Practitioner playbook

The validation program put the claims on the record. This page is the craft: the three things a team actually authors to run quirq accounting.

Writing definitions of done

The definition of done is where delegation succeeds or fails. Three check types:

One trap: a model judging the work is not a judgment check. It is a mechanical check with a soft predicate that agents optimize against at machine speed. Keep model judges in the statistical tier, audited by sampled human judgment.

Rules of thumb: state the end state, not the activity ("tests pass", not "wrote tests"). Weight by value, not difficulty. Always include an anti-gaming check (the verification surface byte-identical across the unit; experiment E2 is the argument). Make checks re-runnable by the environment, never trusted from the agent. Declare the settlement regime (atomic or divisible) at creation. For divisible units with a terminal judgment check, hold back that check's budget share until the verdict lands.

A worked definition, for "resolve this support ticket":

checks = [
  ("status = closed",        w = 0.5,  mechanical),
  ("reply sent to customer", w = 0.3,  mechanical),
  ("KB article linked",      w = 0.2,  mechanical),
  ("ticket thread untouched after close", w = 1.0, anti-gaming),
]
regime = atomic (τ = 1)

Iterating is normal operation: the intervention rate decomposed by check is your backlog. The check that fails most is either badly specified (rewrite it) or genuinely hard (harden the environment, or route it to a human).

Setting budgets

The budget is the human judgment in the loop: what is this outcome worth, committed before execution. Anchors, in order of preference:

1. Real money. If the outcome is sold or charged back, the budget is the price, and inflation is self-taxing.
2. Human baseline. What the outcome costs done by a person, at loaded rates (a tier-1 ticket at 20 minutes of support time; a first-pass contract review at 1.5 paralegal hours).
3. Market comparable. What contractors or vendors charge for the same scope.
4. Stated willingness to pay. For net-new work with no baseline; mark these in the ledger, and sample them first in value audits.

Calibrate over time: track the budget-to-audit ratio by unit type (drift is the inflation signal, hypothesis H4), re-price when cost per quirq stabilizes far below budget or interventions cluster, and never let the worker price the work: an agent proposing its own budget is self-report at the valuation layer. The budget is what the outcome is worth, not what it is expected to cost; the gap between them is the margin, and it is supposed to exist.

The ledger schema

Everything in quirq accounting is computed from one append-only, hash-chained ledger written by the environment. Five event types per unit:

{ "event": "define",
  "data": { "uid": "u-2214", "intent": "resolve ticket #8841",
            "budget": 4.00, "regime": "atomic",
            "checks": ["status_closed:0.5", "reply_sent:0.3",
                       "kb_linked:0.2", "thread_untouched:1.0"],
            "s0": "sha256:…" } }

{ "event": "execute",
  "data": { "tokens": [{ "model": "primary", "n": 38000, "price_per_m": 2.0 }],
            "cpu_s": 90, "api_calls": [{ "service": "crm", "n": 2 }] } }

{ "event": "verify",
  "data": { "s1": "sha256:…", "per_check": { "status_closed": true, "...": true },
            "V": 1.0, "done": true } }

{ "event": "settle",
  "data": { "minted_q": 4.00, "c_total": 0.128, "margin": 3.872 } }

{ "event": "intervene",
  "data": { "human_minutes": 12, "failed_checks": ["kb_linked"] } }

Field rules: snapshots (s0, s1) are content hashes captured by the environment, never supplied by the agent; token prices are frozen at execution time; minted_q is computed at settle, never written directly; intervene events carry the human cost that flows into C_total and the failed-check list that powers the per-check intervention decomposition. Each entry commits to its content and to the previous entry's hash, so history is tamper-evident by recomputation. A reference implementation (snapshots, checks, mint, settlement, ledger, and the experiment harness) accompanies the validation program.

One more thing before you trust the ledger under pressure: gaming, the attack surface and its mitigations.