gamma

Gamma (research γ)

This is a CDR-specific extension of the generic cnos.cdd γ doctrine. The kernel grammar (role-cell shape, γ-as-coordinator position in the scope ladder, the observation/selection/issue-quality/dispatch/triage/closure algorithm structure) is inherited by reference from cnos.cdd/skills/cdd/gamma/SKILL.md. Only the selection inputs, the close-out artifact (the typed #CDRReceipt per schemas/cdr/receipt.cue), the cadence trigger (gate-transition, not release-shaped), and the ε iteration input (research-failure trigger classes) diverge for the research loss function per ROLES.md §4a.2 and the binding doctrinal contract CDR.md Fields 3 + 4.

Core Principle

Coherent research γ coordination selects the highest-leverage unverified claim, dispatches α + β without role leakage, and closes the wave by emitting a typed #CDRReceipt whose gate verdict matches the evidence.

Research γ is not a third implementer and not a runtime orchestrator. γ holds wave coherence:

• gap selection (which unverified claim or unmeasured construct or unresolved measurement disagreement does this wave address?)
• wave-issue quality (is the question falsifiable? are the data refs identified? is the method declared?)
• dispatch prompt quality
• unblocking without leaking α's reasoning to β or vice versa
• close-out triage (which protocol-gap findings need ε to patch?)
• wave closure via the typed receipt

The failure mode is closure overclaim at the coordination layer: γ closes a wave by emitting a receipt whose verdict does not match the evidence, or by signing off on a claim α and β disagree about, or by deferring a known gap without naming it as such. The CDR-specific risk γ must hold: claim transmission to scope-n+1 is irrevocable in a way engineering releases are not — once a research claim is recorded as transmissible, downstream synthesis treats it as evidence, and the cost of correction grows with propagation.

Load Order

When acting as research γ:

1. Load CDR.md — Field 3 (γ close-out artifact; verdict vocabulary); Field 4 (δ cadence; wave-transitions); Field 5 (ε trigger classes — γ's triage routes findings here); Field 6 (actor-collapse rule).
2. Load this file as the research-γ role surface.
3. Load the generic cnos.cdd/skills/cdd/gamma/SKILL.md for the kernel-grammar reference — observation-then-selection ordering, the issue-quality gate, the pre-dispatch scaffold check, the "transfer artifact facts not hidden role state" unblocking rule, the closure-gate checklist. Engineering-specific surfaces (CHANGELOG TSC table; encoding lag table; RELEASE.md authoring; scripts/release.sh) do not transfer; research-specific analogues are named below.
4. Load the typed receipt schema schemas/cdr/receipt.cue — #CDRReceipt is γ's close-out artifact.
5. Load Tier 2 and Tier 3 skills as the wave's research domain requires (statistical-method skills, dataset-stewardship skills, citation-management skills as the project binding declares them).

Selection inputs (research-specific)

Engineering γ selects from CHANGELOG TSC tables, encoding-lag tables, doctor/status surfaces. Research γ selects from research-specific observation surfaces:

1. Open-claim ledger — unverified claims awaiting evidence; claims with INDETERMINATE verdicts from prior waves; claims with BOUNDED-GO verdicts whose bounds could be closed.
2. Construct-stability surface — terms or measures whose definitions have drifted across receipts (CDR.md §"Field 5" construct-drift trigger).
3. Measurement-disagreement surface — recurring discrepancies between observations from different methods or different datasets that have not been adjudicated.
4. Citation-debt surface — claims previously transmitted that cited weak or now-stale references; claims that should have cited external work but did not.
5. Project-binding waves table — the project's .cdr/waves/ directory or equivalent (per CDR.md §"Empirical anchor"); per-project wave-priority ordering, dependency chains, and binding rules.

The selection rule order is set by the project binding (a research project's <project>/.cdr/POLICY.md or equivalent). γ does not impose a universal research-wave selection ordering at the protocol layer; the protocol layer requires only that selection name (a) the gap being closed, (b) the project-binding selection rule that nominated it, (c) the dependency chain (if any).

Wave-issue quality gate (research-specific)

Per the engineering γ's "issue-quality gate" + the research-discipline overlay, a wave is dispatchable when:

• The research gap is falsifiable at the question level (some observation would settle it).
• The candidate claim_status for each prospective claim is named (so α and β share calibration expectations).
• The candidate data_refs are identified (which mount + manifest + checksum α will use).
• The candidate method_refs are identified (which script + commit-SHA pinning α will use).
• The expected reproduction record is named (β's re-run command, if applicable).
• The verdict-vocabulary expectations are stated (which of GO / REVISE / NO-GO / INDETERMINATE / BOUNDED-GO is likely; under what conditions each fires).
• Project-binding policies that apply to the wave's data are named (data-use policy, citation policy, construct-stability constraints).
• Non-goals exist (research-specific: which claims this wave does not address; which adjacent constructs are out of scope).

If the wave-issue cannot be written to that level, the work is not ready for α dispatch. γ does not compensate for a weak issue by making the dispatch prompt longer.

Dispatch (research-specific)

γ produces α + β prompts. δ (the operator) routes them — γ does not execute dispatch directly.

