Field notes
from the oracle.
Engineering deep-dives, regulatory analysis, and customer stories — written by the founders and the Oracle Council. One monthly digest, no filler.
Why we built audit manifests in Rust, not Python.
The non-obvious choice that made bit-for-bit reproducibility actually possible — a deep dive into determinism, type safety, and what regulated AI demands of its substrate.
A world model makes the fork. Calibration makes it a signal.
Just-in-time world modeling is about generating the shape of the future when it is needed. The fork is necessary. The signal is earned.
EngineeringThe audit ledger is the schema migration you never write
Every schema change eventually runs two versions of the system against one database. The append-only answer is not a cleverer migration — it is a decision to never rewrite the past.
OracleSearch returns rankings. An ontology returns structure.
A neurosymbolic search engine claims a 2.7× accuracy win over e-commerce giants. The mechanism is the lesson: force the model to honor a graph, and you can audit what it did.
OracleCones are belief states that know how to update
A language model that compresses its own history is performing a belief update. It fails in exactly the way a forecast must not: information lost, and nobody charged for it.
CustomerSovereign AI procurement: the audit clause that matters
A government buyer writes reproducibility into the RFP. Anonymized, illustrative.
CustomerThe design partner shapes the POC
Building with, not for — the partner examiner defines the manifest. Anonