Gate weak retrieval
before it answers.
Ravel scores retrieval confidence and freshness on the request path, holds answers that miss the bar, and appends each decision to an immutable audit ledger you can replay later.
Confidence gating
Below-threshold retrieval never reaches the user unreviewed.
Immutable audit log
Each decision written once, exportable, and replayable.
Audit exports
Evidence packs formatted for auditors and risk reviews.
trust_ledger.log
| Time | Source | Conf | Age | Gate |
|---|---|---|---|---|
| 14:02:41.118 | policy_handbook_v9trc_8f21ad | 0.94 | 2d | passed |
| 14:02:43.507 | rate_sheet_q3trc_8f21b0 | 0.61 | 184d | gated |
| 14:02:46.902 | claims_sop_2026trc_8f21c4 | 0.89 | 6h | passed |
| 14:02:49.334 | legacy_wiki_exporttrc_8f21d1 | 0.42 | 412d | gated |
| 14:02:52.771 | contract_addendum_11trc_8f21e6 | 0.97 | 31m | passed |
| 14:02:55.019 | kb_pricing_notestrc_8f21f2 | 0.73 | 96d | gated |
Built for
Why now
Silent RAG failures stop being acceptable in production
Thin or stale retrieval that still answers confidently was tolerable in pilots. In production and under audit, that is the failure mode buyers are paid to prevent.
01
RAG sits on the answer path
Assistants quote rates, policies, and procedures to clients and staff. A stale chunk is a decision with liability attached, not an internal annoyance.
02
Audit needs a frozen decision record
Risk and compliance teams need a replayable record of why an answer shipped. Dashboards do not freeze the retrieval decision that produced it.
03
Scrutiny of generative systems is rising
EU AI Act-style transparency rules and sector exams raise the bar. If you cannot show when the model should have abstained, you own that failure.
How it works
Most RAG stacks never say when to doubt the answer
A retrieval can be close, stale, or thin, and the model still answers with full confidence. Ravel scores the evidence first, then gates the decision.
Retrieve & score
Ravel sits on the retrieval path and scores each chunk for semantic confidence and freshness against your source-of-truth policy.
confidence · freshness · provenance
Gate the response
Below-threshold or stale evidence is held, sent to human review, or answered with an explicit abstention.
pass · gate · escalate
Write to the ledger
The decision, inputs, thresholds, and outcome are appended to an audit log you can export and replay months later.
append-only · exportable · replayable
Live demo
Run the trust gate on a real request
Pick a scenario or edit the retrieval. Ravel scores this submission and returns a pass or gate decision with rationale.
gate_input
gate_result
Ready to evaluate
Run the gate on the context and question to the left. You get a confidence score, freshness check, release decision, and rationale for this submission.
Capabilities
Controls that decide whether an answer ships
Three runtime controls first. Expand for routing, policy versioning, and decision replay.
Confidence gating
Per-route thresholds decide when retrieval is strong enough to answer, and when the model must abstain or hand off.
Auditable trust ledger
Append-only records of retrieval, score, threshold, and outcome. The evidence trail risk teams request in review.
Staleness detection
Freshness windows per source. A two-year-old rate sheet cannot be quoted as current policy.
Compliance
An auditable ledger for every RAG output
When a regulator or internal reviewer asks why your assistant said something eight months ago, you should open a record, not rebuild the story from logs.
Written once, never edited
Entries are appended once and never edited in place. Cryptographic hash-chaining ships after pilots prove demand.
Provenance on every decision
Source identifiers, scores, freshness, and the thresholds in force at the time are captured with the outcome.
Evidence packs for review
Export a scoped range for internal audit, risk review, or an external examiner.
ledger_entry · sealed
- entry
- 00048112
- trace_id
- trc_8f21b0
- question
- commercial loan rate, 24mo term
- source
- rate_sheet_q3
- confidence
- 0.61
- freshness
- 184d (window 45d)
- policy
- finance.quote@v14
- decision
- GATED → reviewer_queue
- reviewer
- assigned 14:02:44Z
- prev_hash
- 9c4a…e17b
- entry_hash
- f20d…41c8
chain verified · 48,112 entries
Who this is for
What we can stand behind before launch
Ravel is pre-launch, so there is no customer logo wall. Today we can state the ICP, the evidence model, and how early teams engage.
Built for regulated RAG
Finance, healthcare, legal, and insurance knowledge assistants, where a wrong citation costs more than a bad UX score.
