# ARCHON SFT v2 Phase 6.5 — Readiness Validation Report

Date: 2026-06-02
Validator: Claude Sonnet 4.6 (subagent)
Scope: 7 Phase 6.5 artifacts, cross-validation of 5 criteria

---

## VERDICT: NOT READY — 3 blocking conflicts, 1 data source defect

Synthetic generation must NOT launch until conflicts C1 and C2 are resolved.
Conflict C3 and defect D1 can be fixed in parallel with generation prep.

---

## VALIDATION RESULTS

### V1 — HF Search Results

Status: PASS (with noted risk)

Single usable peer-equal dataset identified: `Multi-Agent-LLMs/DEBATE` (apache-2.0).
7,200 usable rows across debate + relay + memory + report paradigm configs.
Peer-equal check confirmed: no supervisor/manager/orchestrator pattern in sampled rows.
Rejection of camel-ai datasets on NC licence + structural hierarchy grounds is correct.
UltraInteract_sft classification as N1-only is correct (SFT split is flat, no multi-agent turns).

Risk noted: DEBATE tasks are trivia/QA domains (manta ray biology, seating puzzles), not
ARCHON-domain (network/cybersec/Etherlink). Format transfer is accepted but domain gap is real.
Mitigation already documented in adversarial check section — adequate.

### V2 — Synthetic Generation Plan Feasibility

Status: PARTIAL PASS — cost reasonable, hours reasonable, sample targets INCONSISTENT (see C1/C2)

Cost: $220-250 USD (generation plan estimate with prompt caching is the reliable figure;
the YAML $125 figure underestimates by using a lower $/call approximation).
Hours: 28-35h sequential, ~18-20h with 2 parallel workers. Feasible.
Pilot gates (V1-V4) are well-designed and proportionate.
Parallel launch strategy (N1+N4 terminal 1, N2 terminal 2, N3 terminal 3) is sound.

### V3 — MTP Task Profiles Consistency

Status: PASS

4 new task types added correctly:
  family_consult     [1.0, 0.65, 0.50, 0.40, 0.30]  (N2 structured — moderately predictable)
  family_negotiate   [1.0, 0.55, 0.40, 0.30, 0.20]  (N3 debate — mid-predictability)
  family_swarm       [1.0, 0.40, 0.25, 0.18, 0.12]  (N4 emergent — low predictability)
  family_role_rotate [1.0, 0.70, 0.55, 0.45, 0.35]  (meta escalation — structured)

Predictability gradient N2>N3>N4 is correct and rationale is documented.
`general_conv_fr` was NOT duplicated (already present). Comment confirms this.
Phase 6 submix sum assertion: 0.01+0.04+0.06+0.04 = 0.15 — PASS.
Curriculum continuity: 7 phases, steps 0-50000, no gaps, no overlaps — PASS.
DPO sub-window 42000-50000 aligned in both Python file and design doc — PASS.
Legacy `multi_agent_coord`/`delegation_chain` kept at low weight (0.8) for rehearsal — correct.

### V4 — Mix Distribution Target Coherence

Status: FAIL — sample target inconsistency across three documents (see C1, C2)

### V5 — METATRON / "Master" Connotation Audit

Status: CONDITIONAL PASS — one spec doc has a METATRON leak (see C3)

"master" usage in all Phase 6.5 artifacts is definitional negation only (describing what
to avoid). No content encodes master-subordinate behavior. File name
`_out_synthetic_prompts_master.md` is a title, not a connotation issue.

Generator script: METATRON correctly replaced by `[trading-specialist]` in ROSTER,
SYSTEM_BASE, and anti-pattern list. Forbidden-pattern validator checks "metatron" string.
Generation plan gate V2 checks "metatron" absence. PASS.

`_out_phase_6_5_design.md` line 75: schema example lists `METATRON` as a participant:
  "participants": ["ARCHON", "<1-3 peers from NEXUS/CYPHER/AETHER/METATRON>"]
This is a design doc inconsistency — the actual generator script never uses METATRON.
Risk: low (design doc is not read by the generator), but must be corrected before sharing.

### V6 — Coverage: HF + Synthetic vs Total Target

Status: FAIL — N4 coverage is severely under-specified (see C2)

---

## CONFLICTS (blocking unless noted)

### C1 — BLOCKING: N4 sample count three-way inconsistency

| Document | N4 target |
|---|---|
| _out_multi_agent_sources.yaml | 10,000 (to_generate) |
| _out_synthetic_generation_plan.md | 3,000 |
| _out_mix_distribution_target_phase65.json | 20,000 (target_samples in N4_family_swarm) |

The mix JSON is the authoritative training target: 20,000 N4 samples.
The YAML says 10,000 to generate (which could mean HF covers 10K — but HF has 0 N4 rows).
The generation plan says only 3,000 — this is the most severe undercount.

