REVIEW — TFW-60 / Phase AA: Portable Delivery (revision 3)

Date: 2026-08-28 Author: Claude Code (Reviewer), on_behalf_of: saubakirov, via: claude Verdict: ✅ APPROVE — by owner decision, 2026-08-29, over the reviewer's 🔄 REVISE. Two of the three blocking items were closed by the coordinator before the ruling landed (ab7093e, edab067); the third stands as TD-199. The phase's own closure commit 9570402 records this as a waiver rather than an approval, and that is the more precise word — it is kept here rather than smoothed over. The analysis that produced REVISE is left standing in §4 and in the stage files: the verdict changed, the finding did not. Reviews the R4 corrective pass — TS revision 4, AC-15, twelve items TS: TS Phase AA, revision 4 RF: RF Phase AA — at revision 2. It does not describe this pass Field report: second external update — filed, verbatim, not authored here Contract baseline: master HL at 2123de1 (after amendment A4) Reviewed at: b75bef1. Commits: 22de861 · d62fd26 · 2d269e4 · f7c9dfe · 94f02a6 · fd85b7c · 5e9b0a1 · b75bef1 Stage files: review/rev3/map.md · review/rev3/verify.md · review/rev3/judge.md Earlier passes stand: first — REVISE · revision 2 — APPROVE Owner's stated emphasis: quality, architecture, value, goal. Counts are not the question and are not reopened.

---

1. Map

The awaited evidence arrived and it brought a design error with it.

innoforce-ai-first — a real external project, a different session, no access to this phase's reasoning — ran the full 1.3.0 → 2.0.0-dirty.2 update including the board migration and wrote its own retro. On the phase's purpose the answer is unambiguous: the first consumer "spent the rest of the session reconstructing what to do and in what order"; this one spent nothing. Zero unrecognized directories, zero unaccounted rows, and the identity gate refused a bad value rather than swallowing it.

Then it recorded what refusing cost. The operator wrote the obvious actor: claude-code, was refused, read the validator's source, and invented claude-20260828a. Two projects out of two were pushed into minting a profile per agent session; one later deleted them and left its build gate red permanently — events are immutable, profiles are not, and no action inside that project could ever clear it.

The owner's ruling names the cause rather than the symptom: actor was carrying two unrelated jobs — say who wrote this and make the filename unique — and the two contradict each other. A distinct writer needs a distinct value; a declared handle needs a profile. Remove the field until TFW-54. The pass then closes eleven further items, every one traced to a measured failure in a real project.

2. Verify

Every claim re-derived. Nothing accepted on the strength of a commit message.

#	Checked	Result
1	Full suite	260 passed, 1 skipped
2	mkdocs build	exit 0
3	--check index · tasks · project	exit 0 each
4	Retired vocabulary in the adapter layer	0 files
5	actor across every instruction surface	only prose narrating the retirement survives — zero live instructions to write it
6	§10.4 read back against all templates	20 templates, 0 contradicting
7	Every payload path, three reference forms	unresolved: none — TD-193 and TD-194 close
8	Consumer payload vs v2.0.0-dirty.2	68 files, 0 missing, 0 stray, 65 byte-identical; the 3 that differ are CHANGELOG.md (one commit behind the tag's own record) and the two project-owned files
9	--check tasks on the live consumer, current code	15 tasks validate — it was "2 problems across 15 tasks" under the shipped code
10	git status --porcelain in both consumers, before and after	empty — my verification wrote nothing to a project that is not ours

The architectural question, answered

Is removing actor a loss of capability? No — it is the removal of a false claim.

Measured, the field never carried an independent fact: upstream, 21 consecutive events repeat on_behalf_of in it; in the consumer it is via plus a session number. And the frozen master HL §3.1 asks that a team be able to answer "who/what acted" — two questions. Two fields answer them: on_behalf_of — who — and via — what. The three-field model was the over-claim; the two-field model is closer to the frozen text, not further from it.

What is genuinely given up — telling two concurrent sessions of one tool apart inside the journal — was never delivered; it is what forced the per-session profiles. Git commit attribution (D55) still answers it where it was always answered. The canon says the gap is open and gives it an owner: "A writer is not named yet, and saying so is the point."

