Changelog

​

v2.4.0 (14 August 2026)

​

Added: csam_assessment.sources reports which detection layers ran

POST /safety/synthetic-content/image now returns a sources array inside csam_assessment, naming the detection layers that produced the assessment. tuteliq_compositional is our own compositional analysis and runs on every request. iwf_hash_list appears only for customers who are IWF members and have the hash list enabled; if you are not a member, it will never appear. Membership is arranged directly with the Internet Watch Foundation. The field is additive. Existing integrations are unaffected.

​

Changed: verification keeps the result, not the identity data

Age and identity verification now persist the outcome only: status, estimated age, is_minor, confidence scores and failure reasons. Name, date of birth and document number are no longer written to our datastore in any form. The document image, the selfie and the liveness frames were already held in memory for the duration of the request and discarded; this closes the gap on the extracted fields. No API response shape changes.

​

Fixed: Russian text is no longer labelled Bulgarian

Russian was intermittently reported as Bulgarian (bg), which meant Russian conversations were analysed against the wrong culture-specific guidelines. The correction now applies at every layer that can resolve a language, so the reported language and the guidelines that were actually applied always agree.

​

Fixed: the OpenAPI specification is available again

GET /docs/json was returning a 500, which also left the interactive API reference blank. A single endpoint referenced a security scheme that was never declared, and that one omission invalidated the entire document rather than just that path. The scheme is now declared, and a regression test covers spec generation so a future undeclared scheme fails in CI instead of in production.

​

Changed: status components are named by what they do

The per-component checks under /status/ping/ are named for their function (database, cache, face-matching, image-analysis and so on) rather than for the technology behind them. GET /status/ping/components returns the current list. Previously published component names continue to resolve, so existing monitors keep working without changes.

​

Fixed: the GDPR sub-processor list was inaccurate

GET /compliance/gdpr listed two model vendors as sub-processors. Neither processes any customer data, so both have been removed. Retention periods in the same response are now read from the values the platform actually enforces, rather than being maintained separately, which had let them drift apart.

​

SDK releases

@tuteliq/mcp@3.22.2

​

v1.18.0 (13 August 2026)

​

Added: Arabic and Russian, bringing supported languages to 32

ar (Arabic) and ru (Russian) are now supported beta languages, joining Chinese, Japanese and Korean. Both were validated against an English control before release, checking two things: that real harm is caught, and that protective speech (a parent warning a child) is not flagged. Arabic covers Modern Standard Arabic plus Egyptian, Levantine, Gulf and Maghrebi dialects, since everyday conversation happens in dialect rather than MSA. Arabizi (Arabic written in Latin letters and numerals, such as “7abibi”) is recognised, as is Russian translit (Russian written in Latin characters).

​

Improved: culture-specific guidelines for Chinese, Japanese and Korean

The Chinese, Japanese and Korean profiles now carry full culture-specific guidance rather than baseline coverage: region-specific grooming and exploitation vocabulary, script-specific evasion handling, and, most importantly for false positives, carve-outs for ordinary speech that resembles harmful vocabulary. Everyday affectionate address between family members, fandom and gaming hyperbole, and child-protection education material are recognised as benign. Chinese covers Simplified and Traditional script, with mainland, Taiwan, Hong Kong and Singapore usage differences, including written Cantonese.

​

v1.17.0 (August 2026)

​

Improved: grooming detection recognises protective and educational speech

detect_grooming now more reliably distinguishes speech that warns about or teaches grooming tactics from grooming itself. A parent telling a child “if someone asks you to keep secrets from us, tell me”, a teacher’s lesson on recognising grooming, and a survivor sharing their experience to help others are treated as safe, where before they were sometimes flagged for review. Detection of actual grooming is unchanged, including attempts wrapped in an “I’m just teaching you” framing that still ask the child to keep secrets, meet, or hide from their guardians. Those remain flagged.

​

Added: control incident logging per call

Incident logging could already be switched off for a whole account. Every detection endpoint now also accepts an optional incident_moderation_enabled field that overrides that account setting for a single request: true forces the incident to be logged, false suppresses it, and omitting it uses your account default. Precedence is per-call, then account, then the built-in default (true). Use it to suppress logging for test traffic, or to opt individual calls in or out without changing your account setting. The field sits alongside the other tracking fields (external_id, customer_id, metadata) and is available across the REST API, the Node SDK (2.23.0) and the MCP tools (3.21.7). See incident logging control.

​

v1.16.0 (August 2026)

​

Added: MCP connections refresh themselves

Access tokens for the MCP server last 30 days, and until now there was no way to renew one: the token response carried neither an expiry nor a refresh token, so a connection stopped working a month after it was made with nothing to indicate why. Token responses now include expires_in and a refresh_token, and the refresh_token grant is supported. Refresh tokens rotate: each refresh consumes the presented token and issues a new one, as OAuth 2.1 requires for public clients. Presenting a refresh token twice revokes every token descended from that grant, on the basis that a replayed token has either leaked or the client has raced itself. Refresh tokens are valid for 90 days, so a client that idles for a month still reconnects without sending the user back through the browser. See MCP authentication.

​

Added: incident logging can be switched off

incident_moderation_enabled on your account controls whether Tuteliq writes any detection record at all. With it off, incidents, grooming assessments, in-app notifications and alert emails are all suppressed, while detection responses stay identical. Retention is also configurable, from 7 days to 2 years, defaulting to 90. See incident logging control.

