Pattern-069: Coarse Triggers Causing False-Positive Triage Cost

Status

Proven — Promoted from Emerging 2026-08-25 by CIO. Filed 2026-05-11 by CIO under self-approval authority per methodology-audit-policy-updates-2026-03-16.md, with PM directive to close the loop on the May 10 disposition (elevate from “tactical observation” to formal Emerging). Slot renumber 068→069 same session as cascade from Pattern-068 (Silent State Mutation) renumber per Lead Dev + Architect coordination — see Pattern-068’s Status note for full slot-conflict-resolution context. Surfaced by Code agent’s May 10 PreCompact-hook second-incident addendum; HOST May 10 concurred on naming, deferred proto-pattern-vs-tactical-observation call to CIO.

⚠️ The original promotion window (2 weeks from filing, ~May 25) lapsed unresolved — nobody checked it, including the pattern’s own author. Found 2026-08-25 during an unrelated innovation-backlog audit: the Status field had sat unrevisited for three months past its own stated deadline, the exact “a tracker line is a claim about the world” staleness this cohort has repeatedly caught elsewhere this month (cio-standing-items.md’s 08-23 sweep; the freeze-watchdog false-alarm chase itself, below). Promoting now on the strength of real evidence found in hand, not on the calendar deadline having passed.

Evolution: 2026-08-25 — cross-mechanism recurrence confirmed, via the freeze-watchdog (CIO). The 2026-08-17 escalation (dev/active/cio-to-host-exec-watchdog-pattern-2026-08-17.md) is an independent instance of exactly this shape, on a mechanism with zero code relationship to the PreCompact hook: 5 stall alerts across 4 of 6 days, 100% self-resolved by the time anyone read them — 2 confirmed resolved in 3-6 minutes via the affected role’s own next heartbeat, 3 more confirmed resolved by that evening’s STOP. The watchdog’s verdict was correct every time (the role really was stale at the moment of measurement); the failure was that the alert carried no signal about how likely self-resolution was, so triage cost was paid identically whether the stall was real or about to evaporate on its own. This satisfies the promotion criterion precisely — a genuinely different hook/mechanism (liveness detection, not compaction safety) producing the same “correct detection, unweighted stakes, compounding triage tax” shape — just discovered three months after the window that was supposed to check for it.

Product Relevance

Process-only — Discipline-mechanism design property. Piper’s users will not encounter this; teams building agent-coordination infrastructure with automated guards (hooks, gates, validators) will.

Context

When a discipline mechanism — a hook, a gate, an alert, a validator — fires correctly by its own internal logic but the actual stakes of the situation it detected are low, the cumulative triage cost (human attention or PM-helper sessions burned verifying the alert) compounds faster than the mechanism’s load-bearing-catch benefit. The failure is not in what the mechanism detects; it is in how it weights what it detects.

This pattern is distinct from the triggering-failure-mode patterns (062 family, 045, 046, 047, etc.), which name failures the mechanism is designed to catch. Pattern-069 names a failure of the mechanism’s design — specifically, the absence of severity tiering or context-sensitivity in what would otherwise be a working detector.

Problem

The Failure Mode

Discipline mechanism M is designed to detect condition C
M is calibrated to fire whenever C is present
C has a wide variance in actual stakes (sometimes high, sometimes near-zero)

Scenario A (load-bearing): C is present, stakes are real → M fires correctly → triage produces value
Scenario B (false-positive): C is present, stakes are near-zero → M fires correctly → triage produces no value, but costs the same attention

Over time: B incidents outnumber A incidents
Cumulative triage cost > cumulative caught-real-problem benefit
Net effect: M is now a tax on the cohort, not a guardrail

The mechanism’s verdict is correct each time — that’s what makes the failure subtle. M doesn’t have bugs; M doesn’t fire spuriously; M’s logic doesn’t drift. The failure is that the calibration was set at “detect the condition” without an additional layer of “weight the condition by stakes.”

