Gå til hovedinnhold

C-10: Redaction and clearance

Summary​

Redaction replaces sensitive spans (people, unreleased product names, internal roles, secrets) with protected placeholders before content is sent to a model or to an external tool, and restores the originals afterwards. The defining property is locality: the original value never leaves the machine. Detection runs offline by default, the replacement carries only a coarse category, and the original-to-placeholder mapping lives in a local vault that is never serialized into the content handed to a tool, a prompt, or an exchange file.

The capability is framework-native, a peer of pseudo-translation and search-and-replace: it lives under core/redaction/, as two pipeline tools (redact / unredact), a built-in secure-translate flow, and integration with extract, merge and ingest.

Context​

Content is increasingly translated with cloud models and external tools, but some source must not be disclosed to either: pre-announcement product names, named individuals, confidential roles. The requirement is two-sided. The sensitive text must be absent from anything that leaves the machine, and the finished translation must read naturally with the originals back in place.

A find-and-replace cannot meet this. It either leaks, because the mapping travels with the document, or it loses the ability to restore, because there is no record of what was replaced. The content model already has the right primitive, the placeholder run, whose semantics name redaction as a use case, and the streaming pipeline already preserves inline codes across translation.

Decision​

The placeholder model​

A redacted span becomes a model.PlaceholderRun with a Type of the form redaction:<category>, mirroring the entity: prefix convention of the semantic vocabulary (F-02). The run carries a stable id (the token), a visible stand-in string (default template [REDACTED:{category}]), and constraints marking it non-deletable and non-cloneable. Categories are free-form strings; a recommended set is surfaced in the defaults.

The original text is not present on the run in any field.

The locality guarantee​

The original lives only in a vault, never in the content. Three backings, one per lifetime:

BackingWhereFor
An in-process secret annotation on the blockmemorysingle-run flows
A per-batch JSON sidecar.kapi/vault/batches/<batch-id>.jsonthe extract → external tool → merge round trip, a project's redacted .kpz packages among them
A package sidecar<package>.kpz.vault, beside the packagea redacted .kpz extracted from source files outside a project
The project vault.kapi/vault/redaction.jsoncontinuous ingest, which has no batch id to look under

The in-process annotation is keyed under a name no format writer serializes, and unredact deletes it after restoring, so it cannot reach an output file. The sidecars and the vault are written with restrictive permissions.

A sidecar or the project vault keys each original by document, block ID and token (redaction.RedactedValue.Doc, the project-relative path). Block IDs repeat across files and the token counter starts again in every block, so a key without the document lets one file's original overwrite another's. Content that sits in no file (an edit record, a note) names itself in its block ID and records no document. Flush writes version 2 (fileVaultVersion), which records the document and, for a vault a redacted package names, the vault's ID (redaction.FileVault.ID, optional), and redaction.OpenFileVault reads from version 2 (oldestVaultVersion) up to the version it writes; the floor stays put when the written version moves. A file with no version, or below the floor, was written by an earlier kapi, and it is refused with an *OlderVaultError (ErrOlderVault) that names the version found; one above the written version was written by a newer kapi, and it is refused with ErrNewerVault. Either way merge fails for that file, and a write to the workspace home and a read of a kept edition fail. The caller names the command that writes a new vault (redaction.WithRun): kapi extract for a batch sidecar or a package sidecar, whose entries belong to the files that extraction wrote, and, for the project vault, a run after moving the file aside, since the old file holds the only copy of every recorded edit's originals. The vault is the only copy of the originals, never a cache, so kapi leaves a refused file as it is.

vault/ sits beside work/, outside it. work/ is defined by being disposable: losing it costs CPU. The vault is defined by an exclusion: a named destination must never read it, and losing it means redacted content can never be restored. So rm -rf .kapi/work stays free, and kapi never deletes the vault on its own initiative. core/project.Layout.PrepareVault writes the directory owner-only and adds it to .kapi/.gitignore when that file ignores less than the whole of .kapi/, before anything is withheld into it.

For model calls the guarantee is structural, not advisory: a block with inline codes is rendered through the placeholder protocol, which presents each placeholder as an opaque token (the model sees neither the original nor even the visible label), and the return is matched back to the source run by id, so the placeholder survives the round trip with full fidelity (M-05).

