Changelog

Surprise me can no longer roll a "Male" vocal over a soprano seed — the Vocal now matches the seed's voice.

the 2026-10-10 Deep Audit's follow-ups are on the backlog — 5 new rows, 4 folded into existing ones, and the audit's own mistake recorded.

Deep Audit 2026-10-10 filed; the sitemap parses again, the Crucible stops claiming "zero stored", and the pricing and forge pages describe the pipeline that actually runs.

album and Song-Novel track forges now show up in cost telemetry — one row per track.

the admin cover batch can no longer overwrite a good cover with fal's black safety frame; today's builds recorded in the backlog.

the Song-Novel story pass moves to Claude Opus 5.5 — it won the pre-registered blind test 6-0.

the Song-Novel story pass gets its own model task, and its Opus-vs-Sonnet test is pre-registered.

AI cover art no longer saves a black square when fal's safety filter trips — it retries with a people-free prompt, then falls back honestly.

the Tournament Forge's extra drafts read draft 1's prompt cache instead of each re-writing it (~$0.40 off every song).

Forge v5 goes live — every single forge is now the Tournament Forge (3 drafts, blind debiased judging, legibility check, repair only on a named wound).

Forge v5, the Tournament Forge, is built dark and its A/B is pre-registered.

a rejected repair no longer leaves its lyrics in the saved song; the dead cold-reader call is retired.

the gauntlet writer moves to Claude Opus 5.5; refine measured worse on Opus and stays on Sonnet 4.6.

the song writer moves to Claude Opus 5.5 at effort medium — measured first, and only the forge core moves.

make the request layer Claude-5.x-safe — per-model thinking, max_tokens headroom, thinking-free responses, refusal fallback — and arm the Forge Lab to measure Opus 5.5 on the production prompt. No production route changes model.

the forge still 400'd after B4426 — the effort strip never reached the wire. Fixed in the fetch shim and proven against the real API.

forge down in production — roll the model routes back to the last-known-good 4.x generation.

close the two database P0s — scope 12 "service role" RLS policies to service_role, enable RLS on 4 unprotected tables, and stop signed-in users rewriting their own tier/credits.

refresh AUTO-STATE after the cadence sweep

Bet Review #4, External Audit brief #3, and a scoped Deep Pass; all five cadences green

Quality Council + Trust Decay Audit #11, and the seven surfaces B4422 silently falsified

point AUTO-STATE at a real commit

migrate to Claude 5.x generation (Opus 5.5 / Sonnet 5.5 / Haiku 5.5)

new skill songforge-forge-frontend — the /forge client/React conventions the 25-skill knowledge base never covered.

recalibrate the prosody FATAL threshold — the 79%-FATAL /admin/prosody-report reading was a miscalibrated rule, not a catalog-wide prosody emergency.

admin forge-metrics compares real production songs by promptPath — the lean-prompt cutover week-1 review surface (autonomous half).

forge parse-rejects now leave a durable, queryable footprint — the failure rate is measurable in real time instead of blocked on a 30-day Vercel-log export.

the two forge-failure paths fixed — a false-positive regression removed, and the truncation class now auto-retries instead of making the user re-click.

the Substack initiative enters the record — content half done, operator queue tracked.

the newsletter is assembled and claims-cleared — one publication-create away from launch.

the market-and-engage plan enters the record — "Measured Craft", one ICP, ~3 hrs/week, claims-cleared.

the autonomous stretch enters the record — three deep-validity items now wait only on the operator.

the human panel is one Prolific link away — the blinded rating surface ships.

the judge-tier pin lands before anything can breach it, and the counterfactual bracket runner refuses to run unsigned.

the nightly run enters the record — two closes, four partials, the flip-window cluster executed to its unblocked edge.

the lab claims enter the register, the stale flagship claim is struck, and the /vs claim goes on WATCH.

every pre-registered prediction HIT — the corrected gate numbers, the vindicated arbiter, and a dead thread killed for $0.

the unit correction and the pre-registered flip rule — committed BEFORE the measurements run.

the second war room enters the record — 64 agents, 20 rounds, the flip-window cluster and the deep validity pair.

the critic-panel verdict enters the record — parity at worst, a modest edge at best, and it never lost.

the D2b build-half enters the record — the experiment run is the open remainder.

the cross-model critic panel enters the lab — one draft, two wound sources, a blind verdict per cell.

the gate arming enters the record — and this deploy activates SF_GAUNTLET_GATE_AB=1.

the draft-score gauntlet gate ships dark — SF_GAUNTLET_GATE_AB arms the first live A/B in the experiment registry.

the cost program's probe items close against B4399 and the production gauntlet-gate rollout files as MIXED.

the sizing lands and both recommended probes execute — the writer was 12% of the bill, and the program found its halving anyway.

the per-task cost decomposition lands on /admin/cost-per-song — the pipeline-level split the draft-swap prize is sized from.

repair-closure PROVEN — chain output is draft-source-invariant (kimi paired diff literally +0.00 vs the all-Sonnet chain), and the mechanism is selection, not erasure.

D2b files — the operator's cross-model critic panel — and the gauntlet-chain build row closes against B4395.

the gauntlet-chain cell lands — cheap draft + wound-fed house repair, the D2 repair-closure engine — and the Probe B verdict enters the record.

working notes can never be scored as a song again — the scaffolding vocabulary learns the new generation, the lab gains a sectionless-output backstop, and the fallback flag detects the parse path instead of the preamble.

prompt mode is now surfaced at the point of action — run header + start log announce PRODUCTION PROMPT vs neutral.

the SF_FORGE_PROVIDER seam row flips [~] — B4390+B4391 recorded as the shipped build half, routing promotion held open for Rung 5.

production-prompt cell mode lands — Probe B is now a clickable experiment.

the writer-override seam lands in the forge core — dark, attribution-safe, stamp-true — and DeepSeek V4 Pro clears the full screening rule.

DeepSeek V4 Pro outscores the control at 1/24th the cost — the new savings frontrunner — Terra is a cost-parity quality candidate, Luna is eliminated.

KIMI misses strict promotion by half a point — CI lower bound -1.54 vs the pre-registered -1 gate — while GPT confirms a third time and GLM's parity dies under replication.

the candidate-arm row flips [~] — B4386's three arms recorded as the built half, the screening run as the honest remainder.

three candidate arms land — DeepSeek V4 Pro, GPT Terra, GPT Luna — registry 8 -> 11, every price verified live, zero new keys.

the harness gains the replicate-variance decomposition and the Rung-2 promotion rule — the 288-cell verdict becomes one command.