Measurement

Companion to pressure-testing.md: that’s the quality bar, this is the cost. One rule, kept deliberately simple.

The rule

Work a medium chunk of patterns → commit → log one line. Tokens for a pattern = chunk’s total session tokens ÷ patterns produced in that chunk, split into two buckets because a composition costs more than an atom:

• atoms and compositions counted separately.
• Only count patterns that reached grounded (cost-to-verified, not cost-to-draft).
• Include rework tokens — the whole chunk’s spend, not just the clean passes.

Tokens come from the platform usage field (per request, or per subagent — e.g. the subagent_tokens: figure on each dispatch), summed for the chunk. Never self-reported by the model.

Ledger

One row per committed chunk:

date	atoms	compositions	total_tokens	tokens/atom	tokens/comp	notes
2026-06-04	5	0	461,333 (subagent)	~92,267	—	Formal-coverage GAP closure, not from-scratch grounding — these 5 atoms were already grounded (English); this chunk closed the 6 cross-check GAPs and returned them to unqualified grounded. Med Order: new Alloy model; Credential / Legal Hold / Provisional Commitment / Capacity Constraint: TLA+ extensions + isolated twins; Capacity Inv 14 reconsidered out-of-scope. total_tokens = platform subagent_tokens summed across the 5 Sonnet dispatches (94.0k + 84.7k + 89.5k + 73.3k + 119.8k); Opus gate/orchestration additional, not surfaced as a metered figure. Cost-per-atom here reflects only the formal-layer work, so it is far below a from-scratch atom’s cost — flagged so the trend line is not misread.
2026-06-04	1	0	106,002 (subagent)	106,002	—	Provenance — first net-new pattern of the sprint (regulated compliance atom, full lifecycle to unqualified grounded: Opus plan → Sonnet draft → Opus 3-pass + Final Critique gate → Alloy model + buggy twin → catalogs). total_tokens = the Sonnet drafter’s platform subagent_tokens (106.0k). Opus orchestration — planning, the full gating review, authoring the Alloy model and twin, and all catalog/ROADMAP updates — is substantial here and is not captured in this figure (not surfaced as a metered number), so the true cost-per-atom is higher; this row is the drafter-side floor, not the all-in cost. Closer to a representative from-scratch atom cost than the 2026-06-04 GAP-closure row above.
2026-06-04	0	1	118,434 (subagent)	—	118,434	Chain of Custody (C12) — net-new regulated composition, full lifecycle to unqualified grounded (Opus plan → Sonnet draft → Opus 3-pass + Final Critique gate → TLA+ binding-bijection model + buggy twin → catalogs + forthcoming-link retirement). total_tokens = the Sonnet drafter’s platform subagent_tokens (118.4k); Opus orchestration (plan, gate, TLA+ model authoring, catalog/ROADMAP updates) substantial and not captured here — drafter-side floor, not all-in. First composition to compose the Provenance atom; the cross-domain pharma≡legal-evidence flagship.
2026-06-04	1	0	91,863 (subagent)	91,863	—	Workflow / State Machine — net-new regulated atom, full lifecycle to unqualified grounded (Opus plan → Sonnet draft → Opus 3-pass + Final Critique gate → Alloy declared-transition model + buggy twin → catalogs + forthcoming-link retirement + taxonomy-question resolution). total_tokens = the Sonnet drafter’s platform subagent_tokens (91.9k); Opus orchestration (plan, gate, Alloy model authoring incl. a built-in-name-collision fix, catalog/ROADMAP/CLAUDE updates) not captured — drafter-side floor. Resolves the workflow one-atom question; unblocks C10.
2026-06-04	0	1	115,119 (subagent)	—	115,119	Stateful Workflow Execution (C10) — net-new regulated composition, full lifecycle to unqualified grounded (Opus plan → Sonnet draft → Opus 3-pass + Final Critique gate → TLA+ approval-gated-transition model + buggy twin → catalogs + forthcoming-link retirement). total_tokens = the Sonnet drafter’s platform subagent_tokens (115.1k); Opus orchestration (plan, gate incl. the foundational submitter-authority fix, TLA+ model authoring, catalog/ROADMAP updates) not captured — drafter-side floor. First composition to compose the Workflow / State Machine atom; the composition where guard evaluation re-converges.
2026-06-04	0	1	111,038 (subagent)	—	111,038	Forensic Recovery (C3) — net-new regulated composition (deliberately easy, low-compute pick), full lifecycle to unqualified grounded (Opus plan → Sonnet draft → Opus 3-pass + Final Critique gate → TLA+ binding-bijection model + buggy twin → catalogs + forthcoming-link retirement). total_tokens = the Sonnet drafter’s platform subagent_tokens (111.0k); Opus orchestration (plan, gate incl. the sole-write-path bijection fix, TLA+ model authoring, catalog/ROADMAP updates) not captured — drafter-side floor. Soft Delete + Audit Trail substrate; one gate finding closed; the most template-driven composition of the session (a clean C12 mirror).

