# GPT-5.6-sol cross-score of Fable's revised finalists

## Method

This score treats `WIZARD_REFINERY_ADDENDUM_FABLE.md` as an override wherever
it changes the original proposal. I scored the resulting F1R–F5R, not the
pre-addendum versions and not the Refinery as a sixth finalist.

Each dimension is 0–100. The composite is the rounded weighted mean multiplied
by ten: Useful and Pragmatic are 2×, Accretive is 1.5×, and the other seven
dimensions are 1×, for total weight 12.5. Thus a score above 700 means I would
actually prioritize building the option, not merely admire it. The governing
standard is deliberately hostile to role ceremony: a resident LLM must own an
irreducible, bounded, measurable judgment loop that cannot be a formula, lease,
reducer, worktree rule, or mode
([AGENTS.md](/Users/bo/dev/gc-role-duel/AGENTS.md:18)). The mandatory baseline
also includes isolated branches/worktrees, fresh exact-subject validation, one
integration reducer, protected `main`, and revalidation after rebases
([EVIDENCE_PACKET.md](/Users/bo/dev/gc-role-duel/EVIDENCE_PACKET.md:246)).

### Score matrix

| Rank | Revised finalist | Robust | Reliable | Performant | Intuitive | User-friendly | Ergonomic | Useful (2×) | Compelling | Accretive (1.5×) | Pragmatic (2×) | Composite |
|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| 1 | F2R — Mayor + fresh judges + Refinery | 82 | 82 | 78 | 87 | 86 | 83 | 93 | 90 | 84 | 78 | **845** |
| 2 | F1R — Bare City + Refinery | 78 | 76 | 72 | 73 | 55 | 60 | 68 | 60 | 90 | 82 | **727** |
| 3 | F5R — Stigmergy factory + Refinery | 77 | 70 | 78 | 61 | 48 | 63 | 62 | 55 | 85 | 70 | **675** |
| 4 | F3R — Mayor + persistent governor + Refinery | 76 | 74 | 70 | 60 | 75 | 62 | 73 | 74 | 50 | 52 | **653** |
| 5 | F4R — Mayor + quest conductors + Refinery | 70 | 65 | 80 | 50 | 60 | 52 | 62 | 68 | 48 | 44 | **583** |

All five inherit a materially useful delivery design: candidate branches feed
a bounded integration train, one PR represents the train, and every subject
mutation regenerates, checks, and revalidates
([Fable addendum](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:60),
[mutation loop](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:159)).
They also inherit four unresolved seams, priced into Robust, Reliable, and
Pragmatic rather than waved away:

1. The shared bot identity is the standing authority while a TTL lease is held
   by a transient session, but no fencing token or compare-and-swap push rule is
   specified. After reap, a stale process using the same credential can still
   race the replacement
   ([identity](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:27),
   [lease](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:112)).
2. The high-stakes path requests a “2-of-3 family quorum” with author-family
   exclusion, but the stated deployment has only Codex and Claude providers;
   provider is configuration, not another role
   ([quorum rule](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:172),
   [provider boundary](/Users/bo/dev/gc-role-duel/EVIDENCE_PACKET.md:40)).
3. Intake requires “exactly one” PASS while the same design can require several
   judge outputs. It needs an explicit deterministic quorum certificate that is
   singular without erasing the immutable component verdicts
   ([intake](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:134)).
4. The Refinery identity authors the PR, while validator artifacts may satisfy
   required review. Unless the forge policy represents validation as a separate
   required check or a distinct reviewer identity, this can collapse author and
   approver at the repository gate
   ([PR authority](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:27),
   [review gate](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:228)).

These are correctable specification gaps, not reasons to discard the shared
Refinery. They matter because the packet explicitly requires atomic claims,
independent judges, immutable judgments, and a non-bypassable protected-main
transition
([EVIDENCE_PACKET.md](/Users/bo/dev/gc-role-duel/EVIDENCE_PACKET.md:257),
[delivery rules](/Users/bo/dev/gc-role-duel/EVIDENCE_PACKET.md:272)).

