships a subject-only recovery form. --grep is not subject-scoped; a form that filters on %s is. Test candidates and ship one that returns exactly the real freeze commits with no body matches
[ ] The "Known limit" bullet and the human-facing indent instruction are deleted, not shortened — with a subject-only form there is no limit to state
[ ] The MSYS no-leading-slash constraint survives as an instruction; it was learned from a live failure and is not explained away by the new form
[ ] Rule 15 lands at ≤ 60 words against the block's 37-word median. It is 162 today
[ ] The --grep-matches-bodies behaviour and the two-shell asymmetry move to the /tfw-knowledge pass as environment facts, with their measurements intact. conventions.md carries the rule; knowledge carries why
[ ] No other rule in the block grows in this pass

Gate: word-count rule 15 before and after. Run the shipped form under both shells; it must return exactly the real freeze commits, must not return f379c5e, and must not return a commit whose body quotes a conforming prefix — construct that commit locally to prove it if none exists. Evidence: both word counts, both shell transcripts, and the body-match negative test in the EV file. A compression that kept the weaker mechanism would pass the word count and fail the negative test; both are required.

Evidence Artifacts¶

File	Description
`phase-a/evidence/EV__phase-a__contract_in_artifacts.md`	Structured evidence: environment header, per-AC table, verdict (required)
`phase-a/evidence/baseline_recovery.txt`	Output of the documented `git log --grep` recovery command (AC-6)
`phase-a/evidence/classification_exercise.md`	Five RES iter1 rows classified using only the shipped granularity rule (AC-8)

6. Technical Guidance¶

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

The source of truth for the frozen/free split is HL-TFW-53 §3. Do not re-derive it. Both templates/HL.md and conventions.md restate it; if the three ever disagree, the HL wins and the disagreement is a finding for the RF.
A worked example of everything in this phase already exists. HL-TFW-53's own header, §12 (nine rows, all ruled) and §3.1 were written by hand as a prototype. Read them before writing the template — but treat them as a draft to improve, not a spec to copy. The HL says so explicitly in its own header note.
conventions.md is near its attention budget. Every addition should be a numbered rule or a table row. Prose blocks are the failure mode here (F2, D23, process.md F4).
§14 is edited by four phases. A owns the seven entries above; B, C and E append their own. Add, never restructure — a renumbering in this phase creates conflicts in three others.
templates/RES.md line 32 is the exact target of AC-3's deletion. It is the template-side twin of the plan.md Step 6c instruction Phase B rewrites; shipping one without the other reproduces the failure through the surviving channel.
Naming is a deliverable, not decoration (D28). The words that must appear are Contract, FROZEN, Amendment, Amendment Log, EXTEND, SUPERSEDE, RESTRICT, Contract Baseline. (the third Type value was APPLIED — restrictive until amendment A10 renamed it; do not ship the old token) Phase D consolidates the glossary; do not invent synonyms here that Phase D then has to reconcile.
Evidence proportionality. Most ACs here are text edits, correctly marked N/A. Four are not, and they are the ones where a live artifact or live history can be observed: AC-2, AC-6, AC-8, AC-9, AC-11. Those are where the EV file earns its place.

7. Definition of Failure¶

❌ The frozen/free split in templates/HL.md, conventions.md and HL-TFW-53 §3 do not agree exactly.
❌ The contract rules ship as advisory prose without artifact state — the D17/Pattern-B failure, repeated inside the task that exists to prevent it.
❌ templates/RES.md retains "Coordinator applies these" while shipping the two-class split — DoF-1 delivered inside the enforcement site.
❌ The granularity rule is stated so abstractly that the AC-8 classification exercise cannot be completed from the shipped text alone.
❌ Any workflow file (plan.md, review.md, research/base.md) is modified in this phase.
❌ Any Phase C, D or E deliverable is delivered early "while we are in the file".
❌ conventions.md §14 is restructured or renumbered rather than appended to.

8. Phase Risks¶

Risk	Mitigation
Rules land in `conventions.md` but the template does not carry the state, leaving the contract advisory	AC-1 and AC-2 are template ACs with their own gates; DoF names the failure explicitly
The granularity rule is written to sound right rather than to be applied	AC-8's five-row classification exercise against RES iter1 is a falsifiable gate, not a reading
`conventions.md` grows past its attention budget across four phases	AC-8 requires the word-count delta in the RF; B, C and E inherit the same measurement
The prototype in HL-TFW-53 is copied verbatim including its rough edges	Technical Guidance names it a draft to improve; AC-11 turns the HL's own §3.1 into a test subject
Section-level collisions with Phases B, C, E on `conventions.md`	§9 below; append-only discipline in §14; A sequenced before E on §5

9. Cross-Phase Modifications¶

File	Also modified in	Coordination note
`.tfw/templates/HL.md`	Phase C	C adds the north-star pointer field to the same header block. Leave the block extensible; do not close it with a fixed field list
`.tfw/conventions.md` §5	Phase E	A rewrites REJECT branch (a); E adds the `❌ REJECTED` status row. Different lines. Sequence A before E, or coordinate at E's TS time
`.tfw/conventions.md` §14	Phases B, C, E	Append-only. Each phase adds its own entries; nobody renumbers
`.tfw/conventions.md` §3	Phase C	C adds the Project North Star and PV priority 0/1 correction. A owns the HL Contract subsection; C adds its own
`.tfw/glossary.md`	Phase D	Terms shipped as text here are defined as glossary articles there. Use the canonical words listed in §6 so D has nothing to reconcile

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.

TS — TFW-53 / Phase A: Contract in Artifacts | 2026-08-10