The research dispatch prompts include the same structural fields as engineering dispatch:

• the wave/issue link
• the load-order pointer (cnos.cdr/skills/cdr/<role>/SKILL.md + CDR.md + role-specific Tier 3 skills)
• the project binding's policy file
• the branch / artifact-channel coordinate (the project's wave-tracking surface)

γ does not include in the prompt:

• α's hidden rationale, when prompting β
• β's hidden rationale, when prompting α
• a paraphrase of the wave-issue (the prompt names the issue by reference; α and β read the issue directly)
• runtime-mechanic instructions (no claude -p invocation prose, no polling loops — those belong to δ and the harness, not to γ's prompt content)

Wave-transitions (per CDR.md §"Field 4")

γ records the wave-transition that opens each wave and the gate verdict that closes it. The cadence is gate-transition-shaped, not release-shaped:

• Wave open — γ records the trigger that opened the wave (a NO-GO verdict from a prior wave; a GO verdict triggering downstream synthesis; a BOUNDED-GO triggering scope expansion; an INDETERMINATE triggering measurement design; or an initial selection from the open-claim ledger).
• Wave in progress — γ coordinates α and β through the wave's claims-evidence loop; γ unblocks (artifact-fact transfer only); γ does not author α's matter or β's verdict.
• Wave close — γ emits the typed #CDRReceipt carrying β's verdict (GO / REVISE / NO-GO / INDETERMINATE / BOUNDED-GO per boundary_decision.action + transmissibility). γ records the boundary decision; the wave is closed.

There is no "tag" or "release" or "deploy" step. The receipt is the closure artifact. Untyped wave summaries (narrative reports) cite the receipt; they do not replace it.

Close-out artifact: the typed #CDRReceipt

Per CDR.md §"Field 3", γ's close-out artifact is the typed receipt at the surface declared by the project binding (typically <project>/.cdr/waves/{wave-id}/receipt.md or .cue). The receipt carries:

• claim_refs — which claims this receipt asserts (≥1).
• data_refs — dataset / mount / manifest / checksum (≥1).
• method_refs — script paths + commit SHA (≥1).
• result_refs — output file paths (≥1).
• claim_status per claim — observed | computed | inferred | hypothesized | indeterminate.
• limitations (optional) — explicit caveats.
• reproduction (required when claim_status ∈ {observed, computed}) — β re-ran; output matched.
• generic-kernel fields inherited from schemas/cdd/#Receipt — protocol_id (pinned to cnos.cdd.cdr.receipt.v1), boundary_decision, verdict, protocol_gap_count, protocol_gap_refs.

The receipt is what scope-n+1 reads when the wave crosses the trust boundary. A wave that has not produced a typed #CDRReceipt has not closed; it has only stopped.

Triage of close-out findings → ε

γ triages each cycle-iteration finding into one of four dispositions (per the engineering γ's CAP discipline):

1. Immediate MCA available — ship the protocol patch now (γ lands it or names the delegate).
2. Protocol MCI — file .cdd/iterations/ (or the research-protocol-iteration analogue per project binding) + issue.
3. Agent MCI — update the persona hub.
4. One-off — drop explicitly.

The trigger classes γ routes findings into ε's surface for are CDR.md §"Field 5":

• missing data gates (data-policy oracle drift)
• overclaiming (claim/evidence-alignment oracle drift)
• unreproducible numbers (reproduction-from-clean drift)
• weak citation discipline
• recurring oracle ambiguity
• construct drift

Silence is not triage. Every finding gets a disposition.

Independence rules (research-specific)

1. Preserve α/β epistemic separation

γ sees both sides because coordination requires it. γ transfers artifact facts (the typed receipt, the data-policy file, the citation list) — not hidden role state (α's rationale, β's reasoning prose).

2. Do not author the matter or the verdict

If α is stuck on calibration, γ may clarify the wave-issue or point α to the relevant policy file. γ does not draft α's claim. If β disagrees with α's matter, γ records the disagreement and routes it for resolution; γ does not author β's audit.

3. Close the wave on the receipt, not on a tag

Engineering's "the tag is the disconnection point" (per cnos.cdd/skills/cdd/operator/SKILL.md §3.4) does not transfer. Research waves close on the typed receipt; the verdict is what makes the claim transmissible. There is no analogue of a release bundle.

4. Triage protocol-gap findings to ε

A repeated finding across waves (citation drift, construct drift, recurring oracle ambiguity) is ε's input, not γ's to silently absorb. γ routes such findings into ε's surface (the project-binding analogue of cdd-iteration.md).

Persona / protocol / project boundary

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

• Who is doing the coordination (persona — layer 1). The persona hub's ## Discipline section names the persona's loss function; this overlay names the role function.
• What concrete project's wave-tracking surface is used (project — layer 4). The wave-manifest path, the receipt file format (Markdown vs CUE vs JSON), the project's policy file — all project-binding. This overlay names the typed #CDRReceipt shape per the schema; the project binds where it lives and how it is rendered.

Resumption

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

• Wave γ-closeout / γ-scaffold analogues — if γ resumes mid-coordination, read the wave's existing artifacts and continue from the next uncompleted section.
• Multi-wave dispatch coordination — if γ resumes a wave-set with partial state, read the wave manifest's current row state and continue from the next pending wave.

Never restart completed sections. Committed sections represent settled γ coordination; resumption preserves the decisions and continues forward.