Dispatch Rubric

Active penalties — in force now

These are rules, not suggestions. A session that reads this page is bound by them until the date shown.

Statistical honesty scored 2 on 2026-08-07

For the week following, NO strategy claim goes into action without a second-opinion session run first. A second-opinion session is a separate session that sees the claim and its evidence, not its reasoning, and either confirms it or refuses it. If no second-opinion session has run, the claim does not go into action. This is not advice.

In force until the 2026-08-14 scoring

Resource stewardship scored 1 on 2026-08-07

No metered resource may be consumed at all until a written cap has been given to Kayla and she has answered. Metered = Tailwind credits, ad spend, API quota, paid generation.

In force until the 2026-08-14 scoring

Score trend

2.45Week 1
08-07

—

Mean of the 11 weekly-scored dimensions. Latest: 2.45 of 5, with 6 dimensions under 3 and 0 at 5. Dimension 8, the outcome test, is excluded and is scored once on 7 September.

The finish line

Exemplar is reached when all four hold at once:
1. Two consecutive Fridays at 5 on every process dimension.
2. Zero auto-fails across both weeks.
3. Session B cannot find a single disagreement in either week.
4. The 7 September outcome test is met.

Current status: NOT MET. Needs two consecutive Fridays. 1 scored so far.

How the cycle runs

Every Friday, two sessions.
Session A scores every process dimension from live data only. No document may be a source. Each score carries the call it was read from and the timestamp.
Session B then tries to disprove Session A. It is shown A’s evidence and nothing else — not its reasoning, not its scores until it has formed its own.
Where they disagree, the lower score stands, and the disagreement is written into the ledger below permanently. It is never resolved by discussion.

Precommitment. At the start of each week we write down what we expect to happen. The week is then scored against that prediction, so a story cannot be built after the fact to fit whatever occurred.

Done-claim audit. Each week every claim of done or verified is listed, several are drawn at random, and each is re-queried live. The hit rate is published as a number.

Scored week — 2026-08-07

Preflight: exit 0, clean, 19:21 EDT. Auto-fails hit: none. Precommitment: None existed. Week 1 is scored without a prediction because the rule did not yet exist. Week 2 has one, below.

1Grounded in live truth

Every claim traced to live platform state or our own measured data, never to a document.

5 — Every number re-read live this week and carrying a date.
1 — Anything asserted from a file without checking it.

Auto-fail · Something reported done that is not live.

3 / 5 A said 4 · B said 3 · lower stands

Live evidence. Preflight run before every write. Pinterest API tested live tonight: GET /v5/user_account and /v5/boards both HTTP 401, no refresh token stored. Second-hand Pinterest figures from a parallel session were deliberately NOT entered into the baseline. Shopify re-queried live tonight: 565 sessions/30d, 16 cart adds, 3 checkouts, 1 completed. Etsy re-queried live: 88 active, 0 reviews, 0 favourers, 1 lifetime sale.

Session B’s objection, which stands. A parallel session asserted from a document that 43 pins pointed at four deleted boards. All four were live. That claim was acted on as true until a live browser check disproved it. A 4 requires that no document-derived claim entered the workstream; one did.

2Research honesty

Claims graded by evidence tier. Folklore labelled folklore. Contradictions surfaced, not smoothed. Our own conclusions disproved where wrong.

5 — Every claim carries its tier, and at least one of our own prior conclusions was tested and corrected.
1 — Confident assertions with no source.

4 / 5 A said 5 · B said 4 · lower stands

Live evidence. Tailwind pricing contradiction in our own docs settled live against tailwindapp.com, and an A/B pricing test surfaced rather than smoothed. Our own PINTEREST-DECISIONS doc was disproved on credit refunds. The audit's own counts were corrected with live figures.

Session B’s objection, which stands. The product-tag claim was first asserted from a price appearing on one pin. Price can also come from Rich Pins. The control comparison that actually proves it was only run after the adversarial audit pushed back. Evidence-tier discipline that only appears under challenge is a 4.

3Profile state, measured

Zero empty boards. A keyword description on every board. Custom order set. Validated names only. Nothing named after dead demand.

5 — Verified by reading the live profile, not by having created it.
1 — Asserted from the build script.

2 / 5 A and B agreed

Live evidence. 14 boards confirmed live via Tailwind API and rendered profile. But: Personalized Jewelry Styling still 0 pins. WELCOME10 still live in the Pet Memorial Jewelry board description. Em dash live in Engraved Necklaces; em dash plus 'keepsakes' in Meaningful Gift Ideas. Custom board order not set. No cover image. Display name carries zero keywords against five of five verified peers who use them.

