# Current research decisions

Updated: 2026-09-05

Status: `L10_R0_ACTIVE / DTR_R2_DYNAMIC_RETAINED`

This page owns cross-route priorities and decision criteria. Each route current
owns its latest baseline, exact metrics, source boundaries and experiment status.
Historical gates and successor suggestions apply to named scopes; new hypotheses
or evaluation criteria need a rationale and a useful check, not prior success.

## L10: useful commitment under partial evidence

The present question is how to retain correct target recovery while reducing wrong
commitments without hiding abstentions or observation/reference costs.

Use the [L10 current](../research/active/l10-r0/CURRENT.md) for the paired fixed,
triggered and geometric-verifier baselines. Keep the triggered policy and episode
harness available within their recorded Development limits. An unconditional
verification gate must justify the correct coverage it removes.

The [consumed paired diagnosis](../research/WORKFLOW_UPGRADE_20260905.md) separates
two lost correct bindings with target-box support from three without it; error
reduction has no baseline error opportunities. Explore an evidence representation
that distinguishes availability, contradiction and extent. Integrate only a demonstrated useful
change; seek appropriate unchanged-method confirmation before expanding its claim.
No new street source, threshold rescue or protected outcome access is authorized
by this navigation update. The existing frozen result remains unchanged.

## DTR: task-relevant event comparison and capture readiness

The present question is whether the retained risk representation improves useful
events over credible simple baselines under the same observable information.
Use the [DTR current](../research/active/dtr-r0/CURRENT.md) for algorithm authority,
the interrupted R1 source record and the latest committed engineering evidence.

Separate the next capture/readiness check from the algorithm comparison. A repaired
writer or completed capture is an engineering result; it cannot establish risk
performance. Use representative task-owned smoke/replay inputs and the existing
recovery boundaries before spending a new protected run.

After adequate engineering evidence, the next algorithm comparison should report
event utility, false events, coverage and missingness at a fair input/cost budget.
If added complexity has no useful gain, simplify or change the central hypothesis.
R1 is not retried or resumed by this policy update; any protected successor retains
the versioned protocol and evidence requirements appropriate to its claim.

## Shared execution choices

- **Explore:** test one explanatory hypothesis; necessary coupled edits and disclosed
  consumed Development are allowed. Keep comparisons reconstructable.
- **Confirm:** fix the method, comparison, criteria and retry/access rules before
  outcomes; choose independence appropriate to the claim.
- **Engineering:** resolve observed correctness, runtime or recovery bottlenecks.
  Classify implementation/source failures separately from method evidence.
- Each outcome determines continue, revise, integrate or stop. Complete remaining
  authorized delivery and resource release after the individual experiment ends.
- Choose baselines, controls and ablations for the discrepancy being tested; their
  historical labels alone do not make them universally mandatory.
- L10 and DTR do not wait for, modify or validate one another. Preserve concurrent WIP.

The [research workflow](../research/WORKFLOW.md) contains the brief and command.
Do not add governance or tests without a decision benefit or named material risk.
No local result establishes arrival, handoff, natural-camera reliability, user
benefit or safety beyond its own evidence. `UNKNOWN` remains distinct from CLEAR.

Detailed trajectories remain in owning ledgers/results and Git
`daf5720064d98a93b75336469d18e9a2fe0023e5:docs/CURRENT_DECISION.md`.