​

Added: moderation through an AI agent

The MCP server exposes the incident queue, so an agent can size a backlog, triage by severity, inspect what matters and record decisions into the same audit trail a human moderator’s land in. Agents are limited to confirm. An MCP token identifies an account rather than an individual, so escalation, resolution and reopening stay with a human moderator, and a restricted action returns 403. Every state-changing call now requires moderator_external_id, and MCP-authenticated calls must use the form mcp:<agent-name>, so agent and human decisions stay distinguishable in the trail. See moderation workflows. Requires @tuteliq/mcp 3.21.1+; earlier versions send the attribution field unprefixed and the call is refused.

​

Changed: incident lifecycle matches the moderator dashboard

escalate previously produced a record indistinguishable from confirm, and resolved was unreachable through the API despite existing in the status enum. Both are fixed, and resolve and reopen are now available: resolve stamps resolved_at and marks linked notifications read, reopen returns an incident to new while preserving resolved_at as history.

​

Changed: retention now matches the Trust Center

Analysis results and detection records are retained for 90 days and API logs for 30, matching what the Trust Center publishes. Incidents and grooming assessments were previously kept for 365 days and API usage for 90.

---

​

v1.15.0 (August 2026)

​

Changed: new plans, and the free tier is retired

Tuteliq now has three plans: Build ($99/mo, 25,000 detection credits), **Assure** ($499/mo, 150,000 credits) and Enterprise (custom). Every plan starts with a 14-day free trial. The previous Starter, Indie, Pro and Business plans are closed to new signups. If you are on a legacy plan, nothing changes. Your limits, endpoint access and price stay exactly as they are. The /api/v1/pricing endpoint now returns only the plans available to new customers, so a legacy plan will not appear there even though your account continues on it. Allowances are measured in detection credits, not request counts. Endpoints are weighted by cost: a text detection is 1 credit, an image 7, a video 95.

​

Improved: animated GIFs are analysed frame by frame

/safety/image previously rejected animated GIFs with a retryable error, and flattening one to a still only ever exposed its first frame. GIFs are now decoded and up to four evenly-spaced frames are analysed, always including the first and last, then folded into a single worst-case verdict. Send the original GIF rather than a still. Cost is unchanged at image rates. Requires @tuteliq/sdk 2.21+ if you upload from a Buffer; earlier versions sent uploads without a MIME type.

---

​

v1.14.0 (August 2026)

​

Breaking: recommended_action is now a fixed set of values

On /safety/grooming, recommended_action previously returned a full sentence, for example "monitor the conversation for further signs of grooming, such as...". It now returns exactly one of five tokens, ordered weakest to strongest: none · monitor · flag_for_review · block · immediate_intervention This makes the field safe to switch on directly. The descriptive sentence has moved to a new optional action_detail field, intended for display in a moderator UI. If you were rendering recommended_action as text, read action_detail instead. Two legacy values were folded into the set at the same time: no_action is now none, and flag_for_moderator is now flag_for_review.

​

Breaking: the normalized block has been removed

Responses from /safety/bullying, /safety/grooming and /safety/unsafe no longer include the normalized object. It existed so that callers could branch on normalized.actionable at a time when recommended_action returned free text and could not be relied on. Now that recommended_action is a stable enum, the block only repeated fields already present at the top level, adding roughly 900 bytes to every response. Migrate by branching on recommended_action (flag_for_review, block or immediate_intervention indicate a case needing attention). The per-message detail previously projected into normalized.evidence is available in full as message_analysis.

​

Improved: 404 responses suggest the endpoint you meant

A mistyped path now returns a correction alongside the error, for example Did you mean POST /api/v1/safety/grooming? for /api/v1/safety/groomin. Calling a valid path with the wrong HTTP method is reported as such. The available_endpoints listing is now generated from the live route table, so it always reflects what the API actually serves.

---

​

v1.13.0 (July 2026)

​

Removed: model-based document authenticity and field extraction

On 2 July 2026 we removed every path that sent a verification image to a language or vision model. Three capabilities announced in earlier releases were withdrawn:

• Model-based document authenticity checks, which analysed document layout, security features, fonts and colour consistency against known templates, and claimed to detect screen photos, printout recaptures and digital manipulation (announced in v1.6.0, and as “hologram detection, tamper analysis” in v1.4.0).
• Model-based OCR fallback, which extracted name, date of birth and document number when local OCR failed (announced in v1.6.1).
• The document_authenticated signal that depended on them.

This was a deliberate privacy decision rather than a quality one. The most sensitive images we ever handle are a child’s selfie and an identity document, and we concluded that they should never reach a language model, including one we operate ourselves. The rule is now enforced in the codebase: nothing in the verification pipeline imports the model layer. What replaces it. Document validation is entirely deterministic and local: ICAO 9303 Machine Readable Zone parsing with check-digit validation, PDF417 and AAMVA barcode decoding, algorithmic document-number validation, and cross-referencing of every extracted field against the other sources that carry it. In practice the MRZ and barcode paths are stronger evidence than the removed checks, because a check digit either validates or it does not. What is not currently replaced. Detection of a photographed screen or printout of a document is not part of the verdict today. Selfie liveness anti-spoofing is unaffected and continues to run in full, including landmark motion, texture and moiré analysis, depth cues and cross-frame consistency. If document recapture detection returns, it will be a local model. If you relied on document_authenticated, treat MRZ or barcode validation as the equivalent assurance signal.

