# WIZARD_SCORES_FABLE — Fable's cross-scores of GPT-SOL-5.6's five revised finalists

Scorer: Fable wizard. Inputs read: `WIZARD_IDEAS_GPT_SOL_5_6.md`,
`WIZARD_REFINERY_ADDENDUM_GPT_SOL_5_6.md` (treated as final revision where it
changes the original), `AGENTS.md`, `EVIDENCE_PACKET.md`. No score/reaction/
rebuttal files were read.

**Rubric:** Robust, Reliable, Performant, Intuitive, User-friendly, Ergonomic,
Useful (×2), Compelling, Accretive (×1.5), Pragmatic (×2). Composite = weighted
mean × 10, /1000. Complexity budget, role duplication, authority leakage, and
cathedral risk are priced inside the dimensions (chiefly Reliable, Intuitive,
Pragmatic, and Accretive) and called out per finalist.

## Dimension scores (0-100)

| Dimension (weight) | A. Lean Steward | B. Headless Graph | C. Portfolio/Quest | D. Governor+Court | E. Two-Key Court |
|---|---:|---:|---:|---:|---:|
| Robust (1) | 88 | 88 | 78 | 84 | 90 |
| Reliable (1) | 90 | 88 | 76 | 74 | 76 |
| Performant (1) | 75 | 80 | 84 | 62 | 52 |
| Intuitive (1) | 88 | 90 | 60 | 58 | 84 |
| User-friendly (1) | 90 | 72 | 76 | 68 | 74 |
| Ergonomic (1) | 85 | 78 | 66 | 62 | 72 |
| Useful (2) | 93 | 74 | 66 | 62 | 60 |
| Compelling (1) | 88 | 76 | 70 | 60 | 66 |
| Accretive (1.5) | 90 | 92 | 78 | 64 | 72 |
| Pragmatic (2) | 84 | 88 | 56 | 48 | 54 |
| **Composite /1000** | **874** | **827** | **697** | **627** | **680** |

---

## A. Lean Steward + Fresh Refinery — **874 / BUILD**

- **Verdict:** Build. This crosses 700 honestly: it is the architecture I would
  actually prioritize — persistent Mayor as sole named session (proposals only),
  fresh author-distinct judges, deterministic reducer/lease/worktree spine,
  Refinery as durable logical identity + formula state machine + min-0/max-1
  fresh triage pool.
- **Strongest contribution:** the delivery-plane mechanical rigor. The fenced
  delivery lease with monotonic tokens owned by the *delivery record* rather
  than any session (addendum §"One-writer integration lease"), the
  landed-SHA-verified delivery receipt, and the explicit rejection of the
  pinned pack's "file the pre-existing failure and merge anyway" behavior
  (addendum CI-routing table, last row) are the strongest single-writer
  enforcement in either proposal.
- **Fatal/limiting concern:** validation economics. A requires a fresh
  validation after *every* candidate merge (I1, I2, I3 …) **plus** a mandatory
  fresh Codex+Claude pair on every PR head, re-run on *every* head/base
  mutation including each rebase as `main` moves. On a rebase-churny week this
  is O(candidates + rebases) validator invocations per wave; judge capacity is
  the scarce resource and their own E-kill criterion (validator queue >35% of
  wall time) will likely trip inside A itself. The safety story is airtight;
  the cost story is under-priced (their own Cost score of 78 is generous).
- **Irreducible-role test:** passes for every persistent role. Mayor earns its
  session with an honest kill (80%-mechanical-actions absorption rule);
  Refinery persistence is correctly refused — durable identity + fresh wakes;
  judges stay fresh. Clean.
- **Steal/combine:** fencing tokens on every integration write; delivery
  record as lease owner; per-candidate incremental validation *as a
  policy-selectable mode* (it localizes the first incompatible candidate,
  which my per-train validation cannot); draft dependent PRs with
  `depends_on_pr` metadata for pipeline visibility; the
  attest-current-main-and-PR-head step before enabling auto-merge; the
  `degraded_provider_independence` receipt.
- **±100 condition:** −100 if the universal final two-provider pair plus
  per-merge intermediate validation measurably consumes >35% of wall time or
  ~2× cost per admitted node in the proof week (their own E-kill number,
  which A's addendum policy quietly imports). +100 if intermediate validation
  is made stakes-tiered (deterministic classification, pair only on
  high-stakes/final heads) with no escaped defect in the proof week —
  that fix makes this a ~950 architecture.

