epsilon

Epsilon (research ε)

This is the CDR-specific instantiation of the generic ε doctrine declared in ROLES.md §4b. The kernel grammar — ε-as-protocol-reviewer in the scope ladder, the watched receipt fields (protocol_gap_count, protocol_gap_refs), the gap-class instantiation pattern ({protocol}-{axis}-gap), the iteration-artifact rule (required only when protocol_gap_count > 0), the MCA discipline (ship-now / next-MCA / no-patch), and the ε=δ collapse rule for small-protocol regimes — is inherited by reference. Only the trigger classes (research-failure classes per CDR.md Field 5, not engineering-failure classes) and the output-artifact location (the CDR-iteration artifact under the project binding's wave surface) diverge for the research loss function per ROLES.md §4a.2.

§1 ε's CDR-side scope

ε's domain in CDR is protocol-iteration: the work of observing whether the CDR protocol is itself coherent — whether it selects the right research gaps, closes them with correctly-calibrated typed receipts, and produces a research-claim transmission discipline that learns from its own waves rather than repeating the same overclaim class.

The canonical output artifact is the CDR-iteration record (the research analogue of cdd-iteration.md). Per the inherited cadence rule (ROLES.md §4b.4), the artifact is required only when the wave's receipt has protocol_gap_count > 0 — i.e. when the wave produced ≥1 finding tagged with one of the cdr-*-gap classes below. The receipt's protocol_gap_count == 0 is the no-gap signal; no iteration file is required when the wave ran cleanly. The artifact's concrete location is project-binding-dependent (typically <project>/.cdr/waves/{wave-id}/cdr-iteration.md or an equivalent under the project's wave surface); the protocol-overlay layer declares the artifact's existence and shape, not its file path.

For waves that produce findings, the trigger classes the artifact enumerates are CDR-specific (per CDR.md §"Field 5"):

• cdr-data-gate-gap — receipts shipping observed or computed claims with under-specified data_refs (no manifest, no checksum, no data-use compliance record). Signals the data-policy oracle (Field 2) is not being enforced.
• cdr-overclaim-gap — receipts whose claim_status: observed evidence supports only inferred. Signals β's claim/evidence-alignment oracle is too lenient or under-specified; the procedural skill (cnos.cdr/skills/cdr/beta/SKILL.md) needs a sharper calibration-determination procedure.
• cdr-reproduction-gap — receipts with reproduction.output_match: false, or receipts that should have a reproduction record (per Field 2 + per cnos.cdr/skills/cdr/alpha/SKILL.md §"Evidence-ref completeness rule") but lack one. Signals reproduction-from-clean is not running or is failing silently.
• cdr-citation-gap — receipts with claims that should cite external work but do not, or citations that do not support the claim they are invoked for. Signals the citation-integrity oracle is drifting.
• cdr-oracle-ambiguity-gap — the same review-oracle interpretation question appears across multiple waves (β cannot decide; γ has to adjudicate). Signals oracle under-specification at the protocol layer; CDR.md or the per-role skills need additional procedural detail.
• cdr-construct-drift-gap — a key research construct's definition shifts across receipts without explicit deprecation. Signals the project binding (cph or downstream) needs a construct-stability artifact (a glossary, a deprecation registry); the protocol-layer surface ε patches may declare a placeholder convention.

These are the research-failure trigger classes, distinct from engineering's cdd-skill-gap, cdd-protocol-gap, cdd-tooling-gap, cdd-metric-gap. The MCA-discipline shape (ship now / next-MCA / drop) is identical.

The MCA discipline governs what ε does with each finding (per the generic doctrine):

• MCA available (skill patch, gate, schema refinement) → ship it now as an immediate output in the same session.
• No MCA yet, pattern real → file a protocol MCI (entry in the project binding's iteration INDEX + an issue against cnos.cdr if the patch crosses the protocol boundary).
• One-off, no pattern → drop; note the drop explicitly.

ε cross-references ROLES.md §1 row 5 (ε iterates the δ-discipline) and CDR.md §"Field 5" (the authoring procedure and per-finding shape for the CDR-iteration artifact).

§2 ε's relationship to δ

ε and δ are often the same actor for research, just as for engineering. In small-protocol regimes — one active research operator running a handful of waves — protocol-iteration volume does not justify a dedicated reviewer of the protocol. The operator (δ) naturally accumulates the longitudinal view ε requires and performs ε work as part of wave close-out triage.

Separation becomes warranted when protocol-iteration volume justifies dedicated attention: when the CDR-iteration artifact is written on most waves with non-empty findings; when the iteration INDEX accumulates faster than one actor can triage; when the operator's operational load crowds out the reflective work. At that point, ε may be a distinct actor — a second agent or a dedicated human — who reads receipts across waves and drives protocol patches independently.

No claim is made here that ε is required as a separate human or agent. ε is a structural role in the scope-escalation ladder (ROLES.md §1), not a headcount requirement. The role may collapse onto δ indefinitely in small-protocol regimes; the obligation is that ε work happens and is attributed, not that it is performed by a distinct person.

See ROLES.md §4 for the general principle governing role collapse rules, and CDR.md §"Field 6" for CDR's specific collapse rule — note that the α=β collapse is never safe for research-claim cycles (overclaim is the precise failure mode α cannot self-detect), while ε=δ and γ=δ are conditionally permitted.

§3 Persona / protocol / project boundary

This overlay declares what research ε does at the protocol layer. It does not declare:

• Who enacts ε (persona — layer 1; lives in <persona-hub>/spec/).
• What concrete iteration-INDEX path or wave directory ε writes to (project — layer 4; lives in <project>/.cdr/). The trigger classes are protocol-bound (cdr-*-gap enumerated above); the storage path is project-bound.

If a project chooses additional trigger classes (e.g. cdr-ethics-gap for waves whose data raises ethics-review issues, or cdr-stats-gap for statistical-method drift), those are project-binding extensions; this overlay declares the minimum set per CDR.md §"Field 5".