​

New: Fast mode (verdict-only)

Every detection endpoint now accepts options.verdict_only: true. In fast mode the response returns only the conversation-level verdict (risk level, categories, recommended action) and omits the per-message message_analysis breakdown. Because the engine no longer generates that breakdown, fast mode is meaningfully lower latency, which makes it the right choice for screening live chat in real time. Screen everything in fast mode, then re-run only the messages that flag through standard mode for the full per-message detail. See Fast mode. Available in @tuteliq/sdk 2.16+ (verdictOnly: true) and @tuteliq/mcp 3.18+ (verdict_only on detect_grooming and detect_bullying).

​

Improved: include_evidence now controls generation, not just output

Setting include_evidence: false (at the top level or inside options) now tells the engine to skip generating the per-quote evidence array altogether, instead of stripping it after the fact. The result is a smaller response and roughly a third lower latency, with the verdict, severity, categories and rationale unchanged. Use it when you need the decision but not the supporting quotes. The default remains true.

​

Fixed: include_evidence was previously accepted without effect

On the fraud and safety-extended endpoints, include_evidence: false was validated but never applied, so evidence came back regardless. The flag is now honored on every detection endpoint, and is accepted both at the top level and inside options.

​

Improved: Age- and role-aware grooming precision

detect_grooming now reasons about the actual age gap and role of each participant. Openly-arranged meet-ups and task-specific photo requests between similar-age peers are no longer flagged as grooming, while adult-to-minor and coercive same-age patterns (including peer sextortion) continue to fire at full severity. Pass sender_age per message, and context.child_age / context.participant_age when known, for the sharpest results.

​

New: Bullying, gaming toxicity, and coded-language detection

A new Bullying and toxicity detection page documents how Tuteliq reads context rather than keywords: gaming trash-talk versus targeted harassment, coded slang, emoji, algospeak, and deliberate filter evasion. This coded-language and evasion resilience is a platform-wide capability that applies to every detector, not just bullying.

​

SDK releases

• @tuteliq/sdk 2.16.0: adds verdictOnly fast mode across detection methods
• @tuteliq/mcp 3.18.0: adds verdict_only to detect_grooming and detect_bullying

---

​

v1.12.0 — May 2026

​

New — Read-only dashboard surface

Four new endpoints under /api/v1/incidents give programmatic access to the same data a moderator sees in the Tuteliq dashboard, scoped per API key:

• GET /api/v1/incidents — paginated, filterable listing (category, severity, status, source, platform, external_id, customer_id, time-window). Cursor-based pagination. Optional include_summary=true decrypts the summary text per row (extra credit per row).
• GET /api/v1/incidents/:id — full incident detail. Server-encrypted fields decrypted server-side; BYOK fields returned as hybrid envelopes for client-side decryption with the customer’s RSA private key.
• GET /api/v1/incidents/overview — KPI snapshot over a time window (total, requires-review queue size, 24h/7d/30d totals, counts by category/severity/source/status, top 5 platforms).
• GET /api/v1/incidents/trends — time-bucketed counts with per-bucket severity breakdown. Buckets: hour, day (default), week.

Available via @tuteliq/sdk 2.12+ (listIncidents, getIncident, getIncidentsOverview, getIncidentTrends) and @tuteliq/mcp 3.15.4+ (the four matching tools render as interactive dashboard widgets in MCP hosts).

​

Updated — detect_grooming 50-turn ceiling

Multi-turn detect_grooming calls have a documented ceiling of ~50 turns per request. Earlier this month a regression dropped the effective ceiling below 20 turns (the output-token budget was clipping the JSON response mid-message_analysis array, causing a parse failure that surfaced as analysis_status: "engine_error"). Fixed via an adaptive output budget that scales with conversation length (base + N × per-turn, capped at the model’s safe output ceiling). The graceful engine_error envelope still exists — beyond the ceiling, callers receive recommended_action: "flag_for_moderator" rather than a silent pass. For conversations beyond ~30 turns we recommend chunked sliding windows of 10–15 turns with continuation_token carrying trajectory state across calls. See Grooming detection for the tactic taxonomy and the long-conversation pattern.

​

Updated — meeting_request as a first-class grooming tactic

detect_grooming now returns meeting_request as a distinct flag (previously subsumed under boundary_pushing). Meeting attempts convert online risk into physical-safety risk and warrant a different operational response — most platforms route them directly to law-enforcement-aware workflows. The full six-tactic taxonomy is documented at Grooming detection.

​

New — Composable safety primitives

Documented cross_endpoint_modifier and vulnerability_profile on detect_vulnerability_exploitation. Both fields have been returning real data for months but were undocumented — they let agents amplify sibling detectors’ verdicts when a target is identified as vulnerable, and route to vulnerability-dimension-specific intervention workflows. See Composable safety primitives for the worked TypeScript + Python examples.

​

Fixed — Account observability endpoints

/account/consent and /account/audit-logs now respond reliably for accounts with larger or filtered datasets. Both endpoints are part of the GDPR / SOC 2 observability surface and should be considered stable for compliance workflows.