## B. Headless Graph + Fresh Refinery — **827 / BUILD (as the substrate + control)**

- **Verdict:** Build — with the precise role their own decision rule assigns
  it: the mechanism plane everyone builds first, and the control experiment
  that keeps the Mayor honest. Above 700 because I would genuinely prioritize
  building it: it is A minus one configured agent, and every artifact is
  reused.
- **Strongest contribution:** the collapse invariant — "build A's mechanism
  plane so deleting the Mayor yields B rather than a rewrite" (final decision
  rule). That single sentence is the best anti-cathedral device in the file:
  it makes the persistent role *removable by config*, which converts the
  Mayor debate from doctrine into a measurement.
- **Fatal/limiting concern:** it does not answer the decision question. Bo
  asked to hand a large plan to a Mayor and steer mid-flight; B's replan path
  is human + fresh planner + structured holds, and its own kill criterion
  ("operator repeatedly asks for a continuous conversation surface")
  predicts its most likely outcome. As primary it under-serves; as fallback
  it is exactly right.
- **Irreducible-role test:** vacuously passes — zero persistent LLM roles.
- **Steal/combine:** the structured hold contract ("exact node, digest,
  failed invariant, permissible next actions; not a chat request") — my F1/F5
  escalation surfaces were vaguer; planner-reproducibility and
  graph-revision-churn as first-class metrics.
- **±100 condition:** +100 if the proof week shows Mayor actions were ≥80%
  mechanical mirrors of graph state (their A-kill fires and B becomes the
  *proven* primary). −100 if human clarification time exceeds their own
  15-minute/DAG budget — the topology then fails its own pass bar.

## C. Portfolio/Quest Split + Fresh Refinery — **697 / DEFER (correctly gated by its own author)**

- **Verdict:** Defer. Below 700 and that is the point: as a thing to build
  *now* it fails the complexity budget; as a documented, disabled escalation
  path it is fine. Their own migration step 3-4 (Apollo template ships
  `min_active_sessions = 0`, activated only at ≥3 concurrent quests or
  measured Mayor SLO breach) is the right discipline — but a gated future
  option scored as a finalist must be scored as what it builds.
- **Strongest contribution:** the most disciplined Apollo ever specified in
  this corpus: quest-digest binding, hard TTL, context budget, one-owner
  quest lease, provider-agnostic rollover from durable state, and a
  reversal-rate kill (>15% of proposals reversed by Mayor). If a conductor
  tier is ever justified, this is its contract.
- **Fatal/limiting concern:** history plus arithmetic. Apollo autonomy was
  never production-proven through two repo generations
  (`mt-olympus/docs/canonical/olympus-pantheon-synthesis.md:274`), and the
  split adds three handoff boundaries (Mayor→Apollo→Refinery) around a
  formula executor that already sequences ready nodes. The nested two-DAG
  state model, cross-quest path-ownership transfer edges, and per-quest
  integration digests are a real second machine — role duplication cost
  paid before any queueing evidence exists.
- **Irreducible-role test:** Mayor passes. Apollo passes only conditionally —
  *because* it starts disabled and TTL-bound. As a default-on role it would
  fail test (c): its dispatch half is the formula, its replan half is the
  Mayor's loop at a finer grain. Their own cathedral section concedes this.
- **Steal/combine:** the activation gate itself (≥3 concurrent quests OR
  measured dispatch-latency SLO breach) is a better-quantified trigger than
  my F4 gate ("Mayor context recycles >1×/day"); quest-lease
  duplicate-conductor rejection as a mechanical guard.
- **±100 condition:** −100+ if built before three genuinely concurrent quests
  exist (pure cathedral spend). +100 only after F2/A-shape saturation
  evidence — at which point it is no longer premature and its throughput 84
  starts earning.

## D. Governor and Court + Fresh Refinery — **627 / KILL-CANDIDATE**

- **Verdict:** Reject as a build target; run at most as the shadow-mode
  experiment its own migration path describes. This is the consensus-kill
  candidate.
- **Strongest contribution:** the shadow-mode discipline — sidecars only,
  compare unique-find yield against fresh judges and deterministic audits,
  enable reducer-consumed HOLD only after shadow evidence, seeded-drift
  recall as the pass metric. If Bo insists on trying a persistent Witness,
  this is the only honest way to do it. The seeded system-level defect
  battery (omitted acceptance edge; two locally-valid nodes whose
  combination violates intent) is a genuinely good test design I'd reuse for
  *fresh* audit orders.
- **Fatal/limiting concern:** authority leakage plus context gravity, priced
  simultaneously. Post-addendum, D runs **three** recurring observers (Mayor,
  Governor, Refinery) around one evidence stream — their own words: "plainly
  overbuilt." The Governor's HOLD is blocking authority granted to the one
  role whose value is unproven, and `assurance_degraded` makes Governor
  *availability* a product-admission dependency — a persistent role that must
  be up before work ships is precisely the resident-layer failure mode
  (`mt-olympus/FINAL-POSTMORTEM.md:106`). Reliability and cost both eat it.
- **Irreducible-role test:** **fails for the Governor**, and their own text
  admits it "barely passes." I score it as not passing: longitudinal drift
  detection does not require a *resident* context — a scheduled fresh audit
  over durable digests/verdicts/coverage maps has the same longitudinal
  inputs (state lives in artifacts, not transcripts) without saturation.
  The merge-with-neighbor counterfactual is answered by "fresh audit order,"
  which their table never seriously prices as the neighbor.
- **Steal/combine:** the seeded cross-wave defect suite and the
  false-HOLD-rate <10% / unique-find-yield metrics — attached to my
  *fresh* sampling auditor instead of a resident.
- **±100 condition:** +100 if a real shadow week detects ≥3 of 4 seeded
  system-level defects **plus** ≥1 confirmed unique non-seeded finding that
  fresh judges and deterministic checks all missed (their own pass bar) —
  that is the only evidence that would justify residency. Absent that, this
  stays killed.

