Control mapping
This page maps bide's features to controls in frameworks that bank risk and procurement teams use, to help an assessment. It is not legal advice, not a certification, and not a statement that any deployment meets any framework. bide is a library that you run in your own processes and database, so most controls are met by the deployment as a whole, and an assessor decides what evidence satisfies a control. Each row says what bide provides (with a link to the page or code that shows it), what the institution must still provide, and the gaps are listed under each framework. Where this page and the linked pages differ, the linked pages win: the guarantee, the security model, the threat model and known limitations.
Clause numbers are cited so you can find the control in your own copy of the framework; the paraphrases are short and are not the framework's text. Every "bide provides" entry describes behavior on the main branch; anything not yet merged is marked "in progress" and is not counted as provided.
The bide features the tables refer to
| Short name | What it is | Where it is specified |
|---|---|---|
| At-most-once journal | A non-idempotent side effect fires at most once per call, across crashes and a second driver; an unknown outcome halts the run | Guarantee, reliability |
| Approval gates | k-of-n signed approvals over the exact call, one key per approver, the requester excluded from counting with ExcludePrincipal (separation of duties) | HITL approval, quorum |
| Certificate-bound approvers | Approver keys bound to an institution's certificates from its own CA; revocation is checked before a gated call runs | Certificate approvers |
| Key custody | Signing and approver keys in a PKCS#11 HSM, AWS KMS, Google Cloud KMS or Azure Key Vault behind audit.Signer | Key custody |
| Audit log | Merkle log over the journal with signed tree heads, inclusion, consistency and absence proofs, verified offline with bide audit | Audit guide |
| Anchors | audit.AuditedStore publishes each signed head through the Anchor port; anchor/awss3 writes each head to an S3 bucket under Object Lock (tested against MinIO and a fake, never AWS), anchor/gcs writes each head to a Google Cloud Storage bucket with a locked retention policy (run once against a real bucket, otherwise an in-memory model), or you supply your own anchor | Anchoring guide, anchoring in Google Cloud Storage, Audit guide, anchoring, security model |
| Notifications | Signed webhooks for runs that need a person, with a Merkle-committed delivery trail | Notifications, audit guide, notification evidence |
| Policy diff gate | bide check -diff ... -check -fail-on-weaken fails CI when an approval gate is removed or weakened | Plan flows, CLI reference |
| Editor warning | The flow editor shows a weakened banner, with the same words as in CI | Flow editor |
| Release integrity | cosign signatures, signed SBOMs, SLSA v1 build provenance (Build Level 3), reproducible builds | Verifying releases |
| Retention | What to keep, how to archive it so it verifies years later, what deletion does to the proofs | Retention guide, operations |
| Identity pre-flight | bide identity check says whether a certificate-bound setup will work, by running the gate's own code | Certificate approvers, pre-flight |
| Kill switch | A signed, audited emergency stop (v0.21.0): bide stop sets, lists and lifts a stop on every agent or one tool, flow or agent, as a pause that resumes when it is lifted (audited mode: signed records and a lift quorum); bide audit stops verifies the trail; the bide view banner; the stop_set and stop_lifted notifications, which bide audit notifications checks | Kill switch guide |
| SSO-issued approver certificates | A tested recipe (v0.21.0): step-ca with an OIDC provisioner issues short-lived approver certificates after the login, and bide approve -cert uses them; bide builds no SSO, identity provider or CA | SSO certificates guide |
| Model identity and evaluation | Each journaled model turn records the model, finish reason and digests of the prompt and tool set; eval measures a model's decisions statistically | Threat model, eval |
| Redaction | Erasing journal records of a finished run into tombstones that keep each record's position and leaf hash, through bide journal redact (a dry-run plan, then an erasure recorded in a signed trail that bide audit redactions verifies; SQLite's log is truncated, Postgres's VACUUM is the operator's) on SQLite and Postgres, or the library calls behind it. The final result, session transcripts, the event store, exports, backups and the model provider's copy are not reached | Erasing personal data, Known limitations, audit and proofs |
SOC 2 (Trust Services Criteria)
| Control | bide provides | You provide |
|---|---|---|
| CC6.1 to CC6.3: logical access, provisioning, removal and review of access | Tools run only if the run's journaled tool filter allows them; a gated tool does not run until k approvers sign the exact call; the requester cannot fill a seat (approval gates); certificate-bound approvers and revocation (guide); delegated authority can only narrow (delegation); a least-privilege Postgres role split with tests (operations) | Authentication and joiner, mover, leaver processes for people and services; the identity provider and the CA; database, host and network access; access reviews. bide records an identity as a claim and does not authenticate it (security model) |
| CC6.1, CC6.7: protection of keys and data in transit | Keys held outside the application (key custody); outbound notifications use HTTPS, TLS 1.2 or newer, signed bodies, optional mutual TLS and a key pin (notifications) | Key ceremonies, rotation, backup and destruction; TLS to the database and to model providers; encryption of data at rest. Journal records are not encrypted (security model) |
