ADR-0036 — Outcome-derived labels, censoring and lockbox v1¶
Status: ACCEPTED
Date: 2026-09-20
Context¶
F03 already owns the canonical OutcomeSpec / Outcome boundary, including
explicit forward horizons, COMPLETE | CENSORED_END_OF_DATA |
INSUFFICIENT_COVERAGE, causal availability and exact canonical fraction-text
numeric values. ADR-0008 requires Outcomes to precede Labels and distinguishes
outcome-derived labels from policy-derived labels.
F06 is frozen and complete under ADR-0031. It owns walk-forward candidate
admissibility, causal availability, warm-up/support, exact dependency-overlap
purge, embargo and deterministic classification precedence. ADR-0031
intentionally leaves target/outcome support, label/censoring policy and the
lockbox/hidden-evaluation boundary to F07, while providing
DependencyCutoffRole.FOLD_COMPLETION for a future target dependency.
DG-E therefore has one bounded blocking proposition on the Validation -> Strategy / Supervised-ML path: freeze F07 semantics without reopening F03 or F06 and without importing downstream Strategy/Execution assumptions.
Issue #95 is the governance-only decision-gate scope for this ADR. F08 DSR/PBO is an independent DG-E branch and is not part of this decision.
Decision¶
Adopt F07 labels/censoring/lockbox v1 as the canonical Validation semantic
foundation described below.
1. F07 v1 is outcome-derived only¶
F07 v1 supports outcome-derived labels only.
Policy-derived labels remain a valid future family under ADR-0008, but they
require explicit StrategySpec / ExecutionSpec authority. They are deferred
until those semantics exist and are not approximated inside F07.
F07 does not redefine F03. Outcome horizon, state, realized value, path measurement, censoring and causal availability remain owned by F03.
2. LabelDefinition v1¶
A canonical outcome-derived label definition declares, at minimum:
- a governed
label_key; - an explicit positive semantic version;
- exactly one canonical source
OutcomeSpecidentity; - one v1 transformation kind and its identity-bearing parameters;
- an explicit output value/class schema;
- the v1 censoring policy
REQUIRE_COMPLETE.
All output-affecting semantics above participate in deterministic definition identity. Descriptive text, runtime location, host/process information and execution timestamps do not.
The identity domain is:
3. Source eligibility and censoring¶
F03 source state is preserved, never guessed or collapsed:
Outcome COMPLETE
-> eligible for deterministic label transformation
Outcome CENSORED_END_OF_DATA
-> no label value; source-censored state remains explicit
Outcome INSUFFICIENT_COVERAGE
-> no label value; source-insufficient state remains explicit
A censored or insufficient Outcome is never coerced to zero, negative,
no-touch, false, a neutral class or any other fabricated target.
Only a COMPLETE Outcome carrying its canonical F03 realized_value may
produce a label value.
4. Minimum v1 transformations¶
F07 v1 freezes only two deterministic transformation families.
Exact continuous identity¶
IDENTITY_VALUE preserves the source Outcome's canonical exact rational
realized_value unchanged as the label value. No float conversion or Decimal
rounding context may alter it.
Exact ordered-threshold categorical mapping¶
ORDERED_THRESHOLDS declares:
- a strictly increasing ordered tuple of canonical exact-rational thresholds;
- exactly
N + 1governed class values forNthresholds.
For thresholds t0 < t1 < ... < tN-1, interval semantics are:
Equality therefore belongs to the upper interval. Threshold comparison is exact rational comparison over the F03 value; no epsilon/tolerance or float conversion is allowed.
No generic expression language, callback/plugin label engine, retrospective percentile transform or triple-barrier implementation is introduced by F07 v1. A future path-dependent label that needs semantics not present in F03 must first obtain its own canonical Outcome authority rather than hiding new Outcome semantics inside Label transformation.
5. Label-result identity binds source evidence¶
A label result is traceable to both its LabelDefinition and the exact source
Outcome evidence.
Its deterministic identity must bind at least:
- label-definition identity;
- source Outcome identity;
- source Outcome state;
- source horizon start/end;
- source causal availability;
- canonical source realized value when present;
- resulting label state/value when present.
The result identity domain is:
This explicit evidence binding is required because an opaque source Outcome identity alone must not let two different realized source results collapse to one label-result identity.
6. Label causal availability¶
For the pure deterministic v1 transformations above:
A label can never become available before the Outcome evidence that produces it. Label/lockbox values are evaluation targets and may never be fed backward as features or decision inputs at an earlier decision instant.
7. Canonical target support preserves the Outcome boundary¶
F03 evaluates the source path from horizon_start == event_time through the
boundary observation exactly at horizon_end. That terminal observation is
consumed evidence and therefore belongs to target support.