## 1. F2R — Mayor + fresh judges over a deterministic spine + Refinery

**Score: 845/1000 — STRONG; PRIORITIZE.** I would build this, after closing the
four shared Refinery seams above.

- **Strongest contribution:** it puts the one continuous semantic loop where it
  belongs. The Mayor preserves product conversation, reshapes the experiment
  DAG from verdict feedback, and proposes retries, while fresh plan reviewers,
  fresh verdict judges, and a deterministic reducer retain independence. The
  Mayor cannot implement, judge, or write terminal state
  ([original F2](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:287),
  [authority matrix](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:330)).
  The revised Refinery then owns delivery without moving Git into AgentOps or
  moving delivery into the Mayor
  ([addendum ruling](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:23)).
- **Fatal or limiting concern:** the one-writer promise is not yet a one-writer
  guarantee. A shared Git credential plus a reapable lease needs fencing/CAS at
  every branch update; otherwise a stale fresh-wake session can mutate the train
  after losing ownership. The infeasible quorum wording and ambiguous review
  identity are additional launch blockers, though neither changes the topology.
- **Irreducible-role result:** **yes.** The Mayor passes all five tests: unique
  recurring replan/human-interface judgment, observable absence failure,
  non-mechanical work, proposal-only bounded authority, and explicit SLO/kill
  criteria. Refinery is a persistent authority identity but not a persistent
  LLM; its ambiguous failure-classification work is a min-0/max-1 fresh pool
  ([Refinery runtime](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:27)).
- **What I would steal/combine:** take the capped train—up to five experiments
  or a diff ceiling—and the negative worker-push test, then combine them with my
  fenced delivery record, compare-and-swap branch updates, exact-head dual-
  provider validation, and separate PR-author/policy-check identities
  ([train cap](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:101),
  [proof negatives](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:335)).
- **100+ swing condition:** **lower by at least 150** if the proof week produces
  even one post-reap push from a session that no longer owns the integration
  lease. That is a broken one-writer guarantee, not ordinary tuning.

## 2. F1R — Bare City + Refinery

**Score: 727/1000 — STRONG; PRIORITIZE AS THE BOOTSTRAP/FALLBACK.** This clears
700 because its deterministic substrate is exactly what F2R needs and I would
prioritize building it first. I would not stop here if Bo must converse with a
city about a changing large plan.

- **Strongest contribution:** it is a clean falsification instrument for every
  proposed resident role. A fresh plan compiler, four fresh worker/validator
  pools, formulas, leases, reducer, and the revised Refinery can prove safe fan-
  out and delivery without pre-paying for a government
  ([F1](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:251),
  [shared substrate](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:198)). It
  is also an honest absorption target if the Mayor later fails its week.
- **Fatal or limiting concern:** mid-flight semantic replanning has no owner.
  The proposal explicitly sends every change back through Bo and a new plan-
  compile context; state can remain durable, but conversational intent and the
  reasons for changing the DAG do not become somebody's closed loop
  ([F1 weakness](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:279)). That is
  directly misaligned with the request to hand a large plan to a Mayor and get
  feedback without babysitting
  ([decision question](/Users/bo/dev/gc-role-duel/EVIDENCE_PACKET.md:3)).
- **Irreducible-role result:** **yes, vacuously.** There are no persistent LLM
  roles. The Refinery's durable Git identity is not an LLM session; the fresh
  triage pool remains bounded to delivery classification.
- **What I would steal/combine:** use F1R as the first two-day deployment slice
  and as my winner's kill destination. Keep its formulas, worktree/scope
  enforcement, reducer, and delivery engine unchanged when adding the Mayor.
- **100+ swing condition:** **raise by at least 110** if an adversarial proof
  week includes two material mid-flight plan changes yet Bo spends no more than
  15 minutes per day on dispatch/re-explanation and every fresh plan compile
  preserves prior decision rationale. That would disprove the need for Mayor
  continuity in this workload.

## 3. F5R — Stigmergy factory + Refinery

**Score: 675/1000 — DECENT; EVALUATE, DO NOT PRIORITIZE YET.**

- **Strongest contribution:** pull-based self-claiming plus durable beads makes
  session death cheap and pool scaling natural. The current queue state, not an
  agent's memory, is the coordination surface; the Refinery is likewise queue-
  shaped, so the composition is conceptually accretive
  ([F5](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:438),
  [addendum impact](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:254)).
- **Fatal or limiting concern:** the scheduled fresh “producer tick” is a Mayor
  loop split into stateless invocations without a human interface. It must make
  semantic retry/revision decisions, but no persistent owner carries the
  product conversation or reconciles two individually plausible ticks. This is
  mechanism-shaped orchestration until the first ambiguous failure, then an
  unowned judgment loop
  ([producer tick](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:445),
  [known failure mode](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:461)).
- **Irreducible-role result:** **yes, vacuously.** No persistent LLM is present.
  That does not prove the architecture complete; it exposes the missing
  recurring replan owner.
- **What I would steal/combine:** keep pull claims, durable sidecars, and fresh
  resumption as worker-pool mechanics inside my Mayor topology. Do not make the
  queue itself responsible for semantic replanning.
- **100+ swing condition:** **raise by at least 110** if three independent
  FAIL→revised-intent cycles complete through producer ticks with zero duplicate
  or contradictory intents, zero Bo session attachments, and plan-review PASS
  over each revised DAG. That would show continuity is sufficiently encoded in
  durable state.