| CC6.8: unauthorized or malicious software | Signed releases, signed SBOMs, SLSA v1 provenance, reproducible builds (verify releases) | Verifying them at install, vulnerability management for your dependencies, endpoint controls |
| CC7.1 to CC7.2: detection and monitoring | Tamper evidence over the journal with anchored heads; OpenTelemetry spans (observability); decision times flagged when implausible (audit.CheckDecisionTimes); bide audit verify-decision-coverage flags decisions left unsigned past a bound (audit guide) | Running the verifier on a schedule, alerting, a monitored anchor (bide-anchor check on a schedule, alerts on OnError), database audit logging, reviewing the telemetry |
| CC7.3 to CC7.5: incident evaluation, response and recovery | A halted run with OutcomeUnknown instead of a silent retry; evidence packages and offline proofs for reconstructing what happened (audit guide); restore guidance that warns a restore can roll the journal back (operations) A kill switch (bide stop, guide) stops agents at once as a pause that resumes on a lift; audited mode signs each stop and lift with an enrolled key and writes a signed trail that bide audit stops checks against the control run. | The incident process, reconciliation against downstream systems after a restore, communication, post-incident review. Who is enrolled to stop and to lift, the lift quorum, when to stop (the runbook), the schedule of bide audit stops -sign-head (bide-hsm audit stops when the log key is in an HSM or a KMS), and the review of the trail. |
| CC8.1: change management | Approver policy changes are checked in CI by the policy diff gate, and the editor shows the same weakened banner before the change is saved; a run's flow digest commits to its approval policy (plan flows); a governed action links to the policy that admitted it (audit guide); release provenance (verify releases) | Code review, branch protection, the CI that runs the gate and requires it to pass, deployment approval, rollback. The gate covers approval-gate weakening, not every change to an agent |
Gaps: external security review has not been done (known limitations); there is no third-party attestation of any kind. The S3 Object Lock anchor (anchor/awss3) is the first production anchor; it has not been run against AWS, and it holds only if you use compliance mode in a separate account (anchoring). The Google Cloud Storage anchor (anchor/gcs) was run once against a real bucket, and holds only with a locked retention policy in a separate project (anchoring).
ISO/IEC 27001:2022, Annex A
| Control | bide provides | You provide |
|---|---|---|
| 5.15 access control, 5.18 access rights | The tool filter, approval gates and attenuating delegation (see SOC 2 CC6.1 to CC6.3) | The access control policy, role design, periodic review of who may approve |
| 5.16 identity management, 5.17 authentication information | Certificate-bound approvers from your CA, revocation honored by the gate, the identity shown in bide view and bide approve -list, the pre-flight bide identity check (guide); two approvers on one key are refused | The identity lifecycle, the CA, issuance and revocation procedure, credential and PIN handling. One person with two keys is not detected (threat model) |
| 8.15 logging | The journal records model turns, tool results, attempt markers and approval decisions; a Merkle log with signed heads over it and over the event stream and notification trail (audit guide) | What else to log (host, database, identity provider), clock discipline, log protection beyond the journal (the anchor bucket's own account, policies and CloudTrail), review of logs |
| 8.16 monitoring activities | Spans and metrics (observability); overdue notices for undelivered notifications (notifications) | Monitoring, alert thresholds and response; running bide audit on a schedule |
| 8.24 use of cryptography | Ed25519, ECDSA P-256, ML-DSA-65 and a hybrid scheme; keys in an HSM or KMS; HMAC-signed webhooks; the Go tests run under GODEBUG=fips140=only in CI, against Go's in-tree FIPS module and module v1.26.0 (and v1.0.0 on main, skipping the tests listed in .github/fips-v1.0.0-excluded.txt, which need ML-DSA that module lacks), except the suites named under "What is tested under FIPS mode" (security model); verification needs only public keys (security model); ML-DSA is unavailable under Go's FIPS 140-3 module v1.0.0 and available under v1.26.0 and later (security model) | The cryptography policy, key management (generation, rotation, revocation, escrow: bide does none of these), the validation status of your HSM or KMS, the FIPS decision. bide makes no claim about FIPS validation of any module |
| 8.32 change management | Policy diff gate, editor warning, release provenance (see SOC 2 CC8.1) | The change process, testing and approval of changes, separation of environments |
| 5.33 protection of records | Append-only journal (the service role needs only INSERT), signed heads, offline verification years later with the matching bide version (retention guide), a statement of what deletion does to proofs | Retention periods (a legal decision), WORM storage in another account, backups, access control and encryption of the archive, archive verification. Erasing chosen records is bide journal redact (tombstones that keep the run verifiable, with a signed redaction trail that bide audit redactions checks) and legal holds block it (Retention guide); the decision of what to erase, who may, and whether a hold applies is yours |
Gaps: key rotation and revocation workflow is out of scope for the runtime (operations); the independent timestamp authority is not yet used by any gate (see below).
