📖About

NINAD, 2026-08-04 (CLI-226): "Bilbo's Speech / Image & Video --> gemini, move the workload on gemini". And: "Only codex-gpt-5.6-sol is currently running, because i dont have any money loaded in the gpt api account, and i dont intend to load any money there."

HIS STANDING PROTOCOL, recorded here because a previous session broke it: he says which track runs on which subscription; that is implemented ON THE WORKLOAD; the Carol Intelligence app then SHOWCASES what really runs so he can verify. A session that only updates the display makes his one instrument agree with the request while the estate still calls the old vendor - which destroys the check he was using. His words: "a pure waste of time and tokens.. and a bit annoying."

MEASURED BEFORE TOUCHING ANYTHING, so the work is scoped to what is actually wrong:

• TEXT IS ALREADY DONE. Every metered-OpenAI text call (gpt-5.6-sol, gpt-5-mini, gpt-5.4) stopped on 2026-08-01; today's traffic books provider=codex. Nothing to move.
• PICTURES AND VIDEO ARE ALREADY ON GEMINI, exactly as Ninad said. Proven live today: a picture booked gemini-2.5-flash-image against the media track. The tracks simply declare NO MODEL, which is why they read as empty - a gap in the RECORD, not in the workload.
• SPEECH IS THE ONE THING STILL LEAKING.

WHY IT LEAKS: the voice path takes its engine from a request parameter that DEFAULTS TO OPENAI. Gemini is available but only if the caller asks. Four call sites reach OpenAI speech: the live voice call, its own retry, Carol's WhatsApp voice notes, and Scriber's Logbook video narration.

AND THE GEMINI SIDE IS ALSO DEAD. Measured by calling the vendor: gemini-2.5-flash-native-audio-latest, the model the voice path names for Gemini, returns 404 NOT_FOUND. So "switch the default to Gemini" alone would have moved the workload from an unfunded vendor to a retired model - the same class of failure CAROL-INI-3663 cured this morning for four other capabilities. gemini-2.5-flash-preview-tts also works but appears under NO published SKU, and LAW 1 of the subscriptions runbook forbids enabling a model without rates - an unpriced model makes its track's cap decorative.

SCOPE. 1. Register the working Gemini speech model with all four rates READ FROM GOOGLE'S CATALOG, enabled only after a real call books a non-zero cost (runbook LAW 1: rates before enabling). 2. Retire the 404ing audio model - enabled=0, never deleted (LAW 2). 3. Move the WORKLOAD: all four speech call sites default to Gemini. OpenAI stays reachable but becomes a NAMED, LOGGED, non-default fallback - never a silent vendor swap, since the account is deliberately unfunded and a silent failover is what CAROL-INI-3663 just made illegal. 4. Declare the model on Bilbo's two tracks BY HAND, per Ninad's instruction - never inferred from what they have been running (assign-track-subscription LAW 1). 5. PROVE BY RUNNING: one real synthesis, and the ledger must book Gemini. The Carol Intelligence app is then READ, not edited - if it disagrees, the workload is wrong.

OUT OF SCOPE. The daily budget on any track (Midas's runbook).

WHY THIS IS FILED OVER A DUPLICATE VERDICT. The gate refused it as the same underlying work as open CAROL-INI-3667, whose proof was to be exactly this move. MEASURED at the moment of refiling, which is what rule 713 requires before a duplicate verdict may be set aside - plumbing never buries an UNMET objective:

• 3667 moved the track's OWNERSHIP (Athena -> Bilbo, service -> blogs) and built the runbooks. That half is real and is not being redone.
• The VENDOR did not move. The voice path still defaults to 'openai'; Scriber's narration and Carol's WhatsApp voice notes still call the OpenAI synthesiser; and the last OpenAI speech charge is from TODAY at 08:54. Those three files were last touched on 1 August, before 3667 reached review.
• 3667 is in REVIEWING, i.e. claiming completion on an objective the estate can be measured to have not met.