And the fix costs nobody anything. An already-written actor is tolerated, never required, never rewritten. 28 events here and 46 across two consumers keep their bytes. I re-measured it myself: the same bytes that were red under the shipped code are green under this one, with zero bytes changed in that project.

innoforce-ai-first, same files, read-only

  under 2.0.0-dirty.2   2 problem(s) across 15 tasks   ← permanently red: the profiles
                                                          those events name were deleted
  under this pass       15 tasks validate               ← nothing in that project changed

That is how a schema change earns trust, and it is measured rather than argued.

AC-13 half two — ruled MET

The executor left E51 DEFERRED and routed the ruling here, correctly: "whether the phase's declared outcome is thereby met is a reviewer's and the owner's ruling, not the executor's to grant itself."

The TS's own bullets	Measurement
a real external project, updated by its own operator	a different project, a different session, its own document filed verbatim
zero files hand-carried	0 missing, 0 stray of 68 payload files
zero edits inside .tfw/	65 of 68 byte-identical; the three exceptions are one CHANGELOG commit and the two project-owned files
every delta the first consumer invented is unnecessary	board flags shipped, team/README.md withdrawn, tooling in the payload — all three confirmed
the run records what was confusing	five findings, four of which became AC-15 items

The findings it returned are what the criterion asked for, not evidence that it failed — the bullet requires the run to record what was confusing. And the regress needs naming: b75bef1's message says half two "closes on a third external run." It does not. Requiring a fresh external run to certify each correction the previous run produced defers the phase forever. DoD 19 says an external project completes the update from the payload alone. One did, and I measured it.

3. Judge

#	Check	Status	Evidence
1	DoD met?	⚠️	The engineering meets it and AC-13 half two is now MET. AC-15's twelve items have no RF row and no EV row, and the version marker names a release the CHANGELOG does not describe
2	Purpose · design soundness	(a) ✅ · (b) ✅	(a) serves DoD 19 and the north star's "a session is not an independent person" (PV 3 · D59) — which the framework's own gate was violating. (b) the strongest of the phase: an entity carrying two jobs found, removed, generalized into a DoF rule, and proved on live corpora
3	Tech debt documented	❌	No RF §6 entries for this pass. The word-ceiling breach is recorded only in an evidence file that defers to "the RF"
4	Style & standards	⚠️	Canon, templates and adapter layer clean. update.md 1380 words against a ceiling of 1200
5	Observations collected	❌	Nothing to triage
6	RF completeness	❌	The RF predates the pass entirely
7	Evidence completeness	⚠️	Three good artifacts, none indexed in the EV
8	Evidence sufficiency	✅	Where evidence exists it is unusually strong; every gate re-derived and every one held
9	Backward compatibility	⚠️	The carrier change is handled exactly right — tolerance, not migration. But a consumer acted on a rule this pass retracts (F6)
10	Safety	✅	Both third-party projects read and never written

Purpose Check: Aligned ✅. Reference set re-checked: §3.1's "who/what acted" is two questions and now has two fields; the baseline states that "the exact filenames and aggregation mechanism remain research decisions", so the filename grammar was never frozen. No contract defect — the contract permits this change. Only the TS forbids it, and that is F2.

4. Verdict

✅ APPROVE — by owner decision, 2026-08-29, over the reviewer's 🔄 REVISE.

What actually happened, and it is not what an earlier draft of this section said. This review returned 🔄 REVISE with three items, all record-keeping and none in the engineering. The owner ruled them not worth a round: «отметить как апрув, эти мелочи править не будем».

Two of the three were already closed by then, by a coordinator session working concurrently in the same tree while this review was being written:

Item	Disposition
2 — the TS mandated in AC-15 what §1, §7, §8 and §9 forbade	Closed, ab7093e — the four sentences now state the exception once with its reason, every other prohibition intact, and §9 tells Phases B and C directly that they extend a two-field event
3 — VERSION said 2.0.0-dirty.3 and no changelog entry described it	Closed, edab067 — the entry leads with what an updating project must do about the actor removal: nothing
1 — the R4 pass has no RF	Outstanding, filed as TD-199. Genuinely waived

F4 (update.md at 1380 words) and [TD-197](../../../reference/tech-debt.md) (via unvalidated) were moved into Phase AB rather than waived — both land in files that phase opens anyway.

