Calibrate your A.I.'sTemporal Clock™
One API call. A deterministic, model-aware calibration block that anchors your A.I. to the present moment and closes the gap from your model's training cutoff. Keep it from mixing up instances, sequences, and temporal perspectives.
The problem
Every A.I. model risks temporal confusion — and most teams don't realize how much it costs them.
Date-aware reasoning is the silent failure mode behind a huge class of "the model just made that up" bugs. The symptoms look like hallucinations; the root cause is an unsynchronized clock.
Why smarter models won't fix this:
Linguistic reasoning isn't time-based reasoning.
Language-based reasoning about time rests on pattern recognition — the model maintains an explanation of time, not a measurement of it. You can't calibrate numbers by writing more words; the model needs the gap quantified, not described.
It thinks today is the future.
An A.I.'s training data ends on a specific date — but the model has no internal sense of how much time has passed since.
Temporal drift over time.
Models treat their knowledge cutoff as the present. "This year's" reports, "the latest" version, "recently announced" — all of it anchors to a model's past training year, no matter when you're asking from.
A wall clock alone isn't enough.
Stamping the current instant into the prompt tells the model what time it is right now — but says nothing about how much time has elapsed since its training data ended. The model still reasons as if no time has passed, just with a fresher timestamp pasted on top.
One date stamp can't carry a session.
Even a correct, current date written once at the top of a prompt is a frozen snapshot. Long-running agents, multi-turn conversations, and scheduled jobs need an ongoing sense of how much time has passed since they started — not a string that goes stale the moment it's written.
Basic fixes make it worse.
The workarounds developers implement, leave the underlying discrepancy in place:
- Hard-coding "today is 2026-05-04" at the top of the prompt. Doesn't dislodge the model's prior — output contradicts itself, citing your date alongside facts dated to the cutoff.
- Asking the model what year or date it is. Elicits a confident guess anchored to training, not a measurement.
- Pasting in fresh web text with no temporal framing. The model sees new strings but has no signal that they're newer than what it already "knows."
- Injecting a server-side timestamp (new Date(), wall clock). Tells the model the current instant but nothing about the gap to its training, so it still reasons as if no time has passed.
- RAG that retrieves recent documents. Surfaces fresh content without telling the model how recent it is relative to its own knowledge — old and new get weighted the same.
- System-prompt or fine-tune notes about the model's own training boundary. The model will quote the boundary back to you and then reason past it anyway; static text can't track a moving "now."
The solution
Calibrate and connect with the present moment.
A precision-calibrated block corrects temporal reasoning by re-anchoring a model from its original training cutoff — eliminating the temporal blind spot at the source. A triangulated bridge then injects fresh knowledge to stabilize the prompt. Meanwhile, our Temporal Spiral keeps everything synchronized to a live multi-host time anchor. Fuse them into a complete temporal function call with POST /api/v1/full.
Model-aware
We maintain a curated registry of every major model's training cutoff — OpenAI, Anthropic, Google, Meta, Mistral, Cohere, xAI, DeepSeek, Qwen, and more. Pass a model name; we resolve the right calibration without your team tracking provider release notes.
Deterministic & cacheable
Same inputs, same output, byte-for-byte. Every response carries a weak ETag so your edge cache and your client both behave correctly. No hidden state, no time-dependent randomness — just an auditable function of the inputs you sent.
User-local time, baked in
Pass an IANA timezone and a clock-format hint and the response includes a properly formatted local-time block — so the model's sense of "now" matches the user's wall clock, not the server's.
Stays accurate over time
Models drift. New checkpoints, post-training updates, the occasional surprise. The registry is versioned, every response is stamped, and per-caller overrides exist for the cases where you need to take control.
Versioned & auditable
Every response carries meta.algorithmVersion and an X-Algorithm-Version header — bumped any time the block changes. GET /api/v1/registry/drift reports any divergence between our registry and the public reference dataset. The ETag is bound to both, so when the registry or algorithm changes, your cache invalidates automatically.
Freshness-scored citations
Bridge results carry a freshness score per citation and a meta.freshnessMedian across the set, mirrored in the X-Freshness-Median response header. A briefing built on five-year-old pages can't masquerade as a fresh one just because the citations span a lot of reputable domains — freshness is scored independently and surfaced alongside the source set.
Bridge the cutoff gap
Some questions need knowledge the model never had — release dates, current prices, who-just-shipped-what. POST /api/v1/full pairs the calibration block with an anchored bridge briefing in one call. Bring your own sync key (Brave, SerpAPI, Perplexity, or your existing OpenAI / Anthropic / Google key), get back cited sources the model can ground its answer in instead of confabulating. Idempotency-Key support, dry-run mode for CI, automatic backoff against degraded upstreams.
