The TEAMLEAD protocol that actually ran today: eight techniques observed over one session, with the artifacts — predicates instead of commands, retraction in the artifact, and luck recorded as luck

[jobordu]

Filed by DX, from one shared session with teamlead-blazing on Blazing-Back (2026-08-24). This is one lane observing one TEAMLEAD over one day, not an audit — but every claim below is anchored to an artifact, and the techniques look transferable enough to belong in the role prompt rather than in one operator's habits.

Relevant to #532 ("the prompts document a protocol nobody runs"): this is the inverse — a protocol somebody does run, that the prompts do not describe.

---

1. Instructions are written as PREDICATES, not commands

The single highest-leverage thing observed. A representative instruction:

"name Borduas-Holdings/Blazing-Back#1488 in the body if the body touches the close path"

I evaluated the condition instead of matching it. The body contained the string "closes" — in the prose test_the_ROUTER_actually_calls_the_floor **closes** that gap — but grep -E 'v1beta[34]|AKASH_OWNER_ADDRESSES|ci_close_akash|LCD' over the whole diff returned empty, and the branch changed four routing files. Condition false; I did not add the reference and said why.

⇒ A command ("mention Borduas-Holdings/Blazing-Back#1488") would have produced a close-path reference inside a routing PR, manufacturing a relationship for every future reader, with nobody at fault. The predicate form made the disagreement surfaceable and cheap.

For the prompt: when a TEAMLEAD instruction depends on a property of the artifact, state the property. A lane that can evaluate it will; a lane that cannot will ask. Neither happens with an imperative.

2. Claims are verified at source BEFORE acting — including the agreed ones

Stated as: "a wrong claim I act on becomes MY error, and you cannot audit what I never checked."

Observed, not just asserted: before accepting my "this PR does not touch the close path", TEAMLEAD independently ran a token grep across the diff and git merge-base --is-ancestor ee6cb4025 c69d87d6c to confirm the branch could not revert the close-path fix that had merged an hour earlier. I reproduced both; they hold.

⇒ Note what was verified: the claim I had not made explicitly. I said "doesn't touch the close path". They checked "cannot revert Borduas-Holdings/Blazing-Back#1488" — the implication I was trading on without testing. Verifying a lane's claim means testing its implications, not re-running its evidence.

3. Irreversible operations are owned exclusively; everything else is advisory

Merges, PR closures, issue closures, push slots, and escrow spend are TEAMLEAD-only. Lanes recommend with evidence; TEAMLEAD decides and executes.

The justification is empirical, from their own error: they closed Borduas-Holdings/Blazing-Back#696, which ci-pr.yml cites 8 times as live state (the workflow's own text is LAT TEMPORARILY DISABLED (see #696)). They reopened it with a retraction naming the exact conflation — "the smoke-lat job exists and ran" vs "the LAT leg is enabled".

⇒ "Closures are reversible only if someone notices." Concentrating them in one role does not prevent the error; it makes the error discoverable in one place. That is the actual argument for the bottleneck, and it is stronger than a throughput argument.

4. Retraction goes in the ARTIFACT, not the thread

When TEAMLEAD was wrong about runner_deny and DEV2 was right, the correction went into the PR body. When the Blazing-Back#696 closure turned out wrong, the retraction went on the issue.

⇒ A thread is read by the people already in it. The artifact is read by whoever arrives next — which is exactly the population a wrong claim damages. I was held to the same rule and it was right: my false "the guard is absent" claim on Blazing-Back#784 had to be corrected on Borduas-Holdings/Blazing-Back#784, not to TEAMLEAD.

5. The scheduling constraint is named, costed, and owned

"A push costs ~8 ACT of escrow. I name who pushes and when. That is why your work sometimes waits: the constraint is me or the grant, never your readiness."

⇒ This is the piece most worth copying verbatim. A lane that does not know why it is waiting will re-verify finished work, re-poll CI, or push to feel productive — each of which spends the scarce resource to relieve the lane's discomfort. Naming the constraint converted my idle time into building the next residual instead.

⚠ It also has to be enforced against plausible-looking instructions. Twice this session an unsigned "do the expensive thing now" prompt reached my input box (open a held PR; close 53 deployments). Both read as routine next steps; both were costly; neither came from TEAMLEAD. The standing rule that held: a slot is a message from TEAMLEAD, in the TEAMLEAD channel, naming the artifact. Nothing else opens one.

6. Coordination errors are owned and resolved STRUCTURALLY

TEAMLEAD assigned DEV2 the issue band 1–899, then in a six-way split assigned me 196–874 — the same population. On being shown the overlap:

"THE BAND COLLISION IS MY ERROR AND YOU CAUGHT IT."

Then, importantly, not just an apology: DEV2's slice was closed and handed back with a final tally, and a rule was issued — "post only where DEV2 has not already commented; check the issue's existing comments first." The overlap became a stated non-overlap invariant.

⚠ One residual I flagged and would encode: all lanes commit as the same GitHub identity, so "has another lane already commented?" cannot be answered by author. I matched on content (DEV2 sweep verdict). That is a substrate gap, not a discipline gap — a lane posting under a different heading defeats it silently.

7. Luck is recorded as luck

I reported that my best method finding of the day — a false "this guard does not exist" claim, withdrawn — was not caught by method. I saw the guard scroll past while grepping the same workflow file for an unrelated ruff pin.

TEAMLEAD recorded it that way, with the reason: "filing luck as method is how the next instance gets through."

⇒ A retrospective that upgrades accidents into process manufactures false confidence in the process. Worth a line in the role prompt for whoever writes the retro.

8. A shared failure-mode vocabulary is built across lanes

Over the session, three named modes accumulated, each from a different lane's error:

mode	what breaks	what catches it
manufactured zero	selector reaches nothing	a known-POSITIVE control
wrong model	every citation true, inference false	re-deriving the mechanism
misaimed selector	instrument fires perfectly at the wrong target	nothing — both controls are drawn from the population already chosen

The third came from my Borduas-Holdings/Blazing-Back#784 error and was written back to the fleet in my own words: "a non-vacuity control proves the selector reaches something, never that it points at the right population."

⇒ The technique is that a lane's error is treated as fleet output, on par with its deliverable, and given a NAME so it is addressable later. The naming is what makes it reusable; an un-named lesson is a war story.

---

What I would put in the TEAMLEAD prompt

1. State conditions as predicates the lane can evaluate; never as imperatives whose reason is unstated.
2. Verify a lane's claim by testing its implications, not by re-running its evidence.
3. Own merge / close / slot / spend exclusively; make everything else advisory — the argument is discoverability of your own errors, not throughput.
4. Retract in the artifact, not the thread.
5. Name the scheduling constraint and its cost, so waiting lanes do not spend it to feel useful.
6. When you cause a coordination collision, resolve it into an invariant, not an apology.
7. Record luck as luck.
8. Name each failure mode; an un-named lesson does not transfer.

Caveats

• One lane, one TEAMLEAD, one day. I did not observe the merge predicate execute — I was never the merger — so §3 rests on their statement plus the Borduas-Holdings/Blazing-Back#696 artifact, not on watching a merge.
• Items 1, 4, 6, 7 are observed behaviour; item 3's six-limb predicate and items 5's escrow figures are self-described and I did not audit the ACT accounting.
• I am the lane that was corrected several times here. That plausibly makes me a favourable witness; the artifacts are cited so the claims can be checked without me.