The word the closure commit uses is "waiver, not approval", and it is the better word. Blocking item 1 stands: the phase's largest architectural change has no executor declaration. What survives it is real — TS AC-15 carries the analysis and §2 of this review carries an independent measurement — but neither is an RF.

Nothing in the engineering was ever in question. Every gate was re-derived here, and §4 below already called this the strongest work of the phase.

The reasoning that produced the original verdict follows unchanged.

Let me be exact about what I was returning, because the work itself is the best in this phase.

The architecture is right, and right for the right reason. The report handed over a symptom with two branches — bless per-session agent handles, or say the opposite. Taking the first would have shipped a naming convention and left the contradiction in place. Instead the pass asked what job the field actually does, found two, established that one of them cannot be done honestly until there is a principal that delegates and answers to someone, and removed it. Then it wrote the general rule into the DoF — "an entity whose single job was never named aloud" — and listed four instances from this task against it, two of which the phase had itself committed. That is a finding about how the project fails, not a patch.

The value is delivered and measured. A live external project's permanently-red gate goes green with zero bytes changed inside it. Not argued — measured, by the executor and again by me.

And it refused the small answer twice more. §10.4 was rewritten as a rule rather than patched at its example, with my revision-2 objection quoted rather than overridden — 20 templates, 0 contradicting. PROVIDER_FAMILIES was deleted rather than documented, because with actor gone nothing reads it.

What I was returning is the record, not the work — all three now carried as debt by owner ruling rather than closed in a round:

#	Item	Routes to
1	The pass has no RF. Header says revision 2, cites TS revision 3, fourteen AC rows, no AC-15, no key decision for the two-jobs analysis, no verification block, no observation for the ceiling breach — which r4_gates.txt itself defers to "the RF." Three good evidence artifacts exist and none is indexed in the EV, which stops at E63. The Trust Protocol makes the RF the thing a review is conducted against; for the largest architectural change of the phase there is no declaration to test. In six months the reasoning that makes this defensible lives only in commit messages	executor
2	The TS mandates in AC-15 what it forbids in §1, §7, §8 and §9. §1 "The model itself does not change… no carrier, schema… is touched"; §7 "❌ The model changed: any edit to a carrier schema, the event grammar…"; §8 "A finding about the model is filed, not fixed here"; §9 "changes no carrier they will extend." The "Why no amendment" box is a sound argument about the master HL — no §5 or §6 clause is crossed and I agree — and it never touches the TS's own four sentences. An executor reading §7 before AC-15 is told the pass is a declared failure. No amendment is needed; the TS is not frozen and needs its own text brought into line with the ruling it now carries	coordinator
3	.tfw/VERSION and tfw.version say 2.0.0-dirty.3; the CHANGELOG has no such entry — its newest is [2.0.0-dirty.2] and [Unreleased] says "Nothing pending." Meanwhile eight payload files teach a reader about 2.0.0-dirty.3: the canon, the glossary, the migration guide, the event template, bindings.yaml, the adapter README, update.md, project_config.yaml. A project reading "an event written before 2.0.0-dirty.3 carries actor" and turning to the changelog finds no such release. This is AC-14's own criterion and HL principle 10, and it is the phase's subject inverted: not an instruction naming a missing file, but a version naming a missing record. The entry must also carry F6 — that copies are now the declared adapter model and the "commands never duplicate workflow content" rule is withdrawn, because a consumer has already acted on it	executor (release surface, TS-authorized per ONB Q3)

Two further items, not blocking, to be closed in the same pass or filed:

• F4 — update.md 1165 → 1380 against the §11 ceiling of 1200. This is the workflow a receiving project actually runs, and the one file this phase had deliberately brought under the ceiling by deleting duplication rather than content (D9). Run D9's move again, or get a coordinator ruling. handoff, init, plan and review are also over and were before this pass; update.md is the only one this pass pushed across.
• F5 — via is now validated by nothing while the canon still states it as an enumeration. Under Structural Enforcement, a rule that cannot reveal its own violation is advice. Either say via is free-form provider text, or check it. One deliberate sentence, not drift.

This is not a REJECT and it is not close to one. Purpose is aligned, the contract permits the change, nothing is unsafe, no work is wasted, and no finding touches the design. Three documents have to catch up with what was built.

What this approval covers

AC-13 half two is met and does not need re-running. The phase closes on the evidence already filed: a real external project, updated by an operator who is not the author, 0 files hand-carried and 0 framework files edited inside .tfw/, independently re-measured by this review. What a third external project produces afterwards is Phase AA's successor evidence, not its gate.

What it does not cover — one item, and it is the one that is genuinely waived. TD-199: the R4 pass ships with no RF, so the two-jobs analysis, the verification numbers and the update.md ceiling breach live only in commit messages and three evidence files the EV does not index. TS AC-15 carries the reasoning and §2 of this review carries an independent measurement — both permanent, and neither is an executor's declaration of what was built.

Blocking items 2 and 3 are not in this category: ab7093e brought the TS's four sentences into line with its own ruling, and edab067 wrote the 2.0.0-dirty.3 entry. F4 and [TD-197](../../../reference/tech-debt.md) moved into Phase AB, which opens the same files anyway.

5. Tech Debt Collected

No RF §6 entries exist for this pass, so nothing was triaged from the executor. TD-192, TD-193 and TD-194 close with this pass — §10.4 rewritten as a rule, the path check extended to all three reference forms, the glossary swept. Two new rows from this review:

#	Source	Severity	File	Description	Action
TD-197	REVIEW TFW-60/AA rev3 §4 F5	Low	.tfw/scripts/gen_index.py, .tfw/conventions.md:317	PROVIDER_FAMILIES was deleted at 2.0.0-dirty.3 with sound reasoning — its only reader was the actor gate. But the canon still states via as an enumeration ("provider family — claude, codex, gemini") and nothing validates it. Under the project's own Structural Enforcement value a rule that cannot reveal its own violation is advice	⬜ Open — one deliberate decision: declare via free-form provider text, or check it. Not urgent; the harm is drift, not breakage
TD-198	REVIEW TFW-60/AA rev3 §4 F6	Med	.tfw/adapters/claude-code/README.md, .tfw/CHANGELOG.md	A consumer acted on a rule this release retracts. innoforce-ai-first rewrote 12 command files from copies to thin adapters, citing the shipped sentence "Commands never duplicate workflow content — they reference it." AC-15 item 12 deletes that sentence and declares copies the model. The consumer followed a shipped rule and the rule was withdrawn behind it; its thin adapters now diverge from the declared shape and Step 6's re-sync will not recognize them	⬜ Open — fold into blocking item 3: the 2.0.0-dirty.3 entry states the retraction and what a project that already went thin should do. A retraction reaching a project that acted on it is the class this phase measures

6. Traces Updated

• [x] the phase's status.md — lifecycle: DONE, closed by the coordinator at 9570402 with the outcome recorded as a waiver: "closed by owner 2026-08-29 with review rev3's blocking item 1 outstanding; the pass has no RF of its own"
• [x] TECH_DEBT.md — TD-197, TD-198 (this review) and TD-199 (blocking item 1, filed by the coordinator) added; TD-192, TD-193, TD-194 closed by the R4 pass. There is no TD-200 or TD-201: blocking items 2 and 3 were fixed, not deferred
• [ ] tfw-docs: Pending — TD-186 (KNOWLEDGE.md:22) is its first item, and §1–§3 now need the two-field identity model
• [ ] tfw-knowledge: Pending — this review's two candidates plus the nine standing from earlier passes

7. Fact Candidates

fact-candidates: processed 2026-08-30

#	Category	Candidate	Source	Confidence
1	philosophy	A gate that refuses without naming what it accepts makes the reader fabricate data to get past it. The identity gate correctly rejected actor: claude-code, named no valid form, and the operator read the validator's source and minted a per-session handle — then a second project did the same and, on tidying up, left its build gate red with no action available to it. The refusal was right and the design behind it was wrong: before adding a gate, say what passes it, and if that cannot be said the gate is guarding a field that should not exist	Second external update, F1; owner ruling 2026-08-28	High
2	process	A ruling that changes the spec must sweep the spec. AC-15 was written into a TS whose §1, §7, §8 and §9 still forbid exactly what it mandates — the same class the phase itself prosecutes in the payload, committed in the document that governs the payload. The coordinator caught the shape one commit earlier ("exactly the residue I had just finished telling the executor not to leave behind") and caught it in the header only	Reviewer, this revision	High

---

REVIEW — TFW-60 / Phase AA: Portable Delivery (revision 3) | 2026-08-28, reviewed at b75bef1