~550metered tokens per real engineering decision

Eight results. Each one links to the experiment that produced it. Nothing below is projected except where marked.

Fable matched the operator 80% of the time. Its successor: 68%.

26 real operator approval cards pulled from o8's own governance history — merge requests, tool confirmations, rebase conflicts, file-size overrides; 14 approved / 12 rejected by the human operator, with verdict leakage scrubbed. Each card was replayed as a ~1KB decisions-only artifact through claude-fable-5 and claude-opus-4-8, in three runs: CLI transport (subscription), API transport (exact tokens), and API with cached doctrine.

• The over-rejection pattern, 3 for 3 runs: every single Fable↔Opus disagreement had the same shape — Fable said APPROVE, Opus said REJECT — and the human operator sided with Fable in the large majority (5/6, 4/7, 4/6). The successor model is systematically more trigger-happy on rejection than both the incumbent and the human. Left alone as sole adjudicator, it would bounce work the operator would have shipped.
• Consequence shipped into the product: after Fable's departure (2026-07-07), contested calls in o8 route to a judge panel — the existing Collide machinery — not a single successor model. A 1-in-4 flip rate is too high for a solo gavel.

*The cached run's doctrine prefix measurably shifted BOTH models' decisions — operator-match dropped ~15 points. Standing doctrine steers the adjudicator; the production window's doctrine text is now a tuned artifact, not an afterthought.

[ TOKEN ECONOMICS PER DECISION — API, EXACT ]

[ VERDICT ]18× fewer total tokens, 26× fewer input tokens, per decision. The whole 52-call eval: 19,480 in / 6,522 out — under a dollar.

Does the window nerf the model? 9.2× less context, zero quality loss measured.

The hardest artifact available: adversarial pre-ship review of the entire Fable-mode change batch itself — ~20 files, ~900 added lines, security-sensitive (billing isolation, tool lockouts, rate limits). Two context conditions, same task prompt, same output contract. RAW is the full change batch in context — the “native frontier” pattern. WINDOW is a ~2K-token digest, the shape o8's digest() produces: per-file roles plus key seam excerpts.

That single raw-context call — 32,720 input tokens — billed more input than the entire 52-call windowed decision eval (19,480). One review, one condition, more than a full succession eval. Measured before the account ran dry.

[ QUALITY — SCORED AGAINST 6 GROUND-TRUTH SUBTLETIES ]

Judge disclosure: findings were scored against 6 ground-truth subtleties known to the batch author. The batch author is the judge.

• Unanimous verdicts across all four conditions — the window did not change the decision.
• The windowed Fable review found real issues the raw-context review missed: the raw-transcript escape hatch limits call count but not output size, and the metered compaction ceiling is enforced client-side only. The raw review found its own uniques — a per-request backend-override mismatch in the Brain gate. Neither condition dominated; both were ship-blocking-grade reviews.

The honest caveat: the digest was authored to be faithful — the same standard digest() is held to in production. A bad digest would produce a blind review, which is exactly why o8 meters fetch_raw instead of banning it, reads adjudicated diffs raw, and spot-audits digests against full context.

Caching the doctrine prefix cut input cost by ~76%.

The window's standing adjudication doctrine — ~1.4K tokens of real o8 governance rules — was set as a cache_control: ephemeral system block, with 52 decision calls against it. Result: 70,928 input tokens served from cache at 0.1× price vs 13,356 at full price — a ~76% input-cost reduction on the doctrine-laden shape.

The caveat that matters: Anthropic caching needs a ~1,024-token minimum stable prefix — thin prompts can't cache, which is exactly why the window batches its standing context into one stable block.

Not a lab number. The whole loop ran live in the o8 app.

1. 01Fable — locked at spawn: no file reads, no shell, no web; 10 native tools plus raw-transcript tools denied — answered repo-governance questions correctly with 6 cited sources, because the org-memory Brain does the reading. A fixed-cost model reads so the metered one never pays to.
2. 02It dispatched a real coding task to a Codex worker; the worker's prompt automatically carried Brain access because the orchestrator is metered.
3. 03When the worker's launcher broke (an environment bug, since root-fixed), Fable didn't retry-spam. It investigated through its dashboards, discovered the operator's suggested fix violated a platform rule — and said so, with citations — parked everything clean, and raised a two-option decision card.
4. 04After the fix it pruned its duplicate packet, redispatched, reviewed the worker's diff RAW (per doctrine — the artifact being approved is never summarized), caught a real nit in the worker's code, passed the 4-check merge gate, and the work shipped as PR #1349.
5. 05Along the way it used fetch_raw — the window's rate-limited raw-content escape — unprompted, exactly as designed.

[ THE NERF VERDICT ]A decisions-only Fable is not a worse Fable. It answered everything a full-tooled session would, made better recovery decisions than it was instructed to make, and its judgment tracked the human operator better than its successor's did.

A benchmark that hides its constraints isn't evidence. These are the ones that keep this honest.

• N=26 decisions and 1 hard task, from a single (real, production) o8 instance — a field measurement, not an academic benchmark suite.
• Cross-model agreement was measured while both models were purchasable — unreproducible after 2026-07-07. That's the point.
• The CLI transport carries Claude Code's ~9.7K-token system prompt per call; on subscription billing this is free, which is itself part of the o8 story: fixed-cost lanes for bulk, a metered lane for decisions only.
• Judge of quality on the hard task: findings were scored against 6 ground-truth subtleties known to the batch author. The batch author is the judge.
• Full raw data: docs/research/fable-succession/run-*.json and hard-task/.

The window, the Brain, the judge panels, and the approval surface measured above ship in o8.

macOS 14+ · Intel build; Apple Silicon via Rosetta