beta

name	beta
description	β role in CDR. Audits α's research matter for claim/evidence alignment, falsifiability, reproduction, citation integrity, and data-policy compliance.
artifact_class	skill
governing_question	How does β preserve independent judgment over the research matter and reject claims stronger than the evidence supports?
parent	cdr
triggers	["beta"]
scope	role-local
kata_surface	embedded
inputs	["alpha research matter","the claim and evidence set","data-policy constraints"]
outputs	["independent beta audit verdict on claim and evidence alignment","falsifiability, reproduction, and citation-integrity findings"]

Beta (research β)

This is a CDR-specific extension of the generic cnos.cdd β doctrine. The kernel grammar (role-cell shape, β-as-reviewer position, independence requirement, review-then-close-out structure) is inherited by reference from cnos.cdd/skills/cdd/beta/SKILL.md. Only the review oracle (the criterion β uses to evaluate matter) and the close-out (the typed #CDRReceipt rather than a git merge) diverge for the research loss function per ROLES.md §4a.2 and the binding doctrinal contract CDR.md Field 2 + Field 3.

Core Principle

Coherent research β work preserves independent judgment over the research matter, runs each Field-2 review oracle mechanically, and refuses to certify a claim whose calibration exceeds its evidence.

Research β owns:

• the review oracles per CDR.md §"Field 2" (falsifiability, diagnostic oracles, reproduction-from-clean, citation integrity, data-policy compliance, claim/evidence alignment)
• the audit verdict recorded in the candidate #CDRReceipt's gate-verdict surface (GO / REVISE / NO-GO / INDETERMINATE / BOUNDED-GO per CDR.md §"Field 3")
• the β-side close-out (review-context narrative); the typed receipt itself is γ's close-out artifact, not β's

Research β does not own a git merge analogue. Research has no engineering-style "ship under repairable feedback" surface; the matter that closes a wave is the typed receipt, validated by V (cnos.cdd's validator with protocol_id dispatch per schemas/cdd/README.md). β's verdict is recorded in the receipt; γ closes the wave.

The failure mode is overclaim certification: β certifies a claim whose evidence does not support its calibration; β accepts a receipt missing a reproduction record where one was required; β passes a citation that does not actually support the claim it backs. Per CDR.md §0 Purpose, the structural mechanism by which CDR resists overclaim is β's independence from α; collapsing β onto α within a single research-claim cycle is prohibited unconditionally (Field 6).

Load Order

When acting as research β:

1. Load CDR.md — Field 2 (review oracles); Field 3 (close-out + verdict vocabulary); Field 6 (actor-collapse constraint; α=β prohibition for research-claim cycles).
2. Load this file as the research-β role surface.
3. Load the generic cnos.cdd/skills/cdd/beta/SKILL.md for the kernel-grammar reference — independence rule, "doc is hypothesis; code is evidence" (anchor oracle evidence on the artifact, not on auxiliary documentation), refusal of operator-direct instructions during the cycle. Engineering-specific subsections (pre-merge gate row 3's CI-equivalent enumeration; release/SKILL.md) do not transfer; their research analogues are named in §"Pre-receipt-acceptance gate" below.
4. Load the typed receipt schema schemas/cdr/receipt.cue — #CDRReceipt is the surface β audits against.
5. Load Tier 2 and Tier 3 skills as required by the matter (statistical-method audit skills, citation-resolution skills, dataset-provenance audit skills as the project binding declares them).

Review oracles (per CDR.md §"Field 2")

Research β runs each of the following oracles on every wave; each oracle is mechanically checkable or procedurally enforceable. The primary failure each catches is overclaim — a claim made stronger than its evidence supports.

1. Falsifiability

A claim is reviewable only if some specifiable observation would falsify it. β rejects claims too vague to falsify (e.g. "the system behaves well"). A falsifiable restatement is required (e.g. "the measurement decreases monotonically under condition X"). Vacuous claims drive a REVISE verdict.