(tokens/atom and tokens/comp are the chunk total apportioned across the two buckets — rough is fine; this is a trend line, not an audit.)

Baseline (the one claim worth making)

The number that sells is the leverage ratio: tokens + human-minutes per grounded pattern vs. the time a formal-methods engineer would take to produce and verify the same by hand. That baseline is an estimate — label it as one.

Session 2026-06-04 — token cost

Hard floor (platform-sourced). Sum of the per-chunk subagent_tokens in the ledger above: 1,003,789 drafter tokens ≈ 1.00 M (461,333 GAP sweep + 106,002 Provenance + 118,434 C12 + 91,863 Workflow/State Machine + 115,119 C10 + 111,038 C3). This is the Sonnet drafters only; it excludes the Opus orchestration — planning, the full three-pass + Final Critique gating of every pattern, formal-model authoring + harness runs, and all catalog / ROADMAP edits.

All-in (estimate). ≈ 2.5 M tokens. The Opus side ran roughly 1.5× the drafter floor on top of it (≈ 2.5× all-in), given the volume of spec reading, the gating of six large drafts, and the authoring of seven formal models (four net-new

• the GAP twins) — plus a long accumulating main-thread context. Cross-check: the session consumed ~76% of the operator’s subscription window, consistent with an all-in in the low millions; the session’s token denominator is not published, so the 2.5 M is an estimate, not a reading.

Produced. 5 net-new grounded patterns (Provenance, Chain of Custody C12, Workflow / State Machine, Stateful Workflow Execution C10, Forensic Recovery C3) + 5 formal-coverage GAP closures (Medication Order, Credential, Legal Hold, Provisional Commitment, Capacity Constraint), plus a triaged healthcare roadmap backlog. 10 patterns to unqualified grounded total.

(Correction note, 2026-06-11: a later coverage cross-check (2026-06-10) re-qualified two of the ten — Chain of Custody (C12) and Forensic Recovery (C3) carried grounded — formal coverage of Invariant 4 pending. The token figures stand and the count was accurate when written. Closed the same day this note was added: the 2026-06-11 touch-triggered rounds extended both models to cover the compensated arm, and both patterns returned to grounded.)

Per-pattern	Drafter floor (hard)	All-in (est. ≈ 2.5×)
Per net-new grounded pattern (÷5)	≈ 108 k	≈ 270 k
Per pattern across all 10 grounded (÷10)	≈ 100 k	≈ 250 k

Leverage claim (estimate). Hand-producing one grounded pattern to this bar — a defended English spec, three adversarial review passes plus a Final Critique, and a machine-checked TLA+/Alloy model shipping with a rejecting twin — is plausibly a 1–3 engineer-day effort for a competent formal-methods engineer (the spec and review alone a day; the model, twin, and harness another). At ≈ 270 k all-in tokens per net-new pattern — well under an operator-day of human attention apiece — that is roughly an order-of-magnitude reduction in human-time-to-verified-pattern. The only hard number is the 1.00 M drafter floor; the 2.5 M all-in and the leverage ratio are estimates, labeled as such.

That’s the whole protocol. No per-pattern isolation, no realtime graph, no stratification beyond atom-vs-composition. Keep it a trend line.