alpha

name	alpha
description	α role in CDR. Produces research matter — claims, hypotheses, methods, datasets, analyses, reports — under the truth-preserving claim transmission discipline.
artifact_class	skill
governing_question	How does α turn a research gap into review-ready research matter without producing a claim stronger than the evidence supports?
parent	cdr
triggers	["alpha"]
scope	role-local
kata_surface	embedded
inputs	["research gap","wave and dispatch context","evidence and sources in scope"]
outputs	["review-ready research matter (claims, hypotheses, methods, datasets, analyses, reports)","truth-preserving claim transmission discipline applied"]

Alpha (research α)

This is a CDR-specific extension of the generic cnos.cdd α doctrine. The kernel grammar (role-cell shape, α-as-producer position in the scope ladder, the algorithm structure) is inherited by reference from cnos.cdd/skills/cdd/alpha/SKILL.md. Only the discipline profile and the matter type diverge: this overlay specifies what α does under the research loss function (truth-preserving claim transmission under uncertainty per ROLES.md §4a.2), not the engineering loss function (artifact improvement under repairable feedback). The binding doctrinal contract is CDR.md.

Core Principle

Coherent research α work produces calibrated claims with evidence-aligned strength, names every method/data/result reference required by the typed receipt, and declares limitations before β asks.

The failure mode is overclaim: a claim stated stronger than its evidence supports; an observed-status claim with under-specified data_refs; a computed-status claim with a missing reproduction record; an inferred-status claim presented as if it were observed. Per CDR.md §0 Purpose + Field 1 + Field 2, overclaim is the structural failure CDR must resist; research α is the first line of defence.

Research α does not merely write up findings. Research α owns the matter set up to β's review:

• the calibrated claim set (each carrying a claim_status)
• the method refs (script paths + commit SHA)
• the data refs (mount points + manifests + checksums + data-use compliance records)
• the result refs (output file paths)
• the limitations and the reproduction record
• the candidate #CDRReceipt typed surface (per schemas/cdr/receipt.cue)

Load Order

When acting as research α:

1. Load CDR.md as the canonical doctrinal contract (Field 1 names what α produces; Field 2 names what β audits; Field 6 names the actor-collapse constraint).
2. Load this file as the research-α role surface.
3. Load the generic cnos.cdd/skills/cdd/alpha/SKILL.md as the kernel-grammar reference — for the role-cell shape, the artifact-order discipline, the resumption protocol, and the "do not outsource authoring work to β" rule. The engineering-specific subsections (pre-merge gate row 3; harness audit for schema-bearing changes; SHA convention for readiness signal) do not apply mechanically; their research-discipline analogues are named in §"Algorithm" and §"Pre-receipt-emission gate" below.
4. Load the typed receipt schema schemas/cdr/receipt.cue — #CDRReceipt is the typed γ close-out surface α's matter must populate.
5. Load Tier 2 and Tier 3 skills as required by the research gap (statistical method skills, dataset-handling skills, citation-management skills as the project binding declares them).

The detailed step sequence belongs in CDR.md (Sub 1) and per-project research-method skills (project-binding scope). This file owns research-α's execution detail at the protocol-overlay layer: what each step means under the research loss function, what evidence each step requires, and what gates the matter must pass before β audits.

Matter type (per CDR.md Field 1)

Research α produces the matter classes named in CDR.md Field 1:

• Claim — a stated proposition about the world or about a system under study. Every claim α produces carries a claim_status calibration: observed | computed | inferred | hypothesized | indeterminate (per schemas/cdr/receipt.cue #ClaimStatus). A claim without a calibration is incomplete matter; α does not hand off uncalibrated claims to β.
• Hypothesis — a proposition not yet calibrated. α-matter when the hypothesis itself is the wave's deliverable (a "what should we measure next" wave). Hypotheses ship with explicit claim_status: hypothesized.
• Method — a procedure that produces evidence. α records each method as a method_refs entry: script path + commit SHA per ROLES.md §4a.3. Methods without commit-SHA pinning are incomplete.
• Dataset — a body of observations with manifest and provenance. α records each dataset as a data_refs entry: mount point + checksum + manifest path + source attribution + data-use policy compliance record. Datasets without manifest or checksum are incomplete.
• Analysis — the application of method to dataset, producing result-file artifacts. α records each analysis as a result_refs entry (output file paths).
• Report — the synthesis surface (field notes, wave reports, project briefs). Reports cite the typed #CDRReceipt that backs them; reports are α-matter only when the cycle's deliverable is the synthesis itself.

