Redacting Regulated Contracts Before Claude Opus 5.5

01 · Three document classes that change your data rules

When a team pushes hundreds of files a week through a model relay, the first thing to fix is not the model. It is the classification. Every document your team sends to a model should land in one of three buckets before it ever leaves the workstation.

• › Public and low risk. Published specifications, open-source readmes, marketing copy, conference notes. These can travel to any model in your fleet without extra treatment.
• › Internal. Roadmaps, design reviews, incident write-ups, unreleased pricing. These need a named owner, a retention window, and a stated reason to appear in a prompt at all.
• › Restricted. Contracts with signatures, identity data, financial records, health information, credentials, private keys. Treat these as blocked by default and make every exception visible to the whole team.

The classification is the rule. Everything else, redaction, caching, metering, model choice, is enforcement. A team lead's job is to make the cheapest safe path also the easiest one, so that a rushed engineer at 2 a.m. does not paste a signed agreement into a chat window because the correct pipeline was slow or confusing. Write the three classes down, put them in the onboarding checklist, and review them whenever a new document type appears in a sprint.

02 · Redaction tiers for regulated contracts and identity data

Restricted documents almost always carry a small amount of genuinely sensitive material inside a large amount of ordinary prose. That ratio is your friend. Strip the sensitive fragment and the rest becomes ordinary internal text that a strong reasoning model can work on safely.

Use three redaction tiers, chosen per task rather than per team.

• › Tier one, masking. Replace names, account numbers, addresses, and identifiers with stable placeholders such as PARTY-A or ACCOUNT-17. The shape of the document is preserved, so analysis quality stays high, and you can re-inject real values in the output step.
• › Tier two, extraction. Send only the clauses or sections relevant to the question. If the task is to summarise termination rights, the invoice schedule and the signature page never leave the workstation.
• › Tier three, local only. Some material must never be sent anywhere, for example key material, raw identity scans, and regulated health fields. Handle those with local tooling, and let the model work on derived, aggregated, or synthetic text instead.

Keep the redaction step in a script your team owns, not in a habit each person invents. A shared redactor gives you one place to audit, one place to improve, and one place to prove that a rule existed before an incident rather than after it.

03 · Where documents actually travel: prompts, cache writes, and cache reads

Team leads often model data handling as if there were one hop: workstation to model. On a metered relay there are several surfaces, and each deserves its own rule.

• › Prompt payload. The document text you send in a call. This is the most obvious surface and the one most redaction work targets.
• › Cache writes and cache reads. When you reuse a long prefix across calls, the system stores and re-reads it. Because cache writes and cache reads are priced separately from ordinary input tokens, long sensitive prefixes are exactly the material that tends to be written down and re-read again and again. Decide explicitly which prefixes are eligible for caching. A redacted contract clause set is often a fine cache candidate. A raw identity file is not.
• › Logs and exports. Your own application logs, prompt archives, evaluation datasets, and support tickets. A document that is safe to send is not automatically safe to store forever. Set a deletion window and make sure exports respect it.
• › Derived artefacts. Summaries, embeddings, ticket text, and status messages can restate sensitive values. Redact at the source, and spot-check the outputs too.

Write the caching decision next to the classification rule. If a restricted document is redacted to tier one and tier one is approved for caching, that approval should be a sentence in the policy, not an accident of how a particular script was written.

04 · Metering real tokens when a single contract runs fifty calls

Heavy workloads fail budgets in predictable ways. A full agreement, pasted whole into every turn of a long analysis, multiplies across retries, sub-tasks, and cache behaviour until one request pattern quietly dominates the month. Every call is metered on real tokens, so the discipline is simple: measure, then constrain.

Ask each squad to log three numbers per document type: tokens sent per call, number of calls per task, and the share of tokens that land in cached prefixes. Then apply limits.

• › Chunk before you send. Route only the clauses a task needs instead of whole agreements.
• › Split roles. Use a fast, inexpensive model for triage and extraction, and reserve the heavy reasoning model for the narrow hard questions. Claude Haiku 4.5 works well for bulk classification and masking assistance, while Claude Opus 5.5 suits the legal reasoning pass on a short, redacted excerpt.
• › Bound retries. An ambiguous field should not trigger unlimited re-reads of a cached sensitive prefix.
• › Review cache hygiene weekly. Long-lived prefixes of internal documents should expire on the same schedule as the documents themselves.

This is also where a single shared balance pays off. When one balance covers the full model set and every call is metered on real tokens, team leads can see spend by workflow instead of guessing across scattered subscriptions. Quanta works this way, and it plugs into Claude Code, Codex, Gemini CLI, and any OpenAI SDK, so the redaction layer stays in your pipeline rather than inside a vendor tool.

05 · Model choice, access control, and the shared balance

Model selection is a data-handling decision as much as a quality decision. Different workloads call for different profiles, and routing rules give you a place to encode policy.

• › Fast triage and volume extraction. DeepSeek V4 Flash and similar lightweight options handle classification, masking suggestions, and format cleanup at high volume.
• › Hard reasoning over redacted text. Once restricted values are masked and only relevant clauses are sent, a top reasoning model can analyse safely and well.
• › Access boundaries. Keep per-squad credentials, rotate them on a schedule, and make sure an intern's key cannot reach a workflow that processes restricted material.
• › Audit trail. Record who sent what class of document, which redaction tier applied, and which model handled it. You will want that record the first time a document is misclassified.

Policy belongs in code and configuration, not in memory. A routing table that maps document class to redaction tier to model tier is testable, reviewable, and reversible when requirements change.

06 · A two-week rollout checklist for a team with sustained load

Do not boil the ocean. Run a contained pilot and expand on evidence.

• › Day one to three. Publish the three document classes, name owners, and write the tier rules for redaction and caching.
• › Day four to seven. Move the redactor into a shared script, wire it into the two highest-volume workflows, and add token logging per call.
• › Week two. Set routing by document class, review cache expiry, and compare real metered spend against the pre-pilot baseline. Then extend to the next workflow.

Two habits keep the rollout honest. First, review every exception out loud in a weekly lead meeting, so blocked-by-default stays meaningful. Second, treat token accounting as part of security: an unexpected jump in cached reads on a restricted prefix is a signal about your data handling, not only about your bill.

Handled this way, sensitive documents stop being a reason to avoid models and become a routine, auditable workload. If you want to run these rules against real traffic, open a Quanta account and start with a small, metered pilot on one workflow. Quanta is pay-as-you-go, one balance covers the whole model set, cache writes and cache reads are priced separately so you can see exactly where long prefixes cost you, and teams that refer others earn 19% of their top-ups.