22 — Missed opportunities and gate calibration

Strict gates protect the floor. The same strictness can hide real waste. Every blocked candidate is kept; soft-gate blocks can reach the owner; a small exploration budget measures what the gates reject; soft thresholds then move on evidence in both directions. Hard gates never enter that loop.

Humans still promote, dismiss, accept, and execute. L4 does not silently turn a block into an emit.

Purpose#

Specify the opportunity ledger, owner backlog, exploration cards, calibration, and how this feeds the offline improvement loop — without weakening hard stops.

Decisions#

#	Decision	Reason
1	Every gate is `hard` or `soft`; every block records gate id	Audit and calibration need a name, not a vibe
2	Hard gates never backlog, never explore, never tune	Sacred floor (`00-kernel.md`)
3	Soft-gate blocks → ledger + owner backlog; optional exploration	Owner keeps final say; exploration removes selection bias
4	Calibration loosens and tightens from ledger + exploration	Learning only from sent cards drifts cautious (`19-failure-modes.md` section 8)
5	Promotion re-runs hard gates	Backlog cannot bypass the kernel

Rejected: loosening hard stops to catch more opportunities; silent promotion; ranking patterns by closure rate alone; council labels as automatic truth.

Would change this: a new soft-gate kind with a registered threshold location; evidence that exploration caps are too high for a plant’s attention budget.

Hard vs soft#

Canonical gate_id table: 00-kernel.md section 11. Summary:

Class	Examples	Tunable?	Backlog?	Explore?
Hard	Hard stops; constraint violated/unknown-hard; LLM possible conflict; unreferenced rupee; write tool; proof floor; simulator envelope; action-seam disagreement; uncertified detector; supersede-after-accept; staleness hard limit	No	No (audit ledger only)	No
Soft	`routing_agreement`; `one_family_confidence`; `evidence_tier_min`; `freshness_margin`; `attention_budget`; `reproposal_cooldown`; `hypothesis_volume_cap`; `pattern_precision`	Yes — registry	Yes	Only if exploration-eligible below

Wrong-looking hard blocks (e.g. stale constraint row) are raised as data. They never loosen a hard gate.

Opportunity ledger schema#

Stored in the L4 operational store (derived data only). One row per candidate that did not become a normal card (includes L3 Evidence withholds, discovery drops, hypothesis shadow, portfolio holds, cooldown suppressions; as built: Finding withholds).

Field	Meaning
`ledger_entry_id`	Stable id
`decision_case_id` / case snapshot	DecisionCase reference
`candidate_snapshot`	Full candidate set (or winning candidate) at block time: claims, citations, footprint, L3-priced effects per section (Modeled), seam records
`ledger_hash` / PSM snapshot id	As-known-at provenance
`lockfile_hash`	Release pin at block time
`gate_id`	Named gate that blocked
`gate_class`	`hard` \| `soft`
`gate_threshold`	Threshold then in force (soft only; null for hard)
`gate_value`	Candidate’s measured value against the gate
`block_terminal`	`withhold` \| `abstain` \| hold \| shadow-drop (as applicable)
`origin`	Finding / pattern / hypothesis / portfolio
`later_outcome`	Join slot — see below
`created_at`	Recorded time

Later outcome (filled as evidence arrives):

Outcome kind	How
`persistence`	After the verification window, code runs the verification plan on what happened; L3 prices foregone effect on actual data (Modeled, per section). Vanished condition → block likely correct
`later_card`	Same condition key later emitted; ClosureState closure (`ops_confirmed` / rejected / no-change) joined back
`owner_promoted`	Backlog promote → card → closure
`owner_dismissed`	Dismiss reason; may seed a proposed constraint
`explored`	Exploration card closure with `exploration=true` on the learning fact
`none_yet`	Still open

Hard-gate rows keep outcomes for audit and constraint-data debugging. They do not enter backlog selection or exploration sampling.

Owner backlog#

Soft-gate ledger rows surface to the named plant owner as an unassigned backlog:

labelled lower confidence;
gate that held the item in plain words;
candidate summary and priced sections (calculator refs only);
actions: promote, dismiss (with reason), or leave.

Promote: re-run every hard gate and the normal decision path; on pass, create a normal card (not an exploration card unless separately sampled). Cannot bypass hard stops.

Dismiss: label for improvement; may become a proposed constraint row for the constraint owner.

Hard-gate items stay hidden from this backlog (amends “UI hides all withholds”: hard stay hidden; soft may show as backlog — ADR-025).