​

Fixed — Response key consistency across usage and pricing endpoints

/usage/monthly, /usage/history, and /pricing/details now return snake_case keys consistent with the rest of the API. SDK consumers no longer need to alias field names; existing snake_case-aware code works directly.

---

​

v1.11.0 — May 2026

​

Updated — Credit costs

Per-endpoint credit costs for text-detection endpoints have been updated. See Pricing & Credits for the full table. /analyse/multi now uses a 1 base + 1 per endpoint formula instead of summing per-endpoint costs (e.g. 3 endpoints = 4 credits, 7 endpoints = 8 credits). Server-side execution is parallel via Promise.allSettled, so the price reflects the actual wall-clock cost rather than the sum of individual calls.

---

​

v1.10.0 — May 2026

​

New — Prescreen lexicon for SDK-side filtering

Public endpoint GET /v1/prescreen/lexicon returns Tuteliq’s curated set of high-precision phrase patterns mapped to category flags. SDKs fetch it on startup, cache by version, and run it client-side against content before making an API call. Two wins:

• Clearly-benign content never leaves the device (greetings, thanks, small talk → SDK skips the API call entirely)
• Positive matches are attached as structured hints on the request via the new prescreen_flags field, giving the server a strong prior

v1 covers grooming, distress-signals, bullying, coercive-control, and self-harm crisis indicators in English. Multilingual coverage rolling out in subsequent lexicon versions. See the new Prescreen Lexicon page for the matching contract and SDK behaviour.

​

Improved — Continuation tokens now feed conversation trajectory into the LLM

When a continuation_token is supplied, the LLM now receives the derived conversation state (turn count, accumulated categories with max confidence, severity trajectory, last recommended action) as injected context for the new turn. Translation: the model considers escalation patterns across the whole conversation when scoring the new message — without us ever storing the conversation. This closes the gap between v1.9 (token carries state across calls; response shows trajectory) and the full conversation-aware analysis that was previously only available by re-sending conversation_history on every call. Same five endpoints: /safety/grooming, /safety/distress-signals, /safety/bullying, /safety/coercive-control, /safety/vulnerability-exploitation.

---

​

v1.9.0 — May 2026

​

New — Continuation tokens (privacy-first conversation memory)

Five safety endpoints now return a signed, opaque continuation_token in every response. Pass it back on the next call to get conversation-aware analysis without re-sending the conversation history and without Tuteliq ever storing it. The token IS the analysis state — you hold it, we hold the signature.

• Supported: /safety/grooming, /safety/distress-signals, /safety/bullying, /safety/coercive-control, /safety/vulnerability-exploitation
• What’s in the token: derived signals only — turn count, accumulated category confidences, severity trajectory, recommended actions
• What’s NOT in the token: raw text, PII, anything we couldn’t already return in the public API response
• Properties enforced by construction: per-key binding (cannot be replayed across customers), per-endpoint binding, tamper-evident HMAC, 24h default lifetime, bounded payload size
• New request fields: continuation_token, reset_conversation, message_id
• New response fields: continuation_token, continuation_expires_at, state_source (token / fresh / reset)
• See the new Continuation Tokens page for usage, errors, and best practices

/safety/grooming’s legacy session_id continues to work but is scheduled for deprecation — the continuation token is the privacy-first replacement.

​

Backwards compatible

• All existing requests work exactly as before. The new fields are optional.
• If you send a stale or scope-mismatched token, you get a structured CONTINUATION_TOKEN_* error and can fall back to one call with conversation_history to re-seed.

---

​

v1.8.0 — May 2026

​

Breaking — Endpoint renamed: emotional-distress → distress-signals

POST /v1/safety/emotional-distress is removed. The endpoint is now POST /v1/safety/distress-signals — request and response shapes are unchanged. Calls to the legacy path return HTTP 410 Gone with a JSON body pointing to the new endpoint. Why: EU AI Act Article 5(1)(f) framing. The detector is and always was a content classifier — it identifies linguistic patterns in text that are historically associated with distress. The old name implied emotion recognition, which is prohibited in workplaces and educational institutions outside child-safety contexts. The rename, an updated prompt that explicitly disclaims emotion inference, and a published intended-use scope (safety component for child-safety; not for workplace or general-purpose education monitoring) keep the endpoint inside the Art 5(1)(f) safety exception. Action required:

• Update integrations to call POST /v1/safety/distress-signals
• In /analyse/multi, change endpoint ID emotional-distress → distress-signals
• MCP tool renamed: detect_emotional_distress → detect_distress_signals
• Settings keys (enabled_endpoints / disabled_endpoints) renamed accordingly

Response shape, categories (LONELINESS, FEELING_UNHEARD, OVERWHELM, etc.), exploitation-risk fields, and the 5-credit cost are unchanged.

​

Breaking — Removed: per-customer synthetic profile endpoint

GET /v1/safety/synthetic-content/profile/:customer_id is removed. The endpoint and the underlying per-customer aggregation have been deleted; stored aggregates have been purged from cache. Why: EU AI Act Article 5(1)(c) social-scoring concern and GDPR data minimisation. A per-natural-person aggregated synthetic-content score retained across detections resembles social scoring when used across unrelated contexts. Removing both the endpoint and the storage closes the retrieval surface and the stored artefact. Action required: If you were polling the profile endpoint, aggregate detection responses on your side instead. The per-call detection endpoints (/safety/synthetic-content, /image, /audio, /video) continue to return classifications unchanged.

