Notebook entry · 9 November 2026

A completed quality pilot, with limits

Independent review corrected a reproduced regression missed by self-check. The skill did not establish an added quality advantage over manual coordination. All required grading is complete; the interrupted attempts and methodological limits remain visible.

9 November 2026 · Recovered grading completeSame draft, three finishing paths

Six authored public-project tasks each used one genuine initial draft, frozen once and copied identically to A: fresh same-model self-check; B: manual independent review; and C: explicit $agents-collab with matched models. Tasks, rubric and success criteria were frozen before drafting. All six drafts and 18 finishing runs were retained.

The first six anonymous grading calls hit a subscription limit and returned no assessments. That left the experiment unfinished; its interim presentation was withdrawn. After access returned, a reviewed amendment resumed only the missing grades. All six recovered calls completed, without changing participant outputs or frozen criteria.

No added skill quality advantage established over manual coordination

Both review paths fixed a script-invocation regression that self-check missed. Two concrete test weaknesses count minor; C corrected a guardian fixture that B retained. Maintainability ratings are descriptive and never counted as task wins.

The primary advantage is not robust. An untargeted failure of an existing expiry test supplies the second initially material draft required by the rule. A disclosed sensitivity excluding existing-test failures still has two review-versus-self-check wins, but only one initially material draft, below the frozen floor. No general quality, speed, cost or equivalence claim follows.

One task also has conflicting runtime requirements. Recovered grading challenged the earlier operator-only safety claim about quarantine deletion. That runtime claim remains unconfirmed, and its ambiguous checker is excluded consistently. Original failures and earlier observations are retained with corrected dispositions.

Fresh labels, restricted file tools and a read audit were used because old labels had previously been public. This reduces accidental unblinding without proving it impossible. Two assessment responses needed deterministic extraction of a complete JSON object after a prose preface; no model was recalled for that formatting issue.

This study held handoffs fixed. It did not compare alternative packet formats or optimize communication between models. The distributable skill remains prompt-led; the study harness is separate.

Full completed report, criteria and evidence · All 24 preserved artifact states

9 November 2026 · Public sourceActivation and routing are separate claims

We published the protocol and inputs before outcomes, then ran three authored engineering tickets against disposable copies of public ProcLease commit 37975b5. Its project instructions had no model-collaboration policy. The active project, global instructions, installed skill and personal profile were unchanged.

Ordinary gate: 1/3 routes · 36-run comparison not started

The separate diagnostic added only Use $agents-collab for this task. No model names or role assignments were added. All three sessions loaded the skill. The writing case selected claude-opus-5-5 with an accepted High request, then gpt-6-astra / High; the consequential case selected the same reviewer. Native metadata recorded distinct, non-forked ephemeral child sessions. These are interface observations, not runtime effort attestation or filesystem isolation.

External formatting, Clippy and full-suite checks passed for all six outputs. Mutation, filter, compatibility and reverted-fix probes also passed where applicable. Those auxiliary checks were not a blinded full quality score. The symlink defect was deliberately seeded for the ticket; the public ProcLease base already refuses it.

Use supported by this sample: invoke the skill explicitly and let the profile choose collaborators. Ordinary prompts did not activate the required routes here. Three task units, one execution per mode, concurrent local sessions and no disabled-skill comparison cannot establish quality, speed or cost gains. Billed cost is unknown.

Read the full report, times, checks and limits · All 42 run records · Selected sanitized receipts

The public package contains the frozen tickets, seeds, profile, measurement harness and final patches. Raw conversations and account details are not published, so the original runs are not independently auditable end to end. The 36 unstarted comparison rows remain visible with null measurements.

Earlier private pilots · 8 November 2026

Each pilot had a rule fixed in advance: the skill should cut coordinator time or coordinator command count by at least 20%. Reviewers saw unlabeled bundles, but could plausibly infer style; blinding was not guaranteed.

Conditions · 8 November 2026Who did what, at what requested effort

Both pilots used the same roles and requests in both arms, and every output got a paired assessment from both independent reviewers.

“High” was requested and accepted by the interface; independent runtime attestation of the effort actually applied was not established. Writer usage for the timed-out baseline correction is unknown. These were the conditions of these trials on that date, not recommended defaults. The skill and its examples stay model-neutral.

00 · Before the pilotsStructural validation

Privately, 52 unit tests passed for the validator and package assets, and nine simulated initial scenarios plus two focused follow-up scenarios passed. They covered solo work without Git, unsupported required effort, exhausted review caps, resuming a running assignment, delegation restrictions, unenforceable deadlines, records overlap and changed review inputs. Two independent reviewers finished one initial round and one focused follow-up with no remaining material finding.

Capability responses in those scenarios were simulated. That shows a coherent procedure, not production safety. The public test suite drops one private-profile test and adds repository checks; its result will be reported when it is run.

01 · Additive documentation pilotAdding the skill on top of an existing workflow

Five unfinished documentation and decision-support tasks from one project, each done twice: the existing workflow, and the same workflow plus the skill and a private profile. Writers, reviewers, sources, acceptance criteria, word limits and round caps were matched; writer order alternated.

Coordination threshold: not met

Coordinator time and commands both rose. Material findings in both arms sat in a single task and were addressed after one bounded correction round. Because the skill was added alongside existing rules, the benefit that motivated it, removing duplicated policy, was not measured.

02 · Replacement code-task pilotReplacing cleaned rules with the skill

Five real implementation tasks in one application, each from the same frozen code. The baseline kept the existing instructions; the skill arm used cleaned copies with personal model clauses removed, plus the skill and a private profile. Models and reasoning requests were explicitly matched, and both arms got the same paired independent review.

Coordination threshold: not met

* Not a productivity claim. One baseline correction hit its supervised 480.510-second deadline and was stopped, leaving an incomplete usage receipt. Identical tool allowlists gave the two arms different permission outcomes for generated in-writer test commands, so their check workloads differed. With one execution per related task pair, these confounds rule out crediting the skill with the apparent reduction.

• Final readiness required bounded repairs in both arms. A pre-existing integration race in the application stayed open; not every product risk was resolved.
• The skill arm’s coordinator followed the profile’s single-review default; the parent imposed the paired review. Automatic or named-recipe routing was not established.
• Moving personal policy out of project files did not break the workflow, and project-facing handoffs passed a limited scan for model and private-path details. That scan is not a privacy guarantee.

03 · After the pilotsInstallation and discovery

The unchanged package was then installed for personal use with an existing private profile. A fresh session’s skill catalog listed it, the profile passed structural validation and installed file hashes matched. That verifies discovery and structure at the time, not future routing, automatic selection for every task, or runtime effort.

04 · ReadingWhat the evidence supports, and what it doesn’t

How this fed into the cleanup Try it yourself