## 4. F3R — Mayor + persistent Witness governor + Refinery

**Score: 653/1000 — DECENT INHERITED CORE; KILL AS A DISTINCT TOPOLOGY.**

- **Strongest contribution:** it names a legitimate question that simpler
  designs can miss: who notices cross-experiment goal drift and degrading judge
  independence? Its finding-precision and acted-on-finding metrics are better
  than ceremonial “oversight” claims
  ([F3](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:376)).
- **Fatal or limiting concern:** the answer does not require a resident LLM.
  Identity/independence is deterministic, and semantic drift can be sampled by
  fresh read-only audits. A persistent observer adds cost, saturated context,
  another personality to monitor, and an obvious path toward veto/replan/repair
  authority. Fable itself says the role likely fails the test and predicts
  collapse into F2's sampling audit
  ([self-critique](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:386),
  [cathedral argument](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:496)).
  The addendum gives it more PR state to watch but concedes that scheduled fresh
  reads still cover it
  ([revision](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:271)).
- **Irreducible-role result:** **no.** The Mayor passes; the persistent governor
  fails conditions 2 and 3 because its observable outputs can be produced by
  deterministic attestation plus a scheduled fresh audit. Therefore not every
  persistent role passes.
- **What I would steal/combine:** steal the governor's audit prompt, finding-
  precision metric, and system-goal-drift lens; schedule them as fresh assurance
  jobs in my winner. Do not steal the named session or any veto authority.
- **100+ swing condition:** **raise by at least 120** only if, across two real
  plans, the governor finds at least two high-severity issues that are acted on
  and that replayed deterministic checks plus blinded fresh sampling audits both
  fail to find. That would establish unique value from continuity rather than
  prompt quality.

## 5. F4R — persistent product Mayor + per-quest conductors + Refinery

**Score: 583/1000 — DECENT ONLY BEHIND A MEASURED SCALE GATE; DO NOT BUILD NOW.**

- **Strongest contribution:** it is the only finalist with a credible scale-out
  escape hatch for semantic conduction. Conductors are fresh per quest rather
  than permanent area sub-mayors, and Fable correctly gates adoption on an
  observed Mayor scale wall
  ([F4](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:406)). The addendum also
  correctly refuses per-quest Refineries, preserving one integration seam
  ([F4R revision](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:274)).
- **Fatal or limiting concern:** product replanning and quest replanning are the
  same semantic loop at different grain. The product/quest boundary becomes a
  judgmental authority handoff: who decides that a validator result changes
  only a quest rather than the product DAG? The deterministic dispatcher already
  owns mechanical sequencing, so the new tier buys ambiguous authority and
  context-transfer cost before measured saturation. This is precisely the
  unproven Apollo-shaped cathedral risk the proposal itself identifies
  ([F4 limits](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:414),
  [cathedral finding](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:515)).
- **Irreducible-role result:** **yes for persistent roles, with an important
  caveat.** The persistent Mayor passes. Conductors are fresh ephemeral roles,
  so they do not violate the literal resident-session test, but they fail the
  broader duplication test until scale evidence proves that a Mayor prompt mode
  cannot handle the same loop.
- **What I would steal/combine:** keep a dormant `quest-conduct` prompt mode and
  its scale trigger in my winner's runbook. Activate it only after Mayor replan
  latency or context recycling crosses a measured threshold; never deploy it as
  the default tier.
- **100+ swing condition:** **raise by at least 140** if F2R first demonstrates
  a sustained scale wall—more than 20 concurrent lanes and Mayor context recycle
  over once per day—and then three quest conductors complete three nontrivial
  subgraphs without the Mayor reading inner beads, reversing a verdict, or
  repairing their handoffs.

## Final ranking and duel judgment

1. **F2R — 845:** build-worthy winner after delivery-seam hardening.
2. **F1R — 727:** build-worthy deterministic bootstrap and absorption target.
3. **F5R — 675:** promising pool mechanics, incomplete semantic ownership.
4. **F3R — 653:** good F2 core burdened by a role that does not earn residency.
5. **F4R — 583:** retain only as a gated scale experiment.

### Consensus-kill candidate

**F3R is the consensus kill as a distinct topology.** This does not mean its
F2R substrate is bad; it means the only feature distinguishing F3R—the
persistent Judge-Witness governor—should be absorbed into scheduled fresh
assurance. Fable's own irreducible-role analysis says it likely fails and its
own addendum gives the governor no new unique loop
([original verdict](/Users/bo/dev/gc-role-duel/WIZARD_IDEAS_FABLE.md:390),
[addendum verdict](/Users/bo/dev/gc-role-duel/WIZARD_REFINERY_ADDENDUM_FABLE.md:271)).
F4R is a **consensus defer**, not a permanent kill: both proposals preserve the
per-quest mode only if measured Mayor saturation appears.

### Does Fable's winner beat my winner?

**No, but it is close and belongs to the same winning family.** F2R validates
the central conclusion: one persistent proposal-only Mayor, fresh independent
judges, deterministic admission/operations, and a bounded fresh-wake Refinery.
I would prioritize building it over Fable's other four options.

My winner remains preferable because it makes the delivery seam executable,
not merely legible: a durable delivery record carries a fencing epoch; every
integration-branch mutation is lease-epoch checked and compare-and-swap
guarded; the exact final PR head must receive feasible Codex + Claude
independent PASS under separate identities; and the repository gate cannot
interpret the PR author's posting of validator artifacts as self-review. Those
details directly close Fable's shared lease, quorum, intake-certificate, and
review-identity gaps. I would still import Fable's bounded train-size policy,
fresh sampling-audit formulation, and explicit negative tests. That is genuine
convergence with a few high-consequence control-seam corrections, not a reason
to reverse the winner.