Redacted packages​

A .kpz carries each source file as a skeleton (whose entries hold the text of every block, rendered and as written), with the raw bytes under --with-source, and a skeleton has no placeholder form. So a package written under a policy carries a different file: host.redactSourceForKpz runs the redact tool over the source in external mode, a step that draws each redaction placeholder as its visible token (redaction.ShowTokens), and the format writer, which writes the token as text whatever the format. The skeleton, the raw source and the content-memory lookups all read that file. The policy's rules, compiled into a detector, read each memory match and each terms concept as JSON, and one they find a value in is left out of the package, whichever file the value comes from.

The package works on tokens from there on, so merge restores by visible token within the block, in every plural form and select case (redaction.RestoreTree), as it does for PO, keyed by the file's name in the package and the block ID. The block IDs of the redacted file are those of the original: extract reads both and refuses the file when they differ, since merge matches by block. The same detector then reads the redacted bytes and the text of every redacted block, and extract refuses the file when it finds a value: one in the markup or in a field the reader does not extract, which the redact step never sees, or one across an inline code, where redaction.Redact skips a match. A policy that detects entities is refused before anything is read, since extract runs no entity-extract step to find them.

The manifest records redaction.vaultId (kpz.Redaction), and the vault file records the same ID. The vault stays where the route keeps its originals: the project's batch vault, named by its ID, for kapi extract --format kpz, and <package>.vault beside the package for an extract from source files, which refuses to replace a vault already at that path. A workspace cache carries the record in its metadata, with the package's kind, interchange task and each file's project path, and a pack writes them back, so a package that a receiver ran a tool over and packed still merges as what it is. kapi merge refuses a redacted package before it writes anything when the vault is absent or its ID differs (host.packageVault); a vault that records no ID is refused the same way. Merge then restores through the batch-vault path for an interchange package (restoreRedactedBlocks) and an unredact step in external mode for a workspace. A translation that still holds a placeholder prefix after the restore stops the merge with the file and the block: a translator changed the placeholder. The workspace route merges once into the cache to find one before it writes any file.

The interchange task's revisions in the manifest are unsalted hashes of the original runs, as they are in XLIFF 2 and PO, because merge compares them with the source the project holds.

kapi exec never reads the vault: a tool over a redacted package reads and writes tokens, and the originals reach a file only at merge. The refusal of a tool that sends blocks out over a package in a project with a collection's or an item's policy does not apply to a package whose vault is among the project's batch vaults (host.packageVaultHere): kapi extract --format kpz wrote it in this project and withheld each file's values under that file's policy. A vault beside a package proves nothing about which rules withheld its values, so any other redacted package is refused like any other package.

Detection​

Detection produces spans, byte offsets plus a category, consumed by redaction.Redact. Two backends:

  • Rules (default, fully offline): literal terms and regular expressions from a dedicated rules file. Deterministic, and the only backend that preserves the locality guarantee without qualification.

  • Entities (opt-in): redact reads the entity annotations already on the block, produced upstream by entity-extract, and redacts the configured categories. The detection model is the caller's choice: a local model keeps everything on the machine, a cloud model trades that for coverage during the detection step only.

    The categories are the option surface a user picks (redact people, redact dates) through redact's entity-type list, with aliases and the model's entity: prefix normalized and validated. Naming any category enables entity detection, so the user does not also have to list the detector. Dates, times, currencies and measurements are excluded from the defaults, because they usually need locale formatting rather than hiding, but are available.

    Conditional requirement, not a new schema language. Two distinct requirements are in play and neither needs a condition DSL. The resource requirement (named-entity recognition implies a credential) lives statically on entity-extract; redact calls no provider and declares no requirement of its own, so enabling a category adds none to redact. You add the extraction tool to the flow, which is composition. The input requirement (redact needs an entity overlay when entity detection is on) is a config-derived IO contract: a contract resolver flips redact's entity consumed port from optional to required when its config enables entities, so a flow that redacts entities with no upstream producer fails data-flow validation instead of silently leaving the content unredacted and leaking it downstream. With rule-based detection only, redact reads no upstream port and the contract is unchanged.

Redaction is a structured edit​

redact is a transformer (E-03): it produces an edit plan (the span-to-replacement edits plus the originals to vault), and the framework applier rewrites the source. Because the plan is a known span-to-replacement map, the applier rebases the surviving run-anchored source overlays onto the redacted runs in one pass: a term annotation from an upstream tool follows the rewrite and still reaches downstream steps, while a span overlapping a redacted span, including the consumed entity spans, is dropped. The applier vaults the originals as it replaces, so source rewrite and secret capture are atomic.