Exploration cards#

Rule	Detail
Opt-in	Per plant, named owner
Cap	Small weekly budget (registry); attention-aware
Exploration-eligible soft gates only	`evidence_tier_min`, `freshness_margin`, `pattern_precision`, `one_family_confidence` (when in one-family mode). Not eligible: `attention_budget` (hold stays hold), `reproposal_cooldown`, `hypothesis_volume_cap`, `routing_agreement`, any hard gate including `action_seam_disagreement`
Selection	Random within eligible set (unbiased sample)
Origin	Preserved (`l3_finding` / `l4_pattern` / `l4_hypothesis`); `exploration=true` is a flag only
Before send	Re-run every hard gate + card minimizer + kernel re-check; fail → do not send
Label	Lower confidence; learning fact carries `exploration=true`
Closure	Same L5 verify / reject / no-change path

Exploration is the only source that shows what eligible soft gates would have missed without selection bias. Ops watches exploration reject/no-change (16-operations.md); a nuisance spike pauses the budget — it does not loosen hard gates.

Per-gate scorecard#

Metric	Definition
Block precision	Share of blocked items truly not worth sending (vanished, dismissed, rejected when explored)
Miss rate	Share that would have helped (persisted with priced waste, later verified, promoted+verified, explored+verified)
Foregone effect	L3-priced, Modeled, per domain section — never summed across sections
Nuisance cost of loosening	Reject + no-change among explored; owner attention used
Trade-off curve	Replay ledger at alternate thresholds: what X catches vs lets through

Triggers for a calibration proposal: miss rate or foregone effect crosses a registered bar, or block precision drops below a registered floor — or released cards keep getting rejected (tighten).

Calibration (both directions)#

Trigger from scorecard bars or improvement taxonomy labels (over-strict block, over-loose release) (13-improvement.md).
Council (Opus 5.5 + Sol, offline) proposes a threshold change per gate (and per domain/plant when evidence splits). The other model red-teams nuisance cost and hard-gate adjacency. Disagreement → human.
Replay on ledger + holdouts; owner pack shows both curves (20-benchmark.md).
Accept — tech lead (global) or plant owner (plant override).
Shadow → canary → pin (16-operations.md).

Direction is symmetric: loosen on measured misses; tighten on measured nuisance. Starting stance at a new plant: moderate thresholds, measured from week 1.

Hard gates are outside this loop.

Feed into `13-improvement`#

Stream	What the loop learns
Sent cards + closures	Over-loose release, bad route, weak grounding, flooded owner
Opportunity ledger	Over-strict soft blocks; persistence and later-card joins
Backlog promote/dismiss	Owner priors; proposed constraints
Exploration (`exploration=true`)	Unbiased soft-gate error

Without the second stream, self-improvement drifts toward caution: gates hide their own mistakes. Taxonomy labels for over-strict / over-loose become registry proposals (thresholds, prompts, patterns) — still replay-gated. Generator weight training stays off until clean step labels exist (13-improvement.md).

Council judgement on a sample of blocked items prioritises human review; it is not an automatic label and never auto-emits.

Worked sketches (illustrative)#

Loosen: Soft gate evidence_tier_min blocks idle-auxiliary candidates that later show persistence with Modeled foregone kWh (calculator path). Exploration verify rate on that eligible gate is healthy; nuisance low. Owner pack shows the curve; plant owner accepts a lower tier floor for that family; shadow → canary → pin. (Action-seam disagreement is hard — never loosened this way.)

Tighten: Soft gate attention_budget holds were too rare because exception exemption was over-wide; accept rate fine but reject/no-change up on non-exception cards. Domain exemption registry tightened; replay shows lower nuisance; each plant owner accepts the pin.

Numbers in live packs are site-measured or labelled illustrative.

Missed-savings ledger vs opportunity ledger (direction)#

Two ledgers, different questions. Neither replaces the other. Fast-loop design: ../fast-loop/04-messages-budgets-and-missed-savings.md.

	Opportunity ledger (this doc)	Missed-savings ledger (`ledger.missed_savings`)
Question	What did L4's gates block, and would it have helped?	What did an action or signal that was sent but not taken likely cost?
Producer	L4	L3 twin, from follow-through on machine data
Unit	Blocked candidate, by gate id	Action or signal, by procedure and shift
Feeds	Soft-gate calibration; owner backlog	Procedure review and procedure candidates (`15` section 13); L6 view
Never	Tunes hard gates	Shown per person; summed into a headline; counted on partial-data days

Budgets. attention_budget here stays a soft pre-filter on Prescription drafts (as built: cards). Any Prescription or action that reaches a person also counts against that person's single L5 action budget (D16, ADR-036); alerts and signals follow their own L5 rules. When the two disagree, L5 wins at send time.

v1 slice vs later#

v1	Later
Ledger schema, backlog, opt-in exploration, manual scorecards	Automated trade-off curves in CI; paging on miss-rate bars
Moderate soft thresholds at Pilot plants	Per-plant overrides with clear precedence in the lockfile

Change class#

Ledger fields, backlog UX, exploration caps, and soft thresholds are data (registry + L4 store). Changing hard-gate membership or allowing exploration of hard blocks is kernel + ADR (17-change-guide.md) — and is rejected unless a future ADR explicitly overturns ADR-025.