Design-partner waitlist
Pre-launch. Paid pilots at $1,000/mo for teams that give feedback and case studies before enterprise ACV.
Audit log evidence, not charts
Pass/gate decisions written to an append-only log you can export. Hash-chaining comes after pilots prove demand.
Enterprise path when proven
Private/in-VPC and formal compliance packs appear with enterprise accounts that need them — not day-one for every pilot.
Compare
Where freshness-first gating sits next to eval tools
Evaluation and observability platforms are strongest offline. Ravel decides pass or gate inline on every request and keeps the evidence.
- Yes— Supported as a core capability
- Partial— Possible with workarounds or limits
- No— Not part of the product
| Capability | RavelRuntime trust layer | GalileoEvaluation platform | Arize AIML observability | Bedrock GuardrailsOutput filtering |
|---|---|---|---|---|
| Runtime confidence gatingBlocks or escalates a response inline, before it reaches the user. | Yes | Partial | No | Partial |
| Freshness-aware scoringTreats stale evidence as a first-class failure condition. | Yes | No | Partial | No |
| Immutable decision ledgerHash-chained, append-only record of every gating decision. | Yes | No | No | No |
Use cases
Where weak retrieval shows up first
In regulated RAG the pattern is consistent: score the evidence, hold what is weak or stale, keep the record.
Finance
Finance copilots
Assistants that cite filings, rate sheets, and internal memos need a freshness check before a number reaches a client.
Gated: rate sheet 184 days past its freshness window.
Healthcare
Healthcare knowledge assistants
Clinical and operational knowledge bases change often. Ravel abstains when the retrieved protocol has been superseded.
Gated: protocol superseded by 2026 revision.
Legal
Legal & compliance RAG
Contract and obligation answers need provenance. Each release includes which clause was retrieved and why it passed.
Gated: confidence 0.42 below review threshold 0.75.
Pre-launchVerticals shown are the environments Ravel is built for, not customer claims.
Pricing
Design-partner pricing, then enterprise
Paid pilots at $1,000/mo. Convert at $1,000–2,500/mo. Enterprise at $3,500/mo ($42k ACV) after case studies. Every path starts on the same waitlist.
Design partner
Now open$1,000/ month
Paid pilot · steep discount for feedback + case study
For finance and healthcare teams piloting runtime confidence gating. Temporary pilot rate with a defined end date — not a permanent list price.
- Confidence + freshness gating on the request path
- Operational mode surfaced on every gate
- Append-only audit log with export
- Weekly founder feedback loop
- Scoped to what your RAG stack needs
Paid
$1,000–2,500/ month
Negotiated after pilot · by validated value
Conversion band once the pilot proves gate quality and audit usefulness. Priced per account size and volume — not a public SKU.
- Continued runtime gating in production
- Audit log retention and export
- Threshold and freshness policy tuning
- Direct founder support
- Clear path to enterprise ACV
Enterprise
$3,500/ month
$42k ACV · after case studies exist
Scaled enterprise tier once 2–3 paid case studies exist. Private/in-VPC and formal compliance packs when the account requires them.
- Production SLAs and volume commitments
- SSO / SCIM when needed
- Private or in-VPC deployment options
- Evidence packs for audit and risk review
- Named contact once sales capacity exists
trace_cost · estimator
Which band fits
A trace is one gated retrieval decision (score plus audit-log write). Pricing is phased: $1,000/mo pilot → paid conversion → $3,500/mo enterprise. Set a monthly volume to see which band applies.
Suggested band
Design-partner pilot
$1,000/mo
At 50,000 traces/mo, start on a paid design-partner pilot. Temporary discount with a defined end date in exchange for feedback and a case study.
Pilot price is not the forever list price — conversion band is $1,000–2,500/mo.
Pilot pricing is temporary and written into the agreement with an end date. Capacity is capped so founder support stays high. No charge until you confirm a start date.
FAQ
Common questions
Ravel is in a solo bootstrap phase: narrow MVP, then a small number of paid design-partner pilots. Capacity is capped (about three concurrent pilots) so support stays high. Joining the waitlist signals interest; it is not a bill. We do not charge until you confirm a pilot start date.
Waitlist
Join the design-partner waitlist
Paid pilots at $1,000/mo for finance and healthcare teams that want confidence gating and an auditable log. Steep discount, defined pilot window. No charge until you confirm a start date.