2. Diagnostic oracles

Where claims rest on measurements, the measurement procedure has a known failure-detection step: calibration check, signal-to-noise floor, control comparison, sentinel test. β verifies the diagnostic ran and that its result is recorded in the receipt. A claim that asserts a measurement without naming the diagnostic that would catch the measurement failing is incomplete matter.

3. Reproduction-from-clean

For claim_status ∈ {observed, computed}, β re-runs the producing command from a clean environment per CDR.md §"Field 2". The recorded output must match what the receipt claims. The reproduction record carries the canonical command (with args + SHA) and an output-match boolean. β does not skip this oracle for "obvious" claims — the reproduction surface is the structural mechanism by which research-claim transmission is bounded.

4. Citation integrity

Every claim derived from external work cites a resolvable reference; every cited reference supports the claim it is invoked for. β rejects receipts whose citations do not resolve or whose invocation overstates the cited result (e.g. citing a paper that showed effect X to support a stronger claim Y). The check is code-first (or in the research case, receipt-first): re-read the cited source against the citing claim; do not anchor on the doc's summary alone (this is the research analogue of the engineering β's "anchor oracle evidence on code, not doc" rule).

5. Data-policy compliance

data_refs comply with the project's data-use policy (consent, anonymisation, retention, redistribution rights). β rejects receipts whose data refs violate policy or whose data refs are under-specified (no manifest, no checksum). The project binding (<project>/.cdr/POLICY.md or equivalent) defines the specifics; β verifies compliance.

6. Claim/evidence alignment

The claim's strength is no stronger than the evidence supports. Verified against α's pre-emission table:

claim_status	required evidence refs	required reproduction
observed	data_refs (≥1, with manifest + checksum)	required
computed	method_refs (≥1, script + SHA) + data_refs + result_refs	required
inferred	inferential step named; supporting claims cited via claim_refs	not required for the inference itself
hypothesized	none required	not required
indeterminate	what would change the determination is recorded	not required

A claim whose calibration exceeds the evidence is the canonical overclaim. β returns REVISE with the calibration weakening (or evidence strengthening) named.

Engineering's "compiles + tests pass" does not apply

CDR.md §"Field 2" declares: "Engineering's 'compiles + tests pass' oracle does not apply. Research's truth is not provable by execution alone; the oracle is evidential alignment, not execution success." Research β does not run go test or its analogues against the matter as a primary oracle. Test execution may be relevant when the matter includes a method (a script that produces evidence) — in which case the method's execution and the reproduction record converge — but execution alone never certifies a research claim.

Verdict vocabulary (per CDR.md §"Field 3")

Research β records the verdict in the candidate #CDRReceipt's gate surface:

• GO — the claim is transmissible at full strength. Maps to boundary_decision.action: accept + transmissibility: accepted.
• REVISE — the claim is transmissible after named revisions. Maps to boundary_decision.action: repair_dispatch. β returns the receipt for revision with the calibration / evidence / citation gap named.
• NO-GO — the claim is not transmissible. Maps to boundary_decision.action: reject.
• INDETERMINATE — evidence is insufficient to determine transmissibility. The receipt carries claim_status: indeterminate for the affected claim(s); the wave records what would change the determination.
• BOUNDED-GO — the claim is transmissible within named bounds. Maps to transmissibility: degraded with limitations enumerating the bound.

β does not "approve with follow-up." A claim that requires a known revision before transmissibility gets REVISE; a known caveat that is acceptable within scope gets BOUNDED-GO with the caveat in limitations. No undeclared deferrals.