Why It Happens

  1. Detection is easier to design than decision support. Binary “does the condition apply: yes/no” is straightforward. Tiered “given the condition, what’s the appropriate action and severity” requires modeling the stakes-variance distribution, which often isn’t available until after the mechanism has been deployed and observed.

  2. First-incident validation creates anchoring. When M’s first fire catches a real problem (as PreCompact hook did on Docs’s stranded log May 9), the mechanism’s correctness gets validated against high-stakes evidence. Subsequent fires are interpreted through that anchor; the cost-curve drift isn’t visible until enough false-positives accumulate.

  3. No standing review of mechanism cost-curves. Most disciplines have a “this mechanism caught a real problem” feedback loop but no “this mechanism is costing more than it saves” feedback loop. The first is celebrated (visible in retros, narratives); the second is invisible until someone names it.

Concrete Example: PreCompact Hook (May 9–10, 2026)

PreCompact hook deployed May 9 (Lead Dev) — fires on uncommitted changes before /compact, blocks with options.

Fire #1 (May 9 evening, Docs session): Load-bearing catch. Docs had stranded session log + Janus memo in working tree; session was about to compact; uncommitted work was at real risk of session-end loss. Triage produced the cross-agent residue accumulation pattern candidate. Value > cost.

Fire #2 (May 10, PPM session ~3h later): False-positive triage. PPM session was local CLI (files on disk, will survive compaction); 6 uncommitted files were 4 MANIFEST regen + 2 PPM-owned drafts; no actual loss risk. Triggered ~30 min PM-helper session detective work to verify. Cost > value.

Two fires, one day, opposite cost-benefit math. The hook’s logic was correct both times; the weighting was identical despite stakes being radically different. The Code agent author’s proposed refinements (locality differentiation; severity tiering; “safe to compact” explicit option) move the hook from pure detection toward decision support without rolling back the load-bearing fire-#1 catch.

Solution

At the methodology layer

1. Distinguish detection from decision support at mechanism design time. When proposing a new hook/gate/alert, name explicitly which stance is intended. Detection mechanisms have predictable behavior (good for trust) but treat every instance with equal weight (bad for attention budget). Decision-support mechanisms tier severity to stakes, preserving trust while respecting attention.

2. Build cost-curve feedback into mechanism rollout. A new discipline mechanism should have a planned ~5-fire (or ~30-day) review checkpoint: did the fires produce value at a cost the cohort tolerated? If false-positive-to-load-bearing ratio exceeds ~2:1 at the checkpoint, refine the trigger criterion before the mechanism becomes a tax.

3. Reserve “tiered severity” as a refinement path, not a v1 requirement. Most mechanisms can start as detection-only; severity tiering is added after observation reveals the stakes-variance distribution. The May 9 PreCompact-hook v1 was correct to ship as detection-only; the May 10 second-incident is the signal that refinement is warranted.

At the tooling layer (PreCompact hook specifically)

Routed to Docs per HOST’s May 10 framing (Docs owns the hook script). Refinement options ranked by leverage:

  1. Locality differentiation (highest-leverage): hard warning for remote/sandboxed sessions (real loss risk); soft reminder for local CLI sessions (rediscovery cost, not loss).
  2. Explicit “safe to compact” path (documentation-only): add a fourth pick-one option for users whose situation is benign.
  3. Severity reduction for known-safe patterns: MANIFEST regen, .DS_Store, gitignore noise as “tidy-but-not-critical” rather than “substantive-and-stranded.”

Anti-Pattern Indicators

The following signals suggest Pattern-069 may be present:

The last is the strongest late-stage signal: when a mechanism’s fires are increasingly ignored, the cost-curve has already inverted; the mechanism is no longer functioning as a guardrail.

Cross-References

References


Formalized: 2026-05-11 by CIO. PM ratification on May 11 same-day directive. Promoted to Proven 2026-08-25 by CIO on cross-mechanism recurrence evidence (freeze-watchdog, 08-17) — three months after the original two-week window, found during an unrelated tracker audit rather than a scheduled check.