This is the same shape as the mistake Ninad named this session: the record moved and the workload did not. 3667's own work is left untouched and its runs are left alone.

⚖️Decisions

• Elrond's bypass methodology checklist (a reminder, not a gate -- you've got this): 0. File it requested_mode='bypass' (planner-vs-bypass is a deliberate choice). bypass_start REFUSES a non-bypass initiative (CAROL-INI-1846), and the dispatcher only skips the bypass lane when the mode says bypass -- a 'planner' mistag lets Merlin's pipeline grab the placeholder step and block your finished work. 1. Filed as planned status -- let the bypass claim/activate it; never file active. 2. Open the bypass (bypass_start) with your droid id + the remediation answer (remediates_initiative_id=NNN, or remediates_nothing=True). 3. Work the blocks for your work-type: template -> design -> code -> test -> review. Do the real work; record decisions on the initiative as you make them. 4. Reality is recorded for you at close -- code (files changed), each decision, and the twin-review verdict become real activities tied to this initiative and show in the Activity Tracker like a planner run (CAROL-INI-1840). No dummy rows. 5. Keep the initiative status moving; it parks in 'reviewing' and is tagged uat-pending for you at close (CAROL-INI-1836), so the stuck-watchdog leaves it alone until UAT. 6. Close runs the gates (design/architecture compliance + caller-audit). If a gate flags something pre-existing or unrelated to your change, waive it with a clear written rationale -- audit, don't skip. 7. Bypass skips the planner's auto-orchestration, NOT the standards. Same template checklist, same review, same observability as a planner run. (elrond)
• Current state at filing (Elrond validity check): Text, pictures, and video have already moved off the unfunded OpenAI account, but speech is still leaking: 33 calls to gpt-4o-mini-tts booked provider=openai today. The Speech, Image, and Video tracks declare no model, so the record does not yet prove what runs. Carol Intelligence is registered and running, but the filing requires it to read and confirm the ledger rather than merely display what was requested. (elrond)
• [status-router] planned -> executing | event=bypass_executing | bypass transition (or-bx-01)
• [delivery-check] 5 must-have criteria remain pending at bypass_end — delivery has no mechanical re-performance lane; UAT must grade on live evidence, not checklist silence (CAROL-INI-3020): (no detail) (orion)
• [status-router] executing -> reviewing | event=bypass_reviewing | bypass transition (or-bx-01)
• [status-router] reviewing -> executing | event=bypass_executing | bypass transition (or-bx-01)
• [delivery-check] 5 must-have criteria remain pending at bypass_end — delivery has no mechanical re-performance lane; UAT must grade on live evidence, not checklist silence (CAROL-INI-3020): (no detail) (orion)
• [status-router] executing -> reviewing | event=bypass_reviewing | bypass transition (or-bx-01)
• UAT evidence from full regression run 191: regression/test_ini3662.py now fails its deliberate ambiguity guard because 93 formerly ambiguous historical model labels have been renamed. CAROL-INI-3672 must either prove each rename from its voice/image source and update the superseded 3662 expectation, or restore the ambiguous labels; do not close while the estate test and the intended migration disagree. — This is reviewing-stage feedback caused by 3672 current data, not a new initiative and not part of the Admin change. (orion)
• [status-router] reviewing -> closed | event=operator_signoff | Auto-accepted (CAROL-INI-1859): Orion-initiated, >2 days in reviewing with no objection. (el-srac-01)

✅Success criteria

• A real synthesis books a Gemini model in the cost ledger, with provider=gemini, proven by running it. (must_have)
• No speech call site defaults to OpenAI any more, and the OpenAI path logs loudly when it is used. (must_have)
• The Gemini speech model carries all four rates derived from Googles published catalog before it is enabled. (must_have)
• The retired audio model is disabled, not deleted, and nothing can resolve to it. (must_have)
• Bilbos Speech and Image and Video tracks declare their model, and the Carol Intelligence app is READ to confirm it matches what the ledger shows - never edited to agree. (must_have)