Lukehle Degradation
The gate never gets skipped for lack of capability. It degrades, and it says which tier it ran at.
This matters because a managed or restricted seat may block the Workflow tool, subagents, or both — and the temptation is to skip the review entirely rather than run a weaker one. A weaker gate honestly labelled beats no gate.
Convention, not measurement. The three tiers are a portability design. The corpus measured the Tier 1 shape; Tiers 2 and 3 are reasoned degradations of it. See CALIBRATION.md.
Tier 1 — parallel subagents
The full shape. Miners run concurrently and blind to each other; verifiers run per finding; judges see only the confirmed dossier.
What makes it strongest: genuine independence. Reviewers who cannot see each other's work do not converge, and non-convergence is what produces coverage.
Requires a workflow/orchestration tool. Use workflows/quorum.js or scaffold one with
tools/scaffold.mjs.
Tier 2 — sequential subagents
Subagents available, no parallelism. Same lenses, run one at a time.
The rule that preserves most of the value: each pass gets a fresh context and does not see earlier findings. Independence, not concurrency, is what Tier 1 was buying. A sequential run that hides prior results retains most of the rigour; one that accumulates them collapses into a single perspective that agrees with itself.
Costs versus Tier 1: wall-clock time, and some risk of orchestrator context bleed between passes.
for each lens:
spawn agent with ONLY the lens question + scope + evidence contract
collect findings
(do not pass previous findings in)
dedupe in code
for each finding: spawn a refuter with ONLY that finding
tally in code
judges: one pass each over the confirmed dossier
synthesize
Tier 3 — structured single context
No subagents at all. You run every lens, explicitly and one at a time.
This is meaningfully weaker and must be labelled as such. It is also far stronger than one undirected read, because naming the lens converts a vague "review this" into a specific search.
The discipline that makes it work:
- Write the lens question down verbatim, before looking at the subject. Committing to what you are hunting prevents the search from drifting toward what you happen to notice.
- Hunt only that failure. Do not fix anything, do not assess overall quality, do not jump to a different concern that catches your eye — note it and return to the lens.
- Record findings before switching lenses, including "nothing found for this lens." The empty result is data.
- Never skip a lens because an earlier one found something big. Lenses are orthogonal by construction; a major finding in one says nothing about the others.
- Verify in a separate explicit pass, re-opening the evidence, with the refutation framing. Do not verify while mining — that is the fold-verify-into-judge failure the corpus identified as the difference between strong and mediocre runs.
- Rule at the end, against the pooled list rather than as you go.
The order is doing the work here. A single context that runs mine → verify → rule as distinct passes retains the structural benefit even though it loses independence.
What each tier costs
| Independence | Adversarial verify | Panel | Reproducible | |
|---|---|---|---|---|
| Tier 1 | Full | Per finding, parallel | Yes | Yes, scripted |
| Tier 2 | Good (fresh contexts) | Per finding, sequential | Yes | Mostly |
| Tier 3 | None — one mind | Separate pass, same mind | Self-administered | No |
Tier 3's honest limitation: you cannot fully refute your own finding, because you already believe the reasoning that produced it. It catches evidence that does not reproduce; it catches reasoning errors much less reliably.
Say which tier ran
Every quorum output states its tier and that tier's limitation. A reader must know how strong the gate was.
Tier 2 (sequential, fresh context per lens) — parallel subagents are
unavailable on this seat. Reviews were independent but not concurrent.
Tier 3 (structured single-context) — no subagents available. Four lenses were
run as separate explicit passes with a separate verification pass. This is
weaker than independent review: the same reasoning produced and checked the
findings.
Omitting the tier lets a Tier 3 pass be read as a Tier 1 result, which is the one outcome this whole skill exists to prevent.
Deterministic steps never degrade
Dedup, tallying, ruling, ranking and requorum diffing are code at every tier. tools/quorum-lib.mjs
has no harness dependencies — it runs anywhere Node runs, and by hand in a spreadsheet if it must.
Keeping these in code is what stops Tier 3 from also losing reproducibility in its bookkeeping on top of losing independence in its judgement.
State is checkpointed at round boundaries, not inside a run: a run that dies mid-phase re-runs
its round, and what persists between rounds is the returned ruling JSON (see /requorum for where it
lives). There is no intra-run store on purpose — agent state is harness-owned and not this repo's to
serialize.
Choosing a tier
Workflow/orchestration tool available? -> Tier 1
Subagents available? -> Tier 2
Neither? -> Tier 3
Do not use a lower tier for convenience on high-stakes work. If the stakes justify a quorum and only Tier 3 is available, run Tier 3, label it, and say plainly what a stronger gate would have added.
Related skills
quorum— the orchestratorquorum-verification— the phase most degraded at Tier 3quorum-lenses— naming the lens is what makes Tier 3 work at all