Restoration​

unredact and kapi merge restore through two complementary paths, because formats differ in whether they preserve inline structure on write:

  • By placeholder id, for structure-preserving carriers: in-process pipelines, and XLIFF, where the placeholder is a real inline element.
  • By visible token text, for carriers that flatten the placeholder to its string on write. The visible token is made unique within a block so the match is unambiguous.

Together they cover every supported carrier.

The policy and its readers​

The recipe declares redaction under defaults.redaction, on a collection (named or bare) and on a content item, each pointing at a separate rules file so the sensitive term list stays out of the committed recipe:

defaults:
  redaction:
    enabled: true
    rules: redaction.yaml         # binds any path; keep it out of version control
    detectors: [rules]            # opt in: entities
    placeholder: "[REDACTED:{category}]"
collections:
  - name: legal
    redaction:                    # replaces defaults.redaction for these files
      enabled: true
      rules: legal-redaction.yaml
    content:
      - path: "legal/*.md"

A file's policy is the most specific declaration that covers it: the item that claims the file (the claim ItemForPath makes), else that item's collection, else the defaults (KapiProject.RedactionForPath, which host.RedactionFor wraps). The declaration replaces the others whole. Rule files, detectors and placeholders are not merged across levels, so what a collection declares is the complete policy for its files, and enabled: false takes them out of redaction. One check (RedactionSpec.validateAt) holds every level to the same shape: a declaration that sets rules, detectors or placeholder without enabled: true fails the recipe load. Content in no file (an MCP call's text) falls under the defaults.

A policy with one reader is a policy that fails open everywhere else, so the policy is read at every point content can leave, and each reader asks for the policy of the file it holds:

  1. kapi extract resolves it per source, before it reads any, and applies it, emitting a redacted bilingual file and writing the batch sidecar. The --redact and --redact-rules flags layer over the policy each source resolves. Redaction runs before memory pre-fill, so the content memory is queried with redacted text and pre-fills targets from it, and no sensitive value reaches the emitted file by way of a memory match. On kapi merge the incoming source is always restored, so per-block staleness compares original against original; the target is restored unless --no-restore is set.

  2. Ingest, which is the push to a venue (host/venue/source). host.RedactionFor resolves each file's policy and host.RedactAtIngest applies it as the file is read in, vaulting into the project vault. Ingest is continuous rather than batched, which is why its vault is project-scoped: a later restore has no batch id to look under.

  3. The flow placement gate. A project that declares redaction rejects any flow that sends source to a remote destination without a redact step (FlowDefinition.CheckRedactionCoverage, enforced in flow placement). This is a gate rather than an implicit rewrite: silently inserting a step into a user's flow would change what their flow means, while refusing to run it says so. A project flow's tools are built once for every file of a pass, so the gate holds it when any declaration turns redaction on (KapiProject.DeclaresRedaction); a built-in flow built for one input file is held when a policy covers that file. A project flow's input set splits by the policy that governs each file (groupInputsByBinding), and a redact step that names no rulesPath or rules of its own redacts each part under that part's policy (App.applyFlowRedaction); a step that names its own rules keeps them. kapi run, kapi up and its plan reach the step through the same assembly, and the MCP run_flow tool and the desktop's flow preview build their chains under the policy of the one file they run over. The step detects with its own detectors and the policy's together, and a step that detects entities with no entity-extract step ahead of it is refused. A step that sends source out after unredact, with no redact step between them, fails placement (egress-after-restore).

  4. A single-tool run. kapi exec <tool>, and the MCP tools that run one registry tool over a caller's text, compose no flow of the user's, so the run itself places the steps (host/toolrun_redact.go). When the tool's contract, refined for its config, carries an API call, redact runs ahead of it with the rules of the policy that covers the file, settled per file before any file is read, and unredact after it. A .kpz package is one chain over every file it holds, so under a collection's or an item's policy such a run is refused. The originals ride on the block in between as they do in secure-translate. The tools a run reaches read the whole block, so the step is configured for it (WholeBlock): it redacts blocks marked not translatable, and each target locale's runs under tokens of their own. Detection reads a run sequence's own text runs, the coordinates the carve uses, and every plural form and select case is redacted the same way under a prefix of its own. Unredact puts back an edition whose restored content equals what redact read exactly as it was read, and re-stamps a translation's status and origin, so the two steps around a tool that left a translation alone change nothing about it. A collector that sends what it gathered to a model (voice-infer's draft) reads the blocks between the two steps. A policy that detects entities is refused, because entity detection reads annotations an extraction step leaves and a run of one tool has none. A tool with no declared contract (no registry entry, or a schema with no toolMeta) is refused before it reads a block, because nothing says whether it calls a provider: over a file a policy covers, and over a .kpz package in a project that declares any policy. A tool that calls no provider runs alone. The chain each run builds is held to CheckRedactionCoverage and CheckPlacement, so a run without the redact step fails the gate a flow fails.

