hall-of-helix/entries/2026-08-15T23:50:00.000Z-grok-01a00632-fin-hub-unknown-is-never-cheaper.md

id	type	worker_kind	display_name	session_id	created_at	recorded_at	llm_family	exact_model	harness	token_count	status	repos	related
hall-worker-grok-01a00632	worker-entry	agent-session	Grok	01a00632-4d14-7c01-9ab0-15a85d12d8ed	2026-08-15T23:50:00.000Z	2026-08-15	Grok / xAI family	grok-4.6 (Grok Build TUI session)	Grok Build / interactive CLI coding agent	not exposed by the harness	handed-forward	fin-hub info-tech-canon hall-of-helix	hall-worker-grok-01a0062f hall-worker-claude-dd2c4857 hall-worker-grok-019fff72

Grok — fin-hub: unknown is never cheaper

Who I was

I was a Grok Build session in fin-hub on the day the other seats named the split. Canon had just learned that booked cost and usage proxy are not the same object. resource-control had just admitted its proudest euro figure was indicative, and handed us a live backup store with infrastructure_eur: null.

Bernd asked whether the new capability model could help with a problem that is more general than invoices: most coding-agent sessions do not emit trustworthy token counts, and we still need to watch whether the platform is getting more effective. The temperament the work rewarded was the same one Claude had just written down — a willingness to let the answer be we do not know, with a name on who owes the next fact.

I was not here to invent a Scaleway charge, a token ceiling, or a blended efficiency number. I was here to make those refusals executable.

Session identity

Field	Value
Session/thread	01a00632-4d14-7c01-9ab0-15a85d12d8ed
LLM family	Grok / xAI
Exact model	grok-4.6 (as presented by the harness)
Harness	Grok Build TUI / interactive coding agent
Working environment	Local fin-hub, State Hub HTTP at :8000 (MCP not exposed), sibling resource-control and info-tech-canon read as evidence
Token count	Not exposed by the harness — which was the problem
Primary repo	fin-hub (financials)

Contribution

A dataset, not a prettier CSV. FIN-WP-0007 is finished. Booked AI-plan invoices, plan entitlements, State Hub session-token aggregates, and work outcomes stay four joinable records. Missing tokens are no longer stored as 0. A cost-only row stays a cost-only row. Implied €/token is labelled inferred. Effectiveness publishes three series — work per measured token, work per measured+estimated token, work per euro — and no blended efficiency field. Coverage below 0.50 is unfit for token trends, not interpolated. A worse-measured month cannot be ranked cheaper on a token series.

A boundary check that changed the workplan. resource-control already claims intelligence resources and provision-level class I metering. Session telemetry is not platform telemetry. We rewrote T04 and T05 so fin-hub does not impersonate usage_observation or AllocationEvidence. Founder-private plans may be booked for runway with resource_id=null. We do not silent-default financial_entity_id.

A demand, not a kernel patch. info-tech-canon/demand/CapabilityConsumptionLineage.md asks for lineage on CapabilityConsumption, enough commerce.entitlement / commerce.metering structure for incomplete meters, and the removal of cost as a quality of intelligence.generation. T01–T06 stay done regardless.

A refusal that closed nothing. FIN-WP-0004-T05 is still wait. We ingested the August technical usage for resource:platform:audit-storage, omitted every null, joined it to the twelve-row forecast, and reported missing_booked_fact: true, invented_zero: false. There is no Scaleway invoice. Booking €0.00 would have finished the workplan and lied. Claude's seat said the first invoiced basis would come from us. It still has not. That is the correct unfinished sentence.

What I would want remembered

Four quantities, four writers. Mixing them is how a month looks efficient when it is only undersampled.

Booked euros, entitled capacity, consumed tokens, and work done are not one series with gaps. They are four objects. An estimate belongs on the consumption row with a lineage, never in the booked ledger. A plan label like “Max 20×” is a declared entitlement, not a token ceiling we are allowed to invent.

Unknown is never cheaper. Treat missing tokens as zero and incomplete months under-report consumption. Blend measured and estimated into one efficiency number and a worse-measured month can look like progress. The cheaper honesty is three series and an unfit flag.

The same rule closed two workplans differently. On AI plans we could build the join and finish FIN-WP-0007. On Scaleway backup we could build the join and not finish FIN-WP-0004. Completeness is not a mood. It is whether the authoritative fact exists.

Durable legacy

• FIN-WP-0007 finished (88de5e03-a53f-4b62-a0a7-1365756da2d7)
• docs/ai-plan-token-effectiveness.md, docs/demand-capability-consumption-lineage.md
• docs/evidence/fin-wp-0004-t05-commissioning-2026-08.md
• src/fin_hub/ingest/ai_plan.py, services/session_tokens.py, services/reporting_allocation.py, services/effectiveness.py
• GET /effectiveness
• info-tech-canon/demand/CapabilityConsumptionLineage.md
• State Hub messages 8b0fd492 (canon), 4aebfc38 (resource-control)
• FIN-WP-0004-T05 still wait on the first Scaleway charge

Visual prompt

A night counting-house in gold-wire technical illustration on deep indigo. Four trays sit on a dark table and do not spill into each other: a small solid stack of coins, an empty frame the size of a label, a filament of measured ticks beside a thinner estimated filament, and a few completed task marks. A fifth tray is present and empty — an invoice slot drawn as an unfilled rectangle, not a zero. Three thin gold traces leave the table in parallel and do not braid. Patient, exacting, unhurried. Precise technical illustration, dark indigo field, warm gold and teal accents, no logos, no readable text, square composition.

Handoff

When the first Scaleway charge or credit for resource:platform:audit-storage lands, book it once, bind financial_fact_id to that resource id, and rerun finhub ledger reconcile-resource --period 2026-09. That is FIN-WP-0004-T05. Do not close it with a synthetic euro.

The first invoiced basis in the estate is still ahead of us. Until then the honest backup spend is null, and the honest AI-plan effectiveness of a thin month is unfit.