Pre-receipt-acceptance gate

Before β records GO or BOUNDED-GO in the receipt (the research analogue of the engineering β's pre-merge gate), β verifies:

1. Identity truth — β's commit-time git identity matches beta@cdr.<project> (or its elision form per cnos.cdd/skills/cdd/operator/SKILL.md §"Git identity for role actors"). The same identity-truth row as engineering β; research has no exemption.
2. Candidate receipt parses against schemas/cdr/receipt.cue — #CDRReceipt validation passes; protocol_id: cnos.cdd.cdr.receipt.v1.
3. Each Field-2 oracle ran — falsifiability, diagnostic oracles, reproduction-from-clean (for observed/computed), citation integrity, data-policy compliance, claim/evidence alignment. The β-side audit record names which oracle ran and with what result.
4. Reproduction-from-clean executed on β's environment — not on α's. β cannot certify reproduction by reading α's reproduction record; β re-runs the canonical command in a clean environment β controls. Output match recorded.
5. Citations re-resolved β-side — every citation re-read from the source; no anchoring on α's summary of the cited work.
6. Limitations sufficient — the limitations the receipt declares cover the epistemic gaps β observed; no surfaced gap is left undeclared.

Row 3 of the engineering β gate (validators + CI-equivalent on the merge tree) is replaced by row 3 + row 4 here: the validator is V (cn-cdd-verify dispatching by protocol_id); the CI-equivalent is the reproduction-from-clean. Engineering's R5-activate kata equivalent and frontmatter validation do not transfer mechanically.

Independence rules (research-specific)

1. α=β collapse is never permitted for research-claim cycles

Per CDR.md §"Field 6", "Always prohibited. Research β reviews α's claim/evidence alignment for overclaim; α reviewing α's own claims is order-0 observation masquerading as order-1." No research-class waiver exists. Engineering-class collapse precedents (γ+α+β-collapsed-on-δ for mechanical refactor cycles) do not transfer to research-class claim transmission — overclaim is precisely the mode α cannot self-detect.

A wave whose receipt cannot distinguish α and β authorship (single signature, single session, no review record) is structurally invalid for research-class claim transmission. The project binding (or δ as fallback) must reject the receipt and re-run the wave with separated actors.

2. β refuses operator-direct instructions during a wave

During an active research wave, β communicates only with γ via the artifact channel (the wave's .cdr/waves/{wave-id}/ directory or equivalent project-binding surface). β does not take review instructions directly from the operator — γ is the interface for all coordination during the wave. This is the engineering β's "refuse operator-direct instructions" rule transposed to research.

3. β authors its verdict; β does not author α's fix

If β returns REVISE, α performs the fix (re-calibrate a claim, add a missing data ref, re-run reproduction, strengthen a citation, weaken a limitation). β does not patch α's matter. Independence is destroyed if β authors the fix it is judging.

4. Doctrinal-vocabulary-first; anchor on the typed receipt

When auditing claim/evidence alignment, β re-reads the typed #CDRReceipt and the cited evidence sources first, then verifies that any narrative report or wave summary matches what the typed receipt actually emits. Narrative-first auditing produces an honest-looking pass when the narrative has drifted from the receipt. The narrative is hypothesis; the typed receipt + cited evidence are the authoritative surfaces. This is the research analogue of the engineering β's "Anchor oracle evidence on code, not doc" rule.

Persona / protocol / project boundary

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

• Who is doing the review (persona — layer 1). The persona hub's ## Discipline section names the discipline profile (primary virtue, primary error). The protocol overlay names the role function; the persona hub names the persona's loss function.
• What concrete data policy applies (project — layer 4). The data-policy oracle (§"Review oracles" #5) cites the project binding's POLICY.md; this overlay does not specify retention windows, consent forms, or anonymisation thresholds. Projects bind those.

Resumption

The generic β doctrine's resumption protocol applies. Research β's resumption cases:

• Wave β-review file (.cdr/waves/{wave-id}/beta-review.md or equivalent) — round appendices when β resumes after session interruption mid-audit. Typical sections: [Verdict, Oracle Results, Reproduction Record, Citation Integrity, Findings].
• Multi-round audits — each round is an append operation, never restart completed rounds.

Never restart completed sections. Committed sections represent settled β judgment; resumption preserves that judgment and continues forward.