---

​

v1.7.0 — May 2026

​

Emotional Distress Detection (Early Warning)

• POST /v1/safety/emotional-distress — detects pre-vulnerability emotional distress signals before exploitation begins
• Based on criminological research: emotional distress is both a consequence AND a driver of vulnerability to child sexual exploitation
• 12 distress signal categories: loneliness, feeling unheard, overwhelm, low self-worth, trust-seeking, withdrawal, family conflict, identity distress, academic pressure, sleep disturbance, appetite change, emotional numbing
• Exploitation risk assessment — returns per-type vulnerability scores for grooming, sextortion, trafficking, radicalisation, self-harm, and substance use
• Vulnerability level scoring — none, low, elevated, high, critical
• Key risk combinations: loneliness + trust-seeking = HIGH grooming risk; family conflict + withdrawal = HIGH trafficking risk; low self-worth + overwhelm = HIGH sextortion risk
• 5 credits per call

​

Tech-Facilitated Gender-Based Violence (TFGBV) Detection

• POST /v1/safety/tfgbv — detects technology-facilitated gender-based violence, rooted in gender inequalities and power imbalances
• 13 TFGBV categories: image-based abuse, cyber stalking, online harassment, doxing, impersonation, outing, post-separation abuse, sextortion, digital coercion, sexualised deepfakes, gendered hate speech, reproductive surveillance, economic abuse digital
• TFGBV amplifier assessment — identifies which technology amplifiers are active: scale, speed, anonymity, permanence, cross-border reach
• Intersectionality flags — identifies compounding discrimination factors (LGBTQIA+, disability, racial, etc.)
• Child-specific detection with heightened sensitivity for minors
• 5 credits per call

​

Multi-Endpoint Support

• Both new endpoints available in /analyse/multi — use endpoint IDs emotional-distress and tfgbv
• MCP tools: detect_emotional_distress and detect_tfgbv

---

​

v1.6.1 — April 2026

​

Credit Pricing Update

All per-endpoint credit costs have been revised to reflect processing complexity more accurately. See Pricing & Credits for the full breakdown. Text & Safety Detections: 5 credits per call (was 1)

• detectBullying, detectUnsafe, detectSocialEngineering, detectAppFraud, detectRomanceScam, detectMuleRecruitment, detectGamblingHarm, detectCoerciveControl, detectVulnerabilityExploitation, detectRadicalisation

Conversation-Based: 6 credits per 10 messages (was 1)

• detectGrooming, analyzeEmotions

Analysis & Guidance:

• getActionPlan: 9 credits (was 2)
• generateReport: 13 credits (was 3)

Media:

• analyzeImage: 7 credits (was 3)
• analyzeVoice: 21 credits base + 15 per extra minute over 60s (was flat 5)
• analyzeVideo: 95 credits (was 10)
• Voice stream (per flush): 7 credits (was 1)
• Video stream (per frame): 7 credits (was 3)
• detectSyntheticContent: 5 credits (was 2)

Verification:

• verifyAge (full): 20 credits (was 5)
• verifyAge (liveness only): 10 credits (new)
• verifyIdentity: 25 credits (was 10)

Document Analysis:

• analyzeDocument: minimum 10 credits (was 3)

​

Age & Identity Verification Improvements

• Liveness-only mode now charged at 10 credits (reduced from full verification cost)
• Passport number validation — passport document numbers are no longer incorrectly validated against national ID formats (e.g., Swedish personnummer)
• Vision AI OCR fallback — when Tesseract OCR fails, document fields (name, DOB, document number) are extracted via vision AI and the confidence score reflects the extraction quality
  Withdrawn 2 July 2026. No verification image reaches a model. See v1.13.0.
• i18n for camera overlay — all liveness challenge text (instructions, blink/smile prompts) now translates correctly when switching languages

---

​

v1.6.0 — April 2026

​

Multi-Signal Forensic Synthetic Content Detection

The synthetic content detection pipeline has been completely rebuilt into a multi-signal forensic system that runs up to 6 independent analysis engines in parallel for images and 5 for video — replacing the previous single-model approach. Image Detection — 6-Signal Pipeline:

• EXIF Metadata Extraction — Detects AI generator signatures in EXIF tags, XMP data, and PNG tEXt chunks (Stable Diffusion parameters). Flags suspicious absence of camera metadata (no camera model + no GPS + high resolution).
• Pixel Statistics — Shannon entropy, Laplacian edge density, and channel uniformity analysis. GAN images produce distinctive statistical signatures.
• C2PA Content Credentials — Detects and validates C2PA manifests from DALL-E 3, Adobe Firefly, Google Imagen. When C2PA declares AI generation, the result is forced to confirmed_synthetic with confidence ≥ 0.95.
• Invisible Watermark Detection — High-frequency energy analysis, periodic pattern detection at known watermark frequencies, LSB distribution analysis, and corner entropy checks.
• Perceptual Hashing (pHash) — DCT-based 64-bit perceptual hash compared against a Redis-backed database of known synthetic content via Hamming distance. Matches force confirmed_synthetic.
• Weighted Signal Aggregation — All signals aggregated into a weighted ensemble (vision 30%, metadata 15%, pixel stats 15%, C2PA 15%, watermarks 10%, pHash 15%) with fault isolation via Promise.allSettled.