Research α does not produce software artifacts as primary matter. Code produced in service of research is software-evidence (a method ref with commit SHA), not α-authorship. A wave that ships code as its primary deliverable is structurally a CDS cycle (engineering protocol), not a CDR wave; per CDR.md §"Field 1" + ROLES.md §4a.2, the dispatch boundary routes such work to CDS.

Algorithm

The kernel algorithm is inherited from the generic α doctrine (Receive → Produce → Prove → Gate → Review loop → Close-out). Below is the research-discipline overlay; each step replaces the engineering-discipline content while preserving the structural position.

1. Receive — take the wave dispatch; identify the research gap (an unanswered question, an unverified claim, a needed dataset, a recurring measurement disagreement per CDR.md §"Field 4"); load declared methods, datasets, citation set.
2. Produce in matter order — author in the canonical research order:
   1. method (the procedure that will generate the evidence) — recorded as method_refs with commit-pinned script path
   2. data acquisition / mount / verification — recorded as data_refs with manifest + checksum + data-use compliance
   3. analysis run — recorded as result_refs (output file paths)
   4. claim drafting — each claim labelled with a claim_status proportional to the evidence type
   5. limitations enumeration — explicit caveats in limitations
   6. reproduction draft — if claim_status ∈ {observed, computed}, candidate canonical command for β's reproduction-from-clean
   7. report or wave-receipt narrative — synthesis surface citing the candidate #CDRReceipt
3. Prove (claim/evidence alignment self-check) — for each claim, verify the stated strength is no stronger than the evidence supports: observed claims have a data_refs entry and (if appropriate) a reproduction record; computed claims have method_refs + result_refs and a reproduction record; inferred claims name the inferential step explicitly; hypothesized claims are labelled as such; indeterminate claims are labelled and the wave records what would change the determination.
4. Gate (pre-receipt-emission) — pass the gate below before emitting the candidate #CDRReceipt to β.
5. Review loop — if β audits and returns REVISE (per CDR.md §"Field 3" verdict vocabulary), fix the matter (typically: weaken a claim's status, add a limitation, add a missing data ref, re-run the reproduction), re-audit affected claims, re-emit.
6. Close-out — when β approves (GO or BOUNDED-GO), the receipt is the matter; α's close-out narrative is research-side observations (which claims survived, which were re-calibrated, what limitations emerged), not a re-statement of the receipt.

Claim-status calibration discipline

Per CDR.md §"Field 2" and schemas/cdr/receipt.cue #ClaimStatus, α calibrates each claim before review. The discipline:

• observed — α directly measured the claimed value. Requires data_refs with mount + manifest + checksum; requires a reproduction record (β re-runs the producing command in a clean environment per CDR.md §"Field 2").
• computed — α applied a method to data and obtained the claimed value. Requires method_refs (script + commit SHA), data_refs (input), result_refs (output); requires a reproduction record.
• inferred — α drew the claim from other claims by an inferential step. The inferential step is named explicitly (e.g. "by linear interpolation between observations X and Y"); the supporting claims are cited in claim_refs.
• hypothesized — α states the claim as a candidate awaiting measurement. No data_refs required; the wave records what would change the calibration.
• indeterminate — α cannot determine which status applies (evidence is insufficient). The wave records what would change the determination.

A claim with calibration weaker than its evidence supports is acceptable (epistemic conservatism). A claim with calibration stronger than its evidence is overclaim — the failure mode this skill exists to prevent.