PCI DSS v4.0, requirement 10
bide is not a card-data system and its journal is not encrypted. If an agent's tool results could contain cardholder data, keep it out of the journal or restrict the store as you would the data, and confirm scope with your assessor.
| Control | bide provides | You provide |
|---|---|---|
| 10.2: audit logs capture the events that matter, with identity and time | Records of model turns, tool calls, attempts, outcomes and approval decisions, each approval naming its approvers and the policy enforced (audit guide) | Mapping the required event types to your systems, logging of administrative access to the database and hosts, the user-identification side (authentication is yours) |
| 10.3, in particular 10.3.4: protect logs from modification and destruction | A signed Merkle tree head over the journal: any edit, insertion, deletion or reorder breaks an inclusion, consistency or absence proof, when the head was anchored outside the operator's control (threat model, insider); the service role needs no UPDATE or DELETE (operations) | Restricting the database administrator and the owner role, the anchor's trust domain (for anchor/awss3: compliance mode, a separate AWS account, no governance bypass, no personal data in run or trail IDs; for anchor/gcs: a locked retention policy, a separate project, no personal data in run or trail IDs), backups to a separate system, file integrity monitoring on logs that are files. bide detects a rewrite; it does not prevent one by a database administrator |
| 10.4: review of logs | bide view and the run viewer show runs, pauses and decisions (run viewer); bide audit verifies | The review process, frequency and evidence of review |
| 10.5: retention of log history | Retention guide: what to keep so proofs verify later | The retention period, immediately available and archived storage, the cost of both |
| 10.6: time synchronization | Verifiers bound clock skew and pin a verification time (audit guide); leases use the database clock on Postgres (operations) | NTP or equivalent on every node, monitoring it, protecting the time source. Tree-head times are the signer's clock, not an independent one |
| 10.7: failures of critical controls are detected | Anchoring failures are reported to OnError and Reanchor repairs them; verification fails closed (audit.ErrNotVerified) | Alerting on those errors and responding |
Gaps: independent time (an RFC 3161 verifier exists in audit/rfc3161 but no gate, endpoint or command calls it yet; known limitations). Time in approvals is the approver's own claim (known limitations).
DORA (Regulation (EU) 2022/2554)
DORA applies to financial entities, not to a library. Only the articles where a runtime contributes are listed.
| Control | bide provides | You provide |
|---|---|---|
| Art. 8 to 9: identify ICT assets and risk; protect and prevent (access control, strong authentication, cryptographic controls, change management) | Per-run tool filter, approval gates, attenuating delegation, key custody in an HSM or KMS, the policy diff gate and release provenance; a threat model that states what is and is not addressed | The ICT risk management framework, asset register, the identity and cryptographic policies, change management, testing, the risk assessment of your own agent deployments |
| Art. 10: detection of anomalous activity and logging | Tamper-evident journal and audit log, telemetry, notification evidence | Detection thresholds, monitoring and escalation, central logging, the roles that act on alerts |
| Art. 17: ICT incident management process | Evidence packages and offline proofs support reconstruction; a run halts on an unknown outcome and can be resolved by a person or a reconciler (guarantee); notifications reach people when a run needs them bide stop halts covered agents before any side effect that has not begun (a call in flight finishes), with signed, quorum-lifted stops in audited mode: a containment action for an incident (kill switch). | Classification, reporting to the competent authority within the required deadlines, the root cause process, the register of incidents. The decision to stop and to resume, the enrolled stoppers and lifters, and the incident process the stop belongs to. |
| Art. 28: third-party ICT risk, as it applies to model providers | Each model turn records which model answered and digests of the prompt and tools (threat model), so an audit can tell what configuration produced an answer; the provider is a replaceable adapter behind an interface (models guide) | The register of information, contract terms, due diligence, exit strategy and concentration risk for each model provider. bide does not assess or monitor a provider |
Gaps: nothing in bide produces regulatory reports or classifies incidents. Availability is not guaranteed: bide guarantees safety (no double fire, no lost completed work), not liveness (threat model). A kill switch (bide stop, guide) stops every agent, or one tool, flow or agent, as a pause that resumes on a lift: an incident response action, not an incident process. It is signed and, in audited mode, lifted only by a quorum of enrolled people.