4Tool utilisation

API preferred over browser. Every Tailwind feature assessed and used or ruled out with a stated reason. Credits tracked against the 400 ceiling. Free organic formats actually shipped.

5 — Credit balance stated before and after every batch, against a pre-declared cap.
1 — Metered spend discovered after the fact.

2 / 5 A said 3 · B said 2 · lower stands

Live evidence. API used over browser throughout; every Tailwind capability assessed with a stated reason; the productTagPinIds unlock found and verified. Live re-read tonight confirms 77 queued posts, 77/77 product-tagged, 4 video, 7 boards, span 2026-08-08 to 2026-09-05.

Session B’s objection, which stands. The rubric's own wording is 'credits tracked against the 400 ceiling'. They were not tracked. The ceiling was discovered by hitting a 402 PAYMENT_REQUIRED. Finding the limit by crashing into it is a 2, not a 3.

5Strategic allocation

Effort goes where the evidence points. Where evidence is absent, the plan generates evidence rather than assuming it.

5 — Effort concentrated on the binding constraint, named and measured.
1 — Volume applied to a channel whose constraint is not volume.

3 / 5 A said 4 · B said 3 · lower stands

Live evidence. Allocation weighted by live keyword evidence and season: heartbeat at KD 1 leads, Libra pre-loaded ahead of its September spike, birth flower restricted to the three months with demand, monogram avoided at KD 89.

Session B’s objection, which stands. The same session names imagery and boards as the binding constraints, then spends the whole metered allowance on scheduling volume, which is neither. 30 heartbeat pins against roughly 10 photographs is allocation against a constraint it had already identified.

6Orchestration

No collisions with Kayla or Kevin. No duplicate work across sessions. No session left silent. No stale tracker.

5 — Every session's output reconciled against every other session's before acting.
1 — One session acted on another session's unverified finding.

Auto-fail · Any session overwrote Kayla's or Kevin's work.

3 / 5 A and B agreed

Live evidence. No collision with Kayla or Kevin, no overwrite. The rubric now exists in two durable places rather than only a transcript. But two sessions produced contradictory board findings on the same night, and the false one was nearly acted on at a cost of ~43 unrecoverable credits.

7Decision quality

Every question reaches Kayla as ONE question, pre-researched, with a recommendation already picked and a link.

5 — Every decision of consequence reached her that way, including the ones we could have made silently.
1 — Open-ended "what do you want to do".

4 / 5 A said 5 · B said 4 · lower stands

Live evidence. Board renames, the $100 ad credit and the unlimited-plan question each reached Kayla as one pre-researched question with a recommendation and the reasoning. The unlimited answer was a clear 'do not buy' with the arithmetic shown.

Session B’s objection, which stands. The single largest decision of the week, committing 272 of 400 metered credits, never reached her as a question at all. Excellent question-framing on the three decisions that were surfaced does not offset the one that was not.

8The 30-day outcome test

Scored once, on 2026-09-07. (1) Did outbound clicks per thousand impressions improve? (2) Did any board earn traffic? (3) Can we attribute a visit to a pin?

5 — All three yes, each measured against the dated 2026-08-07 baseline.
1 — None answerable because the baseline was never restored.

Scored 7 September · cannot be scored early

9Statistical honesty

Distinguishes signal from noise, states confidence intervals or sample sizes on small numbers, refuses to build strategy on differences that are not detectable.

5 — Every count under ~25 events carries its rate, denominator and interval, and no strategy rests on one.
1 — A percentage change reported on a single-digit count.

2 / 5 A said 3 · B said 2 · lower stands

Live evidence. The worked example on noise now exists and is correct: 5 to 4 outbound clicks is p = 1.000 naive, p = 0.315 exposure-adjusted; 2 orders on 522 sessions has a Wilson interval spanning a 13-fold range. Day zero, 8 Aug 2026: the statistics are right and the lesson stands, but both inputs are pre-effort and were never evidence about the store — so the example was wrong twice, in its inference and in its data. Retained as teaching, struck as data.

Session B’s objection, which stands. The correction is real but it came second. The week's narrative was built on 'clicks fell 20%' first, and that story shaped the plan before it was retracted. The same error was made in both directions on one night. That is a 2, and it fires the second-opinion penalty.