    external-command hands each block to a program kapi runs, which may send it anywhere, so its contract declares the process-exec side effect and the policy covers it as it covers a provider call: a run places redact and unredact around it, a run over a .kpz package is refused under a collection's or an item's policy, and a flow that runs it needs a redact step ahead of it. The program reads each placeholder as its visible token (redaction.VisibleText). The tool compares the tokens of its input with those of the program's output (redaction.CompareTokens); a token the output drops, alters, repeats or adds leaves the block as it was and records a failing external-command.redaction-placeholder finding, which kapi exec reports with exit code 3 and a flow run reports beside its other findings. kapi up collects the same finding over its passes, prints it under the run's summary in place of the up-to-date line, and exits 3.

  5. The deeper look. It sends chunks of the project's block texts to a model (host/contextdeeper.go). A chunk holds blocks of files under one policy only, and each chunk is redacted under that policy before the model reads it. Each chunk then carries that policy and the values read back from the vault, so the rest of its requests is redacted as below.

  6. The operation log. The record of an applied edit travels to every context backend the project shares its context through, so the recorder redacts the runs and the note it keeps of each document under that document's policy, vaulting each run sequence under the edition and revision it belongs to, and leaves the change set as sent out of every record of a project that declares a policy (C-03). A document the recorder knows by key alone is looked up in the document index. One whose file it still cannot name falls under every policy the recipe declares: under one, it is redacted with it; under several, its record keeps revisions and hashes only and the workspace home refuses to keep it, because successive passes of the redact tool number their tokens alike and would overwrite each other's originals in the vault. A translation with no file is kept in the log itself (the workspace home), so its record is redacted the same way and the workspace home restores the originals from the project vault where it reads the edition: on the machine that withheld them the edition reads whole, and anywhere else it reads with its placeholders. A policy that detects entities needs a read's entity annotations, which a record does not carry, so under one a record keeps revisions and hashes only and the workspace home keeps no edition.

Everything a request carries​

A request to a model carries more than the block, and the policy covers all of it. The project strings a request can carry, by where the prompt takes them from:

StringSourcePrompts
The block, its translations, its plural formsthe blockevery prompt
The neighbouring blocksthe documenttranslate
The block's keythe block's nametranslate
The previous source and approved translationthe content memory (prompt.Context.Prior)translate
Each term and its replacement, the do-not-translate termsthe term rules (profile.TermRuleMap)translate, review
Guidance, style, patterns, examplesthe voice profile (profile.RenderVoiceGuide)translate, review, voice-check
The instruction--instruction, the recipe, a fix pass's findingstranslate, segment
The domaina tool presetterm-extract, voice-infer
The known termsthe terms storeentity-extract
The neighbouring linesthe documentmedia-refine

The redact step leaves its policy (redaction.Policy: the rule detector and the placeholder options) on every block it reads in-process, in the block's secret annotation, whether or not the block holds a value to withhold. A tool's view of a block puts a redaction.Egress over that entry on the view's context (tool.ContextBinder). A tool that sends several blocks in one request binds them together: the translate batch (redaction.BindBlocks), and the voice-infer corpus and the media-refine neighbours (redaction.BindCorpus). Each exported AI tool constructor and prompt function in core/ai/tools wraps the provider it is given in aiprovider.Redacting, whoever built the provider; the wrapper reports the provider's limits and does nothing for a context with no Egress. It reads the Egress from the call's context and redacts each text part and section of a chat, and every field of a translate request, before the provider sends it. The recorder sits inside the wrapper, so --explain-prompts shows the request as it was sent.

The Egress matches the policy's rules, and every original the request's blocks already withhold, the latter by literal text, so a value an entity step found is redacted beside the block too. A JSON payload in a prompt (the translate batch) holds each string escaped, so the Egress also reads the unescaped text and replaces each match at its escaped span, with the token escaped the same way. Entity detection itself reads the blocks only: a value named only beside the block is redacted when a rule names it or a block of the request withholds the same text. Tokens are shared within a request:

  • A value a block withholds is drawn as the token the request already shows for it: its placeholder tag when the prompt carries the block's runs as tags (translate), its visible token ([REDACTED:Person]) otherwise. In a batch, the value is also kept in each other block's entry, so it restores in whichever block's answer the model copies it into.
  • A value met only beside the block takes the next token in the block's sequence, under a visible token that no other value of the request uses and that the request's text does not already hold, and is kept in the block's entry. unredact restores it from there by its visible token, so a token that matched text the person wrote would replace that text.
  • A span inside a token the request already shows is left alone.