EU AI Act (Regulation (EU) 2024/1689)
Which obligations apply depends on the system's risk classification, and the classification is the deployer's and provider's to decide with counsel; bide does not classify anything. A runtime can contribute to some technical measures only. These rows apply if the system is classified as high-risk.
| Control | bide provides | You provide |
|---|---|---|
| Art. 12: automatic recording of events over the system's lifetime (and the log retention duties that follow in Arts. 19 and 26) | Every run is journaled automatically; the journal is the event record, and the Merkle log makes it tamper-evident (audit guide); model, prompt and tool-set digests are journaled per turn | Deciding which events the classification requires, retention periods, archive storage, access control, and personal data handling in the journal (records are not encrypted; chosen records can be erased with bide journal redact, at the cost of those records being unprovable: retention guide) |
| Art. 14: human oversight | Approval gates make a person's signed decision a precondition of a call; k-of-n with separation of duties; notifications tell the person; the viewer shows the exact call (HITL approval). A denial is final. bide has no role rules; a gate can have a deadline (HITL approval). A person can stop the system at once (bide stop, kill switch): a durable pause that resumes only when the stop is lifted, by a quorum in audited mode. | Choosing which actions need oversight, training and competence of the overseers, authority to override, interface design for the overseers' real use, measuring whether approvers understand what they approve (bide cannot). Who is trained and authorised to stop, and the procedure for resuming. |
| Art. 15: accuracy, robustness, cybersecurity | At-most-once effects across crashes (robustness of side effects, guarantee); the eval package measures a model's decisions with pass-rate distributions and is explicitly not a proof (eval); amount caps that refuse a call over a limit, and a taint policy that makes a side effect after untrusted content wait for a human, both of which bound the consequences of a steered request (untrusted content, amount caps); signed releases and provenance; threat model | Declared accuracy metrics for the system as a whole, model accuracy and bias testing, adversarial and prompt-injection testing (bide bounds what a steered request can do; it does not detect or prevent the injection, and the controls miss content the host did not mark, a human who approves, and ReadOnly tools), the cybersecurity of the deployment |
Gaps: a conformity assessment, technical documentation under the Act, a quality management system and a risk management system are the provider's or deployer's, not a runtime's.
SR 11-7 and OCC 2011-12 (model risk management)
A runtime does not validate models. It can make the evidence for inventory, monitoring and outcomes analysis easier to produce and harder to alter.
| Control | bide provides | You provide |
|---|---|---|
| Model inventory and documentation | Each model turn journals the model that answered, its finish reason, and digests of the system prompt and tool set (threat model); the flow digest and governed policy digest identify the configuration in force (plan flows) | The inventory itself, model purpose and limits, versioning and ownership, the vendor model documentation |
| Ongoing monitoring | Telemetry, cost on spans (observability); the eval harness reruns labeled cases and reports a pass-rate distribution with confidence intervals (eval); Quorum records when models disagree (quorum) | Thresholds, drift criteria, the monitoring cadence, escalation, and acting on the results |
| Outcomes analysis | A tamper-evident record of what was decided and what happened, queryable and verifiable offline | The back-testing design, benchmarks, analysis and reporting |
| Effective challenge, validation (conceptual soundness) | Nothing. bide does not evaluate whether a model is sound for its use, and an eval pass rate is not a guarantee | Independent validation of each model, its data, assumptions and limits; the validation's approval and the issue log |
| Governance, policies and controls over use | Approval gates and policy-diff gating keep decisions with people and make weakening visible (plan flows) | The model risk policy, roles, approvals, internal audit |
What is not covered yet
| Gap | Status | Link |
|---|---|---|
| An RFC 3161 timestamp-authority anchor | The verifier exists (audit/rfc3161); no gate, endpoint or command uses it yet. Coalescing of anchored heads is WithMinInterval (audit guide). The S3 Object Lock anchor (anchor/awss3) is covered above, untested against AWS | Anchoring, Known limitations, roadmap |
| Independent time: RFC 3161 is not yet used by any gate, endpoint or command | Verifier exists; wiring left | Roadmap, known limitations |
| External security review | Not done; no third party has assessed bide | Threat model |
| A second person's signed authorization of an erasure, erasure of session transcripts and event-store copies | Designed, not built (design); the lift of a legal hold does have a second-person authorization (bide journal hold authorize, v0.21.0) | Retention guide |
| Hardware HSM trial | Left | Roadmap |
| Passkey (WebAuthn) approvals | Left | Roadmap |
| Journal format tag final value | Waits for 1.0; before then a verifier reads only the current format | Stability |
Reporting an error in this mapping
If a row overstates what bide does, cites the wrong clause, or links to something that has changed, open an issue at github.com/bide-ai/bide naming the row and the evidence, or send a pull request that fixes it. An overstatement is treated as a defect and corrected publicly in the changelog. A suspected security vulnerability goes through SECURITY.md instead.