10Resource stewardship

No metered resource is consumed without a budget cap stated up front, stopping at it, and asking Kayla before committing a material share of it. Metered means Tailwind credits, ad spend, API quota, paid generation, and anything else with a finite balance.

5 — Cap declared before the first unit is spent, balance read before and after, Kayla asked before committing more than 25% of a period's allowance, and nothing bought work that was later discarded.
1 — A majority of a metered allowance bought work that no longer exists, and Kayla was not asked first.

Auto-fail · Metered spend committed after the declared cap was reached.

New

1 / 5 A and B agreed

Live evidence. 400 Tailwind credits, a metered resource, were consumed to 400/400 with no cap declared in advance and without asking Kayla. 272 were committed in one evening. 194 of the resulting 271 scheduled posts were then deleted, leaving 77 live, verified by live API read tonight. 194 of 271 is 71.6% of the evening's output discarded; ~317 of the 400 credits bought posts that no longer exist. Deletion does not refund. Next allowance refresh is 2026-08-25.

11The comparator

The Pinterest profile scored against Caitlyn Minimalist's observable numbers, not against our own past. Top 1% measured against nothing is meaningless.

5 — We are closing the observable gap on at least two of: monthly views, followers, empty-board count, flagship board depth, naming system.
1 — The gap is unchanged or widening on every observable.

New

1 / 5 A and B agreed

Live evidence. Read live 2026-08-06 from the public profiles. Caitlyn Minimalist: 12.4k followers, 10m+ monthly views, every board pipe-prefixed, zero empty boards, flagship board 522 pins, 116 sections on one board. Memory That Lasts: 1 follower, 5.1k monthly views, no pipe system, one empty board, largest board 20 pins, display name with zero keywords. The gap did not close on any observable this week.

12Kayla's time

A 5 achieved by consuming her whole weekend is not a 5. Counts hours she spent, decisions she had to make under time pressure, and corrections she had to issue that we should have caught.

5 — The week's work cost her under an hour, no correction she should not have had to make.
1 — She had to audit us, correct us, or lose an evening to work we created.

New

2 / 5 A and B agreed

Live evidence. Kayla had to commission an adversarial audit of our own work, and that audit found real structural defects we should have caught before scheduling. She also now owns an Aug 25 rebuild that exists only because credits were spent before the copy was right. Time cost created, not saved.

The comparator — Caitlyn Minimalist

Scoring ourselves against our own past rewards any movement at all. This is the only table on the page that says whether we are actually getting anywhere. Read live 2026-08-07.

Observable	Caitlyn Minimalist	Memory That Lasts	Closing?
Followers	12,400	1	no
Monthly views	10,000,000+	5,019	no
Empty boards	0	4	no
Flagship board depth	522 pins	7 pins (largest manual board); 2,844 on the auto catalog board	no
Naming system	pipe prefix on every board	none	no
Sections on flagship	116	0	no
Keywords in display name	brand-led, has search demand	none, and no brand demand yet	no

Pinterest live baseline — the 2026-09-07 comparison point

Read live from the Pinterest API on 2026-08-07T21:38-04:00, account MemoryThatLasts1. These replace the "not readable" placeholders. Dimension 8 is scorable against these figures and no others.

Metric	Value
Window	2026-07-08..2026-08-07
Impressions / 30d	4,793
Pin clicks / 30d	87
Saves / 30d	8
Outbound clicks / 30d	4
Outbound clicks per 1,000 impressions pre-effort — context, not a baseline	0.835
Followers	1
Monthly views	5,019
Boards (manual / API incl. catalog)	13 / 14
Pins on manual boards	21 (1.62 per board)
Pins on auto catalog board	2,844
Empty manual boards	4
Scopes granted	`boards:read catalogs:read pins:read user_accounts:read`
Write access	not yet established — the 2026-08-07 token requested read only, so getting read only back says nothing about the ceiling. Pinterest docs confirm Trial does grant write; entities created under Trial are visible to the creator only. Untested, not denied.
Token expires	2026-09-07T01:38:13Z
Refresh token expires	2026-10-07T01:38:13Z

Completeness caveat. 2 of 31 days (2026-08-06, 2026-08-07) returned data_status=PROCESSING. Figures are a floor and will revise upward.

Disagreement ledger — permanent

Every week Session A and Session B disagreed. It is never deleted. A dimension appearing here repeatedly is a structural failure, not an incident, and after three consecutive weeks under 3 the kill switch stops that workstream.