Only the request that produced a translation fills the entry its tokens restore from, so the translate tool neither serves a block that carries an entry from the session's target cache nor keeps its translation there. A later redact step keeps the values an earlier one left in the entry and takes the entry's policy over.

A prompt that carries text alone (review, quality check, term-extract, voice-check, voice-infer, entity-extract, the media-refine neighbours, the deeper look) draws each placeholder as its visible token (redaction.VisibleText), so the model reads a token where the value was. entity-extract reports spans in the text it sent, and redaction.VisibleSource maps each back to the runs; a span on a token names a withheld value and is dropped. The LLM segmenter reads the block as every segmentation engine does, with each inline code masked, because its chunks must align with the runs.

The entry grows in memory only, by at most one value per distinct original the rest of the request carries and the block does not hold, and unredact discards it with the entry. On samples/northsea, a policy over the product and company names adds 250 values to the 23 the redact step recorded, across 136 blocks, at most 2 in one block: the voice profile names both, and every request carries it. The deeper look reads the project vault and writes nothing back to it.

Two commands ask a model outside a run of a registry tool. kapi voice check --ai runs its tool between redact and unredact (host.VoiceCheckTools) under the policy of the file it reads, or defaults.redaction for text given inline. kapi terms expand holds no block, so it redacts each request under the rules of defaults.redaction (host.TermsExpandRedaction); a project whose only policies sit on collections or items gives it none to apply. A withheld term goes out as its token, and the forms the model returns for a token name no term that was asked about, so none is written. The AI actions on Kapi Desktop's review (Fix with AI, Retranslate, Explain, the AI pre-review) run their tool through host.RedactionChain, under the policy of the unit's file.

A machine-translation engine reads a redacted block through semantic HTML: each withheld value is a placeholder element (<span data-type="redaction:person"/>) that maps back to its placeholder by position, and unredact restores it by id, as the translate tool's placeholder tag does. The engine never reads the visible token as text: an engine translates text, so [REDACTED:Person] could come back reworded and match nothing, where it keeps markup in place.

Not covered:

PathWhat reaches the provider
Media payloads (media-refine's image crop, speech clip, video frame)The media as recognised; the rules match text, and a crop or clip holds none
Machine-translation plugin engines (translate with an MT provider)The redacted block, each withheld value a placeholder element; no term rules, voice profile or other project text. The request goes through the plugin's provider, not aiprovider.Redacting, so anything the plugin sends beyond it is outside the policy

CLI surface​

  • kapi run secure-translate -i <file> --target-lang <l>: the in-process flow reader → redact → translate → unredact → writer.
  • kapi run redact-pii -i <file>: the built-in entity flow, entity-extract then redact configured for the common categories. The placement pass keeps redact ahead of any remote-egress step.
  • kapi extract --redact (or --redact-rules <path>): emits a redacted bilingual file and writes the batch sidecar.
  • kapi merge: restores originals from the batch sidecar after applying the returned target.

Not yet built: clearance as a destination ladder​

Redaction is binary at each reader: a span is either withheld or it is not. The design it points at is clearance, governance at a point over a ladder of destinations, where the local working tree sees originals and each further destination needs explicit clearance to. Nothing in the tree expresses a clearance level yet; the exclusion the vault encodes lives in its placement and in this decision rather than in a policy anyone can write down.

Rationale​

Why a placeholder run, not text substitution? Inline codes are already protected from translation, survive the streaming pipeline, and round-trip through bilingual formats. Reusing them means a model and an external tool treat a redaction exactly as they treat any other do-not-touch token.

Why is the original never on the run? So the guarantee is auditable: any serialized artifact can be scanned for the secret and must not contain it. The run carries only a category and a token.

Why dual restoration? Id-based restore is exact but needs the inline structure to survive. Plain-text carriers drop it, so a vault-backed text match on a per-block-unique token is the fallback.

Why rules by default and models opt-in? Rule-based detection is deterministic and fully offline; it cannot itself leak. Model-based detection is more capable but, with a cloud model, discloses source during detection; making it opt-in keeps the default trustworthy.

Why a separate rules file? The term list is itself sensitive. Keeping it out of the committed recipe lets it stay untracked while the recipe still records that redaction is enabled.

See also​