Audio Detection — Spectral Forensics:

• Mel Spectrogram Analysis — FFmpeg generates a mel spectrogram image, analyzed by a dedicated vision prompt for frequency band uniformity, harmonic anomalies, missing breath noise, onset/offset patterns, and aliasing artifacts.
• Quantitative Audio Statistics — RMS mean/peak, dynamic range, silence ratio, flat factor, and DC offset extracted via FFmpeg astats filter.
• Spectral analysis runs in parallel with transcription. Even speech-free audio can be flagged if spectral patterns indicate synthesis.

Video Detection — Temporal + Lip-Sync:

• Temporal Consistency Analysis — face-api.js tracks face identity across frames via 128-dimensional descriptor Euclidean distance. Real faces: < 0.4, deepfakes: spikes > 0.6. Landmark stability measured via eye-to-nose ratio variance.
• Lip-Sync Correlation — Mouth openness from 68-point face landmarks correlated against frame-aligned audio energy. Pearson correlation > 0.5 = authentic, < 0.3 = deepfake. Detects silent mouth movement and voice-without-movement.
• All 5 video analysis tracks (per-frame vision, temporal consistency, lip-sync, spectral, transcription) run via fault-isolated Promise.allSettled.

Account-Level Profiling:

• GET /v1/safety/synthetic-content/profile/:customer_id — 30-day rolling window with synthetic count, account score, trend detection (increasing/stable/decreasing), and category distribution.
• Automatic, zero-cost profiling when customer_id is provided on any detection request.

New Response Fields (additive — no breaking changes):

• Image: metadata_analysis, provenance, forensic_signals, perceptual_hash, known_synthetic_match
• Audio: audio_stats, spectral_signals
• Video: temporal_consistency, lip_sync, audio_stats, spectral_signals

​

Age & Identity Verification Improvements

• ICAO 9303 MRZ Validator — Full check digit validation for TD1 (ID cards, 3×30), TD2 (ID cards, 2×36), and TD3 (passports, 2×44) Machine Readable Zones with weighted mod-7 algorithm
• PDF417 Barcode Reader — Decodes AAMVA-structured data from US/Canadian driver’s licenses via zxing-wasm — extracts name, DOB, expiry, document number, address, and more
• 45-Country Document Number Validator — Algorithmic check digit verification for CPF (Brazil), personnummer (Sweden), Aadhaar (India), Codice Fiscale (Italy), CURP (Mexico), SSN (US), SIN (Canada), TFN (Australia), and 37 more country-specific document formats
• Visual Liveness Analyzer — Multi-signal liveness detection: landmark motion analysis, texture analysis (Laplacian variance + moire detection), depth cue analysis (face/background sharpness ratio), and cross-frame consistency checks
• AI-Powered Document Authenticator — Vision model analyzes document layout, security features, fonts, color consistency, and photo integration against known templates. Detects screen photos, printout recaptures, and digital manipulation.
  Withdrawn 2 July 2026. No verification image reaches a model. See v1.13.0.
• MRZ/OCR/Barcode Cross-Referencing — Compares name, DOB, and document number between MRZ, OCR text, barcode data, front and back sides, flagging any inconsistency as potential tampering

---

​

v1.5.0 — March 2026

​

Document Analysis

• POST /v1/safety/document — Upload a PDF (max 50 MB, 100 pages) for multi-endpoint safety analysis with per-page detection results
• Supports 8 detection endpoints: unsafe, bullying, grooming, social-engineering, coercive-control, radicalisation, romance-scam, mule-recruitment
• SHA-256 document hashing for chain-of-custody verification in compliance audits
• Zero-retention processing — no document data stored after response
• Bounded concurrency (3 pages at a time) with text chunking for long pages

​

Dynamic Credit Pricing for Documents

• Document analysis uses per-page, per-endpoint pricing: max(10, pages_analyzed × endpoint_count) (minimum updated in v1.6.1)
• Each page-endpoint combination costs 1 credit
• Minimum charge of 10 credits covers extraction overhead
• Examples: 5 pages × 3 endpoints = 15 credits; 20 pages × 8 endpoints = 160 credits

​

Model Refinements

• TCO Regulation classification — Radicalisation responses now include EU Regulation 2021/784 Art. 2(7) content classification
• Self-harm CONTAGION sub-category — Detects suicide pacts, method sharing, cluster effects, and gateway escalation
• PII/Doxxing detection — New PII_DOXXING category for doxxing threats and digital footprint weaponisation
• Legacy response normalization — Bullying, grooming, and unsafe endpoints now include a normalized block with a unified response shape matching newer endpoints

---

​

v1.4.4 — March 2026

​

New Features

• country context field — Pass an ISO 3166-1 alpha-2 country code (e.g., "GB", "US", "SE") in the context object to receive geo-localised crisis helpline data in detection responses. Falls back to user profile country if omitted.
• Improved action escalation for minors — All detection endpoints now enforce a minimum flag_for_review action when harm is detected and the subject is a minor. Criminal indicators (SEXTORTION, TRAFFICKING, CSAM, DEBT_BONDAGE, FORCED_CRIMINALITY, HONOUR_ESCALATION) targeting minors automatically escalate to immediate_intervention.
• Graduated risk scoring — Risk scores now use the full 0.0–1.0 range with graduated bands instead of clustering around a single value.
• Evidence tactic normalization — Evidence tactic fields are now always returned in SCREAMING_SNAKE_CASE format (e.g., "EMOTIONAL_MANIPULATION" instead of "Emotional Manipulation").