Resolution required: set N4 target uniformly to 10,000 across all three documents
(20,000 mix target is too aggressive for entirely-synthetic N4; 10K is the YAML intent
and matches the HF search gap analysis). OR keep mix at 20K and update synth plan to 20K
(cost ~$200 for N4 alone via Opus templates + Sonnet expansion, per YAML estimate).

Decision needed from Jescy before launching.

### C2 — BLOCKING: Aggregate N2-N4 coverage gap

Computed from authoritative sources:

| Niveau | HF available | Synth plan generates | Total | Mix JSON target |
|---|---|---|---|---|
| N2 | 2,400 | 5,000 | 7,400 | 20,000 |
| N3 | 3,600 | 7,000 | 10,600 | 30,000 |
| N4 | 0 | 3,000 | 3,000 | 20,000 |
| **Total N2-N4** | **6,000** | **15,000** | **21,000** | **70,000** |

21,000 actual vs 70,000 mix target = 30% coverage. This is a 3.3× shortfall.
Possible resolutions:
  (a) Reduce mix JSON targets to match synth capacity (N2: 7.5K, N3: 10.5K, N4: 3K or 10K).
      Pro: preserves current synth budget. Con: N4 remains thin.
  (b) Scale up synth generation to match mix targets (cost ~$800-1000 total, 3-4 days).
      Pro: full coverage. Con: significant cost and time increase.
  (c) Accept mismatch with oversampling: synth samples x10 weight to fill quota.
      Risk: memorisation at high oversampling factor.

### C3 — NON-BLOCKING: METATRON in design doc schema

`_out_phase_6_5_design.md` line 75 lists METATRON as a possible participant in the
generator schema example. The generator script never actually uses METATRON (correctly
uses `[trading-specialist]`). Fix: replace METATRON with `[trading-specialist]` in the
schema example. One-line edit, no functional impact.

---

## DEFECTS (non-blocking but must fix before mix build)

### D1 — UltraInteract_sft N2 filter is broken

Mix JSON entry for N2_family_consult uses UltraInteract_sft with filter
`trajectory_length >= 4`. The HF search report (section 2) explicitly confirms:
"The parent_id tree is the preference learning split (_pair), NOT this SFT split."
The SFT split is flat (instruction, response) pairs — no `trajectory_length` field exists.
This filter will either fail at runtime or return 0 rows.

Impact: 8,000 planned N2 samples from this source will be unavailable.
Mitigation: remove UltraInteract N2-filtered source from N2 mix entry. Replace with
additional synth generation (+3,000 N2) or oversample DEBATE relay/report configs.
Net effect on C2 gap: worsens N2 shortfall by 8K additional samples.

---

## SUMMARY TABLE

| Check | Status | Action |
|---|---|---|
| HF peer-equal datasets identified | PASS | None |
| Synthetic cost feasible | PASS | Use $220-250 budget |
| Synthetic hours feasible | PASS | 28-35h sequential |
| MTP profiles consistent | PASS | None |
| Phase 6 submix sum = 15% | PASS | None |
| Curriculum continuity 0-50K | PASS | None |
| DPO sub-window alignment | PASS | None |
| N4 sample count aligned | FAIL — C1 | Resolve before launch |
| N2-N4 coverage vs mix targets | FAIL — C2 | Resolve before launch |
| METATRON absent in generated content | PASS | None |
| METATRON in design doc schema | WARN — C3 | One-line fix |
| UltraInteract N2 filter valid | FAIL — D1 | Remove/replace before mix build |
| DPO egalitarian dependency order | PASS | None |

---

## TOTAL COST ESTIMATE (Phase 6.5 complete)

Scenario A — current synth plan (17K samples):
  Synthetic generation:      $220-250
  GPU cost Phase 6 (8K steps, RTX PRO 6000 ~$1.2/hr, ~8-10h): $10-12
  Total Phase 6.5:           ~$230-262

Scenario B — scaled to match mix targets (~54K N2-N4 samples):
  Synthetic generation:      ~$800-1000
  GPU Phase 6:               ~$10-12
  Total Phase 6.5:           ~$810-1012

Recommended: resolve C1/C2 first, then choose scenario.

---

## PRE-LAUNCH CHECKLIST

Before running `python _out_synthetic_generator.py`:
- [ ] C1 resolved: N4 target aligned across all three documents
- [ ] C2 resolved: synth targets or mix JSON targets reconciled
- [ ] C3 fixed: METATRON replaced by [trading-specialist] in _out_phase_6_5_design.md line 75
- [ ] D1 acknowledged: UltraInteract N2 source removed from mix JSON or filter corrected
- [ ] Dry-run passes: `python _out_synthetic_generator.py --niveau all --dry-run`
- [ ] Pilot 50 samples N2 inspected manually before full run
- [ ] ANTHROPIC_API_KEY set in environment
- [ ] synthetic_out/ and logs/ directories created