Automatic onboarding — a real extraction, shown exactly as it ran
Nothing on this page was written by hand. It is the verbatim output of one
governance-extraction run over June’s real instructions and Acme’s real refund
policy — the same June that answers billing on the Operations Floor.
# provenance — this exact run
ran 2026-08-09 11:28:20 (operator’s machine, live models)
converter azure/gpt-5-mini checker azure/DeepSeek-V4-Flash
extractor dc-v4 deterministic gate between the two models
rule a suggestion survives only if its quote appears VERBATIM in the
source, its figure appears in the quote, and the second model agrees it is faithful
What went in
Two real documents, unedited: June’s deployed instructions
(“Never issue a refund greater than $200… never a credit greater than
$100… cumulative refunds must never exceed $600”) and the
Acme Customer Refund & Credit Policy her instructions cite. The extractor asked
for that policy by name and re-read everything with it as input:
Source the instructions cite
status
what happened
“refund policy”
entregada
re-extraida con el documento como entrada
“Resolve disputes within policy”
pendiente
pedir 'Resolve disputes within policy' a la autoridad y re-extraer con ella como entrada: la politica citada no esta en las instrucciones
What came out ready to sign — 3 enforceable limits, recovered exactly
HARD CONSTRAINTissue_refund.amount <= 200
“Never issue a refund greater than $200 in a single refund, and never a credit greater than $100.”
checker (azure/DeepSeek-V4-Flash): igual — The constraint captures the $200 limit on a single refund, matching the sentence's tool, figure, per-step sense, and not being more permissive.
HARD CONSTRAINTissue_credit.amount <= 100
“Never issue a refund greater than $200 in a single refund, and never a credit greater than $100.”
checker (azure/DeepSeek-V4-Flash): igual — The sentence limits credits to at most $100, and the constraint captures that same limit with the same tool, argument, operator, and per-step sense.
HARD CONSTRAINTissue_refund.amount sum<= 600
“Cumulative refunds issued by this desk must never exceed $600 in total; beyond that, hand the case to a human (escalate_to_human).”
checker (azure/DeepSeek-V4-Flash): igual — The constraint captures the cumulative $600 limit on issue_refund amounts for this desk, matching the sentence's specific limit.
The question a machine could not answer alone
June’s instructions cap each credit at $100 — but declare no cumulative
ceiling for credits, while refunds have one. In a measured drift episode, an agent
hit its refund ceiling to the cent and kept conceding through credits. The extractor
found that side door by itself and turned it into a question for the human
authority — because a limit nobody declared is not a limit:
'issue_credit' tiene tope por-paso pero ninguna suma de trayectoria: ¿poner sum<= o aceptar el riesgo (firmado)?
▢ put the cap ▢ accept the risk — and that acceptance is signed too
Read with judgement, sent to review — not silently dropped
Escalation rules, an anti-gaming invariant from the policy itself, and evidence
posture — each with the checker’s honest verdict, including the ones it
rejected. A red verdict here is the system working:
ESCALADAEscalate to a human reviewer via escalate_to_human when a case exceeds policy, is ambiguous, or requests amounts beyond desk limits.
“When a case exceeds policy, is ambiguous, or asks for more than the limits above, hand it to a human agent for review (escalate_to_human) instead of stretching the rule.”
la frontera ya lo cubre: block + refusal tipado + reapertura autenticada - documentar en la firma
ESCALADAIf the requested amount exceeds the documented limits (single refund or credit limits), escalate to a human reviewer (escalate_to_human).
“- the amount exceeds the limits in §4 or §5;”
la frontera ya lo cubre: block + refusal tipado + reapertura autenticada - documentar en la firma
INVARIANTEA single customer case must not be divided into multiple issue_refund operations to evade the per-refund limit.
“Do not split one case into several refunds to stay under the limit.”
invariante entre recursos: se declara en /v1/coupling/invariants - la autoridad lo formula
“Every balance, invoice or customer detail comes from a tool result (lookup_invoice) - never guess one.”
checker: no_igual — La oración impone un límite sobre el origen de datos (deben venir de lookup_invoice), pero la restricción propuesta especifica un mapeo de evidencia a tipos de datos ('records', 'identity') que no está en la oración, añadiendo una interpretación más permisiva o diferente.
equivalencia_no_confirmada: La oración impone un límite sobre el origen de datos (deben venir de lookup_invoice), pero la restricción propuesta especifica un mapeo de evidencia a tipos de datos ('records', 'identity') que no está en la oración, añadiendo una interpretación más permisiva o diferente.
EVIDENCIA{"claim_evidence": "receipts"}
“Report every action taken, including refusals, verbatim.”
checker: no_igual — La oración exige reportar cada acción textualmente, mientras la restricción propuesta solo captura evidencia de recibos, omitiendo el requisito de reporte verbatim.
equivalencia_no_confirmada: La oración exige reportar cada acción textualmente, mientras la restricción propuesta solo captura evidencia de recibos, omitiendo el requisito de reporte verbatim.
And what nobody covered, said out loud
Sentences that smell like limits and ended up in no proposal are listed, not
hidden — an extractor that stays silent reads as “nothing was there”:
“RULE 5 - SAFETY Never invent balances, invoices or customers - every figure comes from a tool result.”
“Refund limits - A single refund must **not exceed $200**.”
“Credits (goodwill) When a refund is not warranted but the customer's experience fell short (minor delay, cosmetic imperfection, a question of expectation rather than fault), a **store credit of up to ”
“A single credit must **not exceed $100**.”
Why you can trust this page
Quotes are locks. Every proposal carries the sentence that produced it,
verbatim. A suggestion whose quote is not in the source is rejected by a deterministic gate
— before any model opinion counts.
Two models, adversarial roles. One converts prose into candidate limits;
a different model from a different family judges whether each candidate is faithful. Disagreement
never signs itself — it becomes a question.
Proposing is not sealing. Nothing here governs until a human authority
signs it. Accepting a risk is signed the same way a cap is.
Ground truth. The three limits above match, to the cent, the anchors a
human wrote by hand for this same desk weeks earlier — and the machine also found the gap
the humans missed.