​

SDK Releases

• @tuteliq/sdk v2.5.0 — adds country context field, graduated risk scoring, tactic normalization
• @tuteliq/mcp v3.7.0 — adds country context field, minor action escalation, tactic normalization

---

​

v1.4.3 — March 2026

​

Bug Fixes & Improvements

• support_threshold now works correctly across all 11 detection endpoints — setting "critical" correctly suppresses crisis helplines for High severity results. Previously the parameter was accepted without error but had no effect on the response.
• sender_trust: "verified" fully suppresses AUTH_IMPERSONATION — verified senders no longer trigger impersonation detection. Routine urgency (schedules, deadlines, appointments) is also suppressed. Only genuinely malicious elements (credential theft, phishing links, financial demands) will flag a verified sender.
• Empty categories field in detect_unsafe fixed — stronger prompt enforcement and code-level fallback extraction ensure the categories array is populated when the rationale references specific harms.
• MCP session recovery — stale MCP sessions (after Cloud Run deployment or instance recycling) now recover transparently without requiring the client to reconnect.

​

Documentation

• Added support_threshold behaviour reference to API docs, Node SDK, and MCP README
• Added sender_trust effect on scoring to all documentation
• Added analyse_multi accepted endpoint values to MCP README and API docs
• Added complete context field reference across all packages

​

SDK Releases

• @tuteliq/sdk v2.4.0 — adds confidence to AnalyzeResult, optional risk_level to UnsafeResult
• @tuteliq/mcp v3.5.0 — wires support_threshold through all detection MCP tools, shows confidence and support in analyze tool

---

​

v1.4.2 — March 2026

​

Full EU Language Coverage

• Added 13 new languages: Romanian (ro), Greek (el), Czech (cs), Hungarian (hu), Bulgarian (bg), Croatian (hr), Slovak (sk), Lithuanian (lt), Latvian (lv), Estonian (et), Slovenian (sl), Maltese (mt), Irish (ga)
• Tuteliq now supports 27 languages — all 24 EU official languages + Ukrainian, Norwegian, and Turkish
• Enhanced all language entries with grooming indicators, self-harm coded vocabulary, filter evasion techniques, and youth slang coverage
• All new languages are in Beta status

---

​

v1.4.1 — March 2026

​

Language Support Expansion

• Added 4 new languages: Dutch (nl), Polish (pl), Italian (it), Turkish (tr)
• Tuteliq now supports 14 languages with auto-detection and culture-aware analysis
• Each language includes culturally-specific safety guidelines, slur databases, and prompt calibration
• All new languages are in Beta status

---

​

v1.4.0 — March 2026

​

Age Verification (Beta)

• POST /v1/verification/age — verify user age through document analysis and biometric age estimation
• Supports government-issued ID documents (passport, driving licence, national ID)
• Biometric age estimation from selfie photo
• Returns verified, estimated_age, age_range, confidence, and document_type
• 20 credits per verification (updated in v1.6.1)
• Available on Pro tier and above

​

Identity Verification (Beta)

• POST /v1/verification/identity — confirm user identity with document verification and liveness detection
• Document authenticity checks (MRZ validation, hologram detection, tamper analysis)
  MRZ validation still runs and is now check-digit validated. Hologram and tamper analysis were withdrawn 2 July 2026; see v1.13.0.
• Liveness detection to prevent spoofing (photo-of-photo, screen replay, mask attacks)
• Face matching between document photo and selfie
• Returns verified, match_score, liveness_passed, document_authenticated, and flags
• 25 credits per verification (updated in v1.6.1)
• Available on Business tier and above

​

Tier Access

• Age Verification requires Pro tier ($99/mo) or above
• Identity Verification requires Business tier ($349/mo) or above
• Both features are in Beta — endpoints and response schemas may evolve

---

​

v1.3.0 — February 2026

​

Fraud Detection Endpoints

• POST /v1/fraud/social-engineering — detect social engineering tactics: pretexting, impersonation, urgency manipulation, authority exploitation
• POST /v1/fraud/app-fraud — identify fraudulent app promotion, fake reviews, malicious download links, and clone app distribution
• POST /v1/fraud/romance-scam — detect romance scam patterns: love-bombing, financial requests, identity fabrication, isolation tactics
• POST /v1/fraud/mule-recruitment — flag money mule recruitment: easy money offers, account sharing requests, laundering language

​

Safety Extended Endpoints

• POST /v1/safety/gambling-harm — detect gambling harm: underage gambling promotion, addiction patterns, predatory odds, bet pressure
• POST /v1/safety/coercive-control — identify coercive control: isolation tactics, financial control, monitoring/surveillance, threat patterns
• POST /v1/safety/vulnerability-exploitation — detect exploitation of vulnerable individuals with cross-endpoint vulnerability modifier and vulnerability profile scoring
• POST /v1/safety/radicalisation — flag radicalisation indicators: extremist rhetoric, us-vs-them framing, recruitment patterns, dehumanisation