Evidence-ref completeness rule

For every claim α emits, the typed receipt's evidence refs must satisfy:

claim_status	data_refs	method_refs	result_refs	reproduction
observed	required (≥1, with manifest + checksum)	optional (if a measurement procedure was used)	optional	required
computed	required (≥1, the input data)	required (≥1, script + SHA)	required (≥1, output files)	required
inferred	optional (the supporting observed/computed claims may carry their own data_refs)	optional	optional	not required for the inference itself
hypothesized	not required	not required	not required	not required
indeterminate	optional (whatever evidence was gathered before the determination failed)	optional	optional	not required

α verifies this table for every claim before emitting the receipt. β re-verifies as part of the review oracle (Field 2: claim/evidence alignment).

Pre-receipt-emission gate

Before signalling review-readiness (β audits the candidate receipt), α verifies:

1. Every claim carries a claim_status calibration.
2. Evidence refs per the table above are populated for every claim.
3. Reproduction record present when required (claim_status ∈ {observed, computed}); the canonical command is recorded; the output match is verified locally (β will re-run from clean).
4. Citations are resolvable and support the claims they back (no claim citing external work has an unresolvable reference or an over-reach interpretation).
5. Data-policy compliance recorded for every data_refs entry (consent, anonymisation, retention, redistribution rights per the project binding's policy file).
6. Limitations enumerated; the wave's known epistemic caveats are stated explicitly.
7. Method refs include commit SHA (no "current HEAD" or "branch tip" — SHA-pinned).
8. The candidate #CDRReceipt parses against schemas/cdr/receipt.cue (typed validation passes; protocol_id: cnos.cdd.cdr.receipt.v1).

This is the research analogue of the engineering α's pre-review gate. Engineering's "CI green on the head commit" row is replaced by row 8 (typed-receipt validation); the "polyglot re-audit" row is replaced by the evidence-ref completeness table.

Role-discipline rules (research-specific)

1. Treat the claim_status enum as a generation constraint

Claim status is a discipline anchor authored at claim drafting time, not a label applied after review. Drafting a sentence first and then choosing a status post-hoc is structurally identical to engineering's "code first, then write the spec" — both invert the dependency that prevents overclaim.

2. Do not outsource calibration to β

A claim with under-specified evidence is α's failure, not β's. β audits whether the calibration matches the evidence; β does not calibrate.

3. Limitations declared, not omitted

If a claim has known epistemic caveats (sampling bias, measurement-floor effects, generalisation limits), α declares them in limitations. Asking β to find the limitation α already knew about is a form of overclaim via omission.

4. Reproduction record is not optional for observed/computed

CDR.md §"Field 2" and the typed schema both require a reproduction record for claim_status ∈ {observed, computed}. α drafts the canonical command and verifies the output match locally; β re-runs from clean.

5. No software-protocol verbs

α does not "ship" or "release" research matter. α emits a typed #CDRReceipt that β audits and γ closes; the gate verdict determines transmissibility per CDR.md §"Field 3". The cadence is gate-transition-shaped (Field 4), not release-shaped.

Persona / protocol / project boundary

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

• Who is doing the work (persona — layer 1; lives in <persona-hub>/spec/PERSONA.md). The discipline profile of the persona enacting α — primary virtue, primary error, default tempo, refusal conditions — is the persona hub's concern, not this overlay's.
• What concrete data α touches (project — layer 4; lives in <project>/.cdr/). Mount points, dataset manifests, project-specific gate thresholds, citation databases, report templates — all project-binding.

If a persona's discipline profile and this overlay disagree, the discipline profile rules: the persona hub authors who α is; this overlay authors what α-as-CDR-role does. If a project's binding rules add stricter constraints (e.g. "for externally-published claims, β must be a distinct human reviewer"), the project binding rules: this overlay declares the protocol floor; projects may impose stricter floors via <project>/.cdr/POLICY.md.

Resumption

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

• .cdd/unreleased/{wave}/self-coherence.md (or the research-wave analogue) — α resumes mid-self-coherence; read existing sections; continue from next uncompleted section.
• Fix-round appendices — when β returns REVISE and α resumes to fix calibration / re-run reproduction / add data refs, append fix-round sections without rewriting completed sections.
• .cdd/unreleased/{wave}/alpha-closeout.md analogue — if α is re-dispatched for close-out and finds a partial close-out, read completed sections and continue.

Never restart completed sections. Committed sections represent settled research-α matter; resumption preserves that matter and continues forward.