Multi-host time anchor (Pro and Enterprise)
The Temporal Spiral doesn't trust any single clock. Every sync cycle queries a fleet of NTP and NTS servers in parallel, weights each response by its round-trip time (or by long-run Allan-deviation stability once enough history accumulates), and combines them into a single BIPM-style ensemble estimate. Pro and Enterprise responses surface the full breakdown on anchor.ensemble (and meta.spiral.ensemble on /v1/full) — per-host samples, weights, the active weighting basis, combined offset, and dispersion — so accuracy-sensitive integrations can audit exactly where the time came from.
Keep a pulse on the present moment.
Temporal Spiral — every scale, sampled live.
Classical clockwork, upgraded. Your model receives a live temporal signal it can actually compute on — not a flat date string it can read but not understand. We encode 'now' as a series of interconnected wheels — scales from seconds to deep cosmological time, each turning at its own steady rate, temporal mapping in real time. As a precision chronometer, every response ships with a measurable confidence band.
The current instant — formatted for you, encoded into your model
In the atom demo, milliseconds and microseconds resolve from the API; nanoseconds, picoseconds, femtoseconds, and attoseconds are derived and visualized in your browser (browsers clamp clock reads to roughly 1 ms).
Running entirely in your browser against your local clock — no API key, no network call. µs renders as a pulse through Spectre mitigation. ms is the fast sweeping hand, visibly tracking each frame. When the hosted block ships from the API, the coarse scales are anchored against an audited external time reference and the finer scales are extrapolated forward with a measurable confidence band — so a calibration eval that hard-codes “this year” doesn’t silently drift as the year-scale hand rotates past it. Time mapping is difficult because linear data misses recurrence, while cyclical misses progress; the spiral is the frequency that ties both together. Temporal Spiral gives the model a continuous coordinate to reason against. The same spiral can be projected through any calendar you reason in — Gregorian for civic time, Holocene for a single positive year axis, Cosmological (anchored at the Big Bang) for deep time, with more options (scientific, historical, and galactic- year mapping) on the way.
The hosted Spiral block, anchored to NIST and surfaced inline on /api/v1/full via includeSpiralBlock, is a paid feature on Standard and above — see the pricing section below for the per-tier breakdown.
Pick a calendar per request: add ?calendar=holocene to /api/v1/spiral or pass "calendar": "cosmological" in the /api/v1/full body. Every successful response echoes a meta.calendar block so the projection is auditable; default stays gregorian if omitted.
Getting started
One request, one calibration, prepended.
The whole API surface — for the common case — is a single POST. Below is everything you need to wire it in.
Get an API key
Join the waitlist. We'll send a `tblk_live_*` key for production traffic and a `tblk_test_*` key for development.
Call /api/v1/calibrate
Pass the model name your application uses (and any optional context — timezone, clock format). You get back a ready-to-prepend calibration block.
Prepend it to your system prompt
Stitch the returned block in front of whatever system prompt you already use. Done. The block is deterministic, so cache it for the lifetime of the inputs.
Scale up to /api/v1/full
Four blocks in one call — a full 4D temporalBLOCK: calibration sync (Calibration Block), anchored time mapping (Spiral Block), distributed event mapping (Causal Block), and sensor stamping (Vision Block). Plan tiers set how deep each block goes.
Pricing
Plans, coming soon.
Pick a plan. Pay-as-you-go covers anything past the included quota.
Lite
Ideal for small scale AI applications and LLM chatbots.
- Unlimited calibration calls
- 100 snippets / month
- 25 briefs / month
- 1 live + 1 test API key
- 60 requests / minute
- Time confidence: basic — spiralCoordinate only (no confidenceMs / breakdown / cyclePhase, no embedded Spiral block)
- Output Pacing (stateless estimate only) — one-shot tokens-to-time or time-to-tokens; live sessions require Standard+
- Skill delivery audit log: 7 days
- Email support
Standard
Production-ready for most apps.
Early adopter perk: 20% off the first 3 months.
- Unlimited calibration calls
- 500 snippets + 500 briefs + 50 deep / month
- "Bring-Your-Own" model key (OpenAI / Anthropic / Gemini) per call — unmetered
- 3 live + 3 test API keys
- 300 requests / minute
- Time confidence: Spiral block + scalar confidenceMs
- Skill delivery audit log: 30 days
- Vision Block (POST /v1/spiral/batch) — up to 100,000 stamps/minute, rate-limited, not metered
- Output Pacing (live sessions) — Spiral-anchored elapsed time, smoothed rate, and projected finish; rate-limited, not metered
- Per-month USD budget cap (opt-in)
- Full support (email + chat)
Pro
Production at scale.
Early adopter perk: 20% off the first 3 months.