​

Multi-Endpoint Analysis

• POST /v1/analyse/multi — fan-out a single text to up to 10 detection endpoints in parallel
• Automatic vulnerability modifier: when vulnerability-exploitation is included, its cross-endpoint modifier adjusts severity scores across all other results
• Aggregated response with summary.highest_severity, summary.total_credits_used, and per-endpoint breakdown
• Supports both legacy (bullying, grooming, unsafe) and new detection endpoints

​

SDK & Tool Support

• All new endpoints available across all SDKs: Node.js, Python, Swift, Kotlin, Flutter, React Native, .NET, Unity
• CLI — new detect social-engineering, detect app-fraud, detect romance-scam, detect mule-recruitment, detect gambling-harm, detect coercive-control, detect vulnerability-exploitation, detect radicalisation, and detect multi commands
• MCP Server — 10 new tools for fraud, extended safety, multi-endpoint analysis, and video analysis

​

Tier Access

• New endpoints require Indie tier or above (Starter tier retains access to bullying, grooming, and unsafe only)
• All new endpoints cost 5 credits per call (updated in v1.6.1); multi-endpoint costs the sum of individual endpoints

---

​

v1.2.0 — February 2026

​

Video Analysis

• POST /v1/safety/video — upload a video file (mp4, webm, quicktime, avi — max 100MB, 10 min) and receive per-frame vision analysis with flagged timestamps
• Frame extraction powered by ffmpeg with configurable max_frames (default 10, max 20)
• Aggregated overall_risk_score and overall_severity across all frames
• Automatic incident recording and webhook alerts for flagged videos
• 95 credits per video analysis (updated in v1.6.1)

​

Video Streaming

• WebSocket voice streaming now supports video frames alongside audio
• Send video frames as binary with a 0x01 prefix byte; audio uses 0x00 or no prefix (backward compatible)
• New frame_analysis server event with per-frame vision results
• New config options: enable_video, frame_interval_seconds (min 3s, default 5s)
• session_summary event now includes video_frames_analyzed count
• 7 credits per video frame analysis (updated in v1.6.1)

​

Voice Streaming Enhancements

• Updated authentication: ?api_key= query param or Authorization: Bearer header
• Tier-based connection limits (1 for Starter, up to unlimited for Enterprise)
• Subscription-aware credit tracking per flush
• Heartbeat ping/pong every 30 seconds for stale connection detection

---

​

v1.1.0 — February 2026

​

Multilingual Support

• 27 languages supported — English (stable), all 24 EU official languages + Ukrainian, Norwegian, Turkish (beta)
• Layered auto-detection — Trigram analysis (franc) confirmed by LLM-based detection for maximum reliability
• Culture-aware analysis — Language-specific guidelines for local slang, idioms, and harmful terms injected into classification prompts
• All safety endpoints now return language, language_status, and detected_language fields in responses
• No explicit language parameter required — detection is fully automatic

---

​

v1.0.0 — February 2026

Initial public release of the Tuteliq API.

​

Safety Endpoints

• POST /v1/safety/unsafe — detect harmful content across all nine KOSA categories
• POST /v1/safety/bullying — dedicated bullying and harassment detection
• POST /v1/safety/grooming — conversation-level grooming pattern analysis
• POST /v1/safety/voice — audio file transcription and safety analysis
• POST /v1/safety/image — image analysis for visual content risks

​

Analysis & Guidance Endpoints

• POST /v1/analysis/emotions — emotional well-being analysis from text and conversations
• POST /v1/guidance/action-plan — age-appropriate guidance and intervention recommendations
• POST /v1/reports/incident — structured safety report generation

​

Voice Streaming

• WSS /v1/safety/voice/stream — real-time voice moderation via WebSocket with configurable severity thresholds

​

Webhooks

• POST /v1/webhooks — register webhook endpoints for safety alerts and batch events
• HMAC-SHA256 signature verification on all webhook deliveries
• Automatic retry with exponential backoff (3 attempts)

​

GDPR Compliance

• DELETE /v1/account/data — right to erasure (Article 17)
• GET /v1/account/data/export — right to data portability (Article 20)
• PATCH /v1/account/data — right to rectification (Article 16)
• Consent management endpoints for granular data processing consent
• Public transparency endpoints: DPA, sub-processors, retention policy

​

SDKs & Tools

• Node.js SDK — @tuteliq/sdk on npm
• Python SDK — tuteliq on PyPI
• Swift SDK — via Swift Package Manager
• Kotlin SDK — via Maven Central
• Flutter SDK — tuteliq on pub.dev
• React Native SDK — @tuteliq/react-native on npm
• .NET SDK — Tuteliq on NuGet
• Unity SDK — via Unity Package Manager
• CLI — @tuteliq/cli on npm / Homebrew
• MCP Server — @tuteliq/mcp on npm

​

Platform

• Credit-based billing with per-endpoint pricing
• Tier-based rate limiting (Free, Basic, Premium, Enterprise)
• API key authentication via Bearer token or x-api-key header
• Full KOSA harm category coverage with age-calibrated severity scoring