## E. Two-Key Court + Fresh Refinery — **680 / REJECT AS DEFAULT, KEEP AS POLICY KNOB**

- **Verdict:** Reject as a universal default; retain as the high-stakes
  admission policy. Below 700 because post-addendum its marginal value
  collapsed *by their own design*: A's delivery policy already mandates a
  fresh Codex+Claude pair on every merge-eligible PR head, so E's remaining
  delta is dual-provider judgment on every intermediate/routine node —
  nearly pure cost (their Cost: 43) purchasing marginal safety on the
  lowest-stakes work.
- **Strongest contribution:** making the council mechanical instead of
  prose — pair assembly, attested identities, PASS/PASS-or-hold, no Mayor
  tie-break, no permanent tie-break provider, re-panel-then-human. That
  deterministic pair-resolution policy is the right *shape* wherever dual
  review is warranted.
- **Fatal/limiting concern:** availability coupling and queue arithmetic.
  Either provider's outage halts *all* admission (their own kill: >4h outage
  blocking routine work); validator capacity is doubled on exactly the nodes
  least likely to carry cross-provider-visible defects. Risk-tiering
  (A's/my quorum-on-demand with deterministic stakes classification) buys
  ~90% of the safety at ~half the judge spend.
- **Irreducible-role test:** passes — Mayor is the only persistent role; the
  court is fresh pools plus a deterministic gate. The flaw is policy cost,
  not role legitimacy.
- **Steal/combine:** the provider agreement matrix and false-fail
  confirmation rate as standing metrics (they make the risk-tiering
  threshold *empirical* — if the cross-provider unique-catch rate on routine
  nodes is ~0 over a month, tiering is vindicated; if not, escalate more).
- **±100 condition:** +100 if a provider-specific defect escapes a
  risk-tiered pipeline into `main` during the proof period (evidence that
  tiering under-triggers — E's premium becomes justified). −100 if provider
  outage blocks routine admission >4h in week one (their own kill fires).

---

## Ranking, consensus-kill, and winner-vs-winner

**Ranking:** A (874) > B (827) > C (697) > E (680) > D (627).

**Consensus-kill candidate: D's persistent Assurance Governor.** Both my
independent analysis (my F3, scored last among my viable options with
cathedral-resistance 45-48) and their own cathedral section ("the most likely
finalist to recreate Olympus… deletion, not prompt refinement") converge:
the persistent read-only Witness fails the irreducible-role test and survives
only as a bounded shadow experiment with a one-week yield bar. That is as
close to consensus as two non-coordinating wizards get.

**Does their winner beat mine?** Their A and my F2R are convergent to a
striking degree — one persistent Mayor (proposal-only, reducer-executed),
fresh author-distinct judges, deterministic scope/lease/reducer spine,
Refinery as durable identity + fresh-wake pool + state-machine formula, wave/
train PRs, no direct `main` pushes, semantic repairs as new AgentOps
experiments. The differences are real but second-order:

- **Where theirs is better:** delivery-plane mechanics. Fencing tokens,
  delivery-record-owned leases, landed-SHA receipts, draft dependent PRs, the
  merge-queue speculative-subject clause, and the pre-existing-failure
  non-waiver rule are more enforceable than my §3/§7 equivalents. Their
  per-candidate incremental validation also localizes the first bad candidate,
  which my per-train validation cannot.
- **Where mine is better:** judgment economics. My deterministic stakes
  classification + quorum-on-demand + per-train validation spends scarce
  validator capacity where risk lives; their A hard-codes a universal final
  dual-provider pair re-run on every rebase plus per-merge intermediate
  validation — the cost profile of their own rejected E leaking into their
  winner.

Candid net: **their A beats my F2R by a small margin on mechanical
enforceability, and loses to it on operating cost; neither dominates.** Scored
on their rubric weights I'd put my F2R at ~880 and their A at 874 — inside
the noise floor of self-scoring bias, so I do not claim a win. The correct
duel outcome is a merge: their delivery state machine, fencing, receipts, and
shadow/collapse disciplines, with my stakes-tiered validation policy and
train-cap sizing replacing their universal pair + per-merge validation. That
merged architecture is the one I would build Monday.