The semantic target support is consequently closed at the right edge:
F06 uses canonical half-open CoverageInterval semantics. When projecting an
F07 target into that seam, the projection must preserve the consumed terminal
instant exactly. With the current nanosecond Instant domain the canonical
projection is:
or an exactly equivalent half-open representation that includes
horizon_end. It is forbidden to shrink this to
[horizon_start, horizon_end): if horizon_end == test_start or
lockbox_start, the target consumed evidence at that held-out boundary and
the development sample is contaminated.
8. Training-target integration with F06¶
For a training/development sample, an F07 target is a fold-completion dependency, not a decision-time feature dependency.
A usable training target must:
- originate from a
COMPLETEF03 Outcome; - carry the exact target support above;
- carry causal availability derived from the source Outcome;
- be represented to the F06 seam with cutoff role
FOLD_COMPLETION; - be complete/available by the F06 fold-completion cutoff;
- not overlap the held-out interval under ADR-0031's exact half-open overlap rule.
Censored/insufficient source Outcomes do not become labeled training samples.
No rule here weakens F06 decision-time availability: feature/observable
inputs still must have been available by the sample decision/reference
instant d.
For test/lockbox samples, the future target is an evaluation output, not a
decision-time dependency of the same sample. Requiring its future label to be
known at d would itself be a temporal error.
9. Lockbox v1¶
A v1 lockbox is one explicitly declared terminal contiguous holdout interval:
It begins after the development folds selected for the same evaluation plan.
Sample membership is determined by the canonical candidate/reference instant
d: d in L means the sample belongs to the lockbox population.
While the lockbox is unrevealed:
- a development sample may not have
dinsideL; - development dependency/target support may not intersect
L; - evidence needed to score lockbox members, including forward target evidence
extending beyond a member's
d(and, where necessary, beyondlockbox_end), is part of hidden evaluation evidence and must not feed the development loop; - lockbox labels, aggregates or diagnostics may not influence feature selection, label thresholds, model/hyperparameter selection, strategy or decision thresholds, trial selection, stopping decisions or any other development choice.
A lockbox evaluation is run only against a candidate/configuration whose relevant development choices are already frozen for that evaluation.
F07 v1 freezes the semantic isolation, not a physical security mechanism. Separate database credentials, encryption, filesystem ACLs, vaults or remote services are not required by this ADR and may be added later only when an operational requirement justifies them.
10. Reveal is explicit and semantically irreversible¶
The lockbox has two semantic visibility states:
Reveal is an explicit action/evidence transition; it is never inferred merely because evaluation bytes exist somewhere.
Once lockbox results have been revealed and observed, that exact population cannot subsequently be claimed as an untouched lockbox for a development or selection decision influenced by those results. It may remain historical evaluation evidence, but another independent untouched population is required for a new lockbox claim.
There is no REVEALED -> UNREVEALED transition.
11. Package/ownership boundary¶
F03 remains Research-owned and F07 remains Validation-owned. The existing
package DAG does not permit quant_platform.validation to import Research
runtime types directly.
A future F07 implementation must therefore consume a narrow Validation-owned projection of the exact F03 Outcome evidence needed by this ADR, supplied by a permitted composition/caller boundary. That projection must preserve the source identities, state, horizon, causal availability and canonical value; it must not reinterpret or recompute F03 semantics.
Concrete local type/helper names remain implementation-local.
Excluded¶
This ADR does not authorize or define:
- F07 runtime implementation;
- F08 DSR/PBO;
- policy-derived labels;
- StrategySpec, backtesting, execution, PnL, fees or fills;
- new F03 Outcome families or hidden triple-barrier/path semantics;
- model training or feature selection algorithms;
- Experiment Study/Trial/Run accounting or persistence;
- physical lockbox security/ACL infrastructure;
- a generic label/validation DSL, registry, plugin system or framework.
Consequences¶
- F07 decision state becomes
FROZEN; implementation remainsMISSING. - The DG-E validation/labeling branch has no remaining semantic blocker for the bounded F07 v1 implementation.
- F08 remains an independent
OPEN_BLOCKINGDG-E branch. - G01 and I04 remain implementation-blocked until F07 runtime is actually
complete; semantic freeze alone does not satisfy their
Requires F07implementation dependency. - F03 and ADR-0031/F06 remain unchanged authorities and are credited rather than re-proved.
Acceptance evidence¶
This governance-only freeze is tracked by issue #95. Existing F03 tests and ADR-0031/F06 adversarial tests are credited evidence for the source Outcome and fold-admissibility propositions respectively; this decision scope does not duplicate those test campaigns or mutate runtime code.