- Unlimited calibration calls
- 1,500 snippets + 1,500 briefs + 150 deep / month
- "Bring-Your-Own" model key (OpenAI / Anthropic / Gemini) per call — unmetered
- 10 live + 10 test API keys
- 600 requests / minute
- Time confidence: Spiral block + asymmetric confidenceBreakdown + syncCyclePhase + signed displacement (recall vs. projection)
- Observer-frame relativistic correction — pass any Earth-relative position (lat / lon / altitude, optional velocity) on `/v1/spiral` and the response carries the applied SR + GR offset. Useful for physics demos, GPS-relativity, drone/balloon telemetry, and any setup where the clock is in motion or operating well beyond MSL
- Tenant-wide defaults
- Skill delivery audit log: 90 days
- Vision Block at scale — up to 500,000 stamps/minute, rate-limited, not metered
- Output Pacing — live sessions + stateless estimate; measured rate stored across sessions for tighter projections
- Per-key analytics + drift alerts
- Per-month USD budget cap (opt-in)
- Priority support
Enterprise
Volume-based, self-hosted, specialized time anchors.
- Security-first by design — modern transport encryption, post-quantum-ready key exchange, SSO, and SOC 2 review on request (full cryptographic detail on the Security and Privacy pages)
- No published cap
- 99.9% uptime SLA
- Time confidence: Pro block + per-tenant access to spiral_anchor_* audit events
- Skill delivery audit log: 365 days
- Vision Block — uncapped stamps/minute (rate-limited, not metered; per-tenant ceilings on request)
- Output Pacing — uncapped sessions; customizable per-tenant rate limits on request
- Custom anchor sources on request — bring your own PTP / White Rabbit / atomic-clock feed; the Spiral is source-agnostic and the confidence band tightens automatically
- Robotics & control-loop integrations: continuous-input Spiral signal designed to feed perception and planning stacks alongside model calls
- Periodic-orbit observers — full per-rung Spiral coordinate across all twelve scales (µs → millennium) with rung-moon kinematics; built for satellites, ground stations, and planetary-mission timelines
- Self-hosted Docker image
- Slack-shared support channel
- Direct founder access
Quotas, per-tier rate limits, and the per-call syncTier field describe planned launch behavior. Backend enforcement is in progress and ships ahead of paid availability. Early adopter discounts apply to the first 3 months of paid service, then revert to the standard published rate.
Educator & researcher offers
Small research teams, student labs, and university programs can request a custom offer — specific to the task. Reach out and tell us what you're working on!
Pay-as-you-go, after included quota
Same rates on every paid tier. Your API keeps working past the included quota — calls are billed per-call at standard published rates and shown on your dashboard before the next invoice. Set an optional per-month USD budget cap from your dashboard and the API returns HTTP 402 monthly_budget_cap_exceeded before any further PAYG charge once the cap is reached.
Lite-tier keys that request gated options fail fast with an HTTP 402 before any upstream search work or billing event: DEEP_REQUIRES_PAID_TIER when syncTier:"deep" is set, and SPIRAL_BLOCK_REQUIRES_PAID_TIER when includeSpiralBlock:true is set on /v1/full. Upgrade to Standard or higher, or drop the gated option — full details in the Error codes table below.
| Call type | Rate |
|---|---|
| Calibration overage | Free |
| Snippet | $0.02/call |
| Brief | $0.04/call |
| Deep | $0.10/call |
| BYO model (per-call choice) | Free (you pay your model bill directly) |
Pick the sync backend per call
The syncTier field on each request lets you trade cost for depth — or route through your own model key when you've added one. Query by query, no plan change required.
Snippet
Raw results. Cheapest — your model synthesizes downstream.
Brief
Pre-synthesized paragraph with citations. Drop straight into a system prompt — no downstream synthesis step.
Deep
Deeper sync with a stronger model behind the synthesis. For citation-heavy queries (legal, medical, financial). Available on Standard, Pro, and Enterprise; not available on Lite.
BYO model
Route the call through your own OpenAI / Anthropic / Gemini key. Unmetered on our side — you pay your model bill directly. Available on any tier where you've added a BYO key.
Time confidence (Spiral block)
Included on every paid tier — no per-call PAYG charge. Depth scales with your plan, not with the per-call syncTier: Lite gets the basic rung (meta.spiralCoordinate only, no embedded Spiral block), Standard adds the embedded Spiral block + scalar confidenceMs, Pro adds the asymmetric confidenceBreakdown, syncCyclePhase, and signed displacement (recall vs. projection), Enterprise adds per-tenant access to spiral_anchor_* audit events. See the tier cards above for the full ladder.
Join the waitlist
Be among the first to ship time-aware A.I. models.
We're inviting teams today. Drop your email and a sentence on what you're building — we'll get you a key.