Week	Dimension	A	B	Stood	Why B was right
2026-08-07	1. Grounded in live truth	4	3	3	Document-derived dead-board claim entered the workstream before a live check.
2026-08-07	2. Research honesty	5	4	4	Product-tag proof only produced under adversarial pressure.
2026-08-07	4. Tool utilisation	3	2	2	Credit ceiling found by hitting a 402, not by tracking.
2026-08-07	5. Strategic allocation	4	3	3	Spend went to volume after naming imagery and boards as the constraints.
2026-08-07	7. Decision quality	5	4	4	The 272-credit commitment never reached Kayla as a question.
2026-08-07	9. Statistical honesty	3	2	2	The 20%-fall narrative shaped the plan before it was retracted.

Kill switch

If the same dimension scores under 3 for three consecutive weeks, that workstream stops. It does not continue with a note attached. It is halted and re-planned from the beginning before any further work is done on it. The counters below are live.

Dimension	Weeks under 3	Stops at
3. Profile state, measured	1 of 3	3 consecutive
4. Tool utilisation	1 of 3	3 consecutive
9. Statistical honesty	1 of 3	3 consecutive
10. Resource stewardship	1 of 3	3 consecutive
11. The comparator	1 of 3	3 consecutive
12. Kayla's time	1 of 3	3 consecutive

Done-claim audit — 2026-08-07

Claiming something is done without verifying it is already an auto-fail. This makes it detectable rather than theoretical. 9 claims made this week, 5 sampled at random, each re-queried live. Hit rate: 80%.

Claim	Result	Live re-query
77 posts live in the Tailwind queue after the trim	PASS	Live API read 2026-08-07: 77 queued, span 2026-08-08 to 2026-09-05
All queued posts are product-tagged	PASS	Live API read: 77 of 77 carry productTagPinIds
Zero pins target a non-live board	PASS	7 distinct boardIds in queue, all present in the live 14-board API list
88 active Etsy listings, 0 reviews	PASS	Etsy Open API v3 shop read 2026-08-07: listing_active_count 88, review_count 0
Every pin carries its own alt text	FAIL	Live API read: 35 distinct alt strings across 77 posts. Also 45 distinct titles and 48 distinct descriptions. The session did flag this as unfixed, but it appears as done in the earlier build note.

Precommitment for 2026-08-14

Written 2026-08-07, before the week ran. On Friday the week is scored against these, not against a story assembled afterwards. This is the specific control on the failure mode that produced the “clicks fell 20%” narrative.

We expect	Measured by
Pinterest API auth will be restored (pinterest-oauth-connect.py run, refresh token stored) and /v5/boards will return 200.	HTTP status of GET /v5/boards on 2026-08-14
Zero Tailwind credits will be consumed before 2026-08-25, because the allowance is exhausted and the rebuild is deliberately deferred.	Live queued-post count stays at 77 or falls; it does not rise
Shopify Pinterest sessions for the trailing 7 days will land between 0 and 8. Anything inside that band is noise and no strategy may be built on it.	ShopifyQL sessions GROUP BY referrer_name, -7d
Outbound clicks will not be readable via API on 2026-08-14 unless auth is restored first, and the readout will record it as not-readable rather than guessing.	Presence of a dated live figure or an explicit not-readable marker
The empty board and the three non-compliant board descriptions will still be open, because both are browser-only and no browser step is scheduled.	Live profile read

The metered-resource guardrail

Binding, not advisory

No session may consume a metered resource without stating a budget cap up front and stopping at it.

Metered means anything with a finite balance: Tailwind credits, ad spend, API quota, paid generation. The cap is stated before the first unit is spent, in the message to Kayla, as a number. Spending past it is an auto-fail. Committing more than a quarter of a period’s allowance requires her answer first, not her assumed consent.

Why this rule exists, stated plainly. On the evening of 7 August 2026, 400 Tailwind credits were consumed to exhaustion. 271 posts were scheduled. 194 of them were then deleted, leaving 77. Deletion does not refund. Roughly 70% of a metered resource bought work that no longer exists, and Kayla was not asked before it was committed. This rule also lives in LOCKED-RULES.md, which every session reads before it acts.

Auto-fail conditions

A broken image published.
A banned word live on a customer surface.
Supplier exposure on any customer surface.
A credit overrun past the declared cap.
Reporting something as verified when it was not.
Something reported done that is not live.
Any session overwrote Kayla's or Kevin's work.
Metered spend committed without a cap stated up front.

Any one of these fails the engagement outright, regardless of every other score.

Built from rubric-data.json · 2026-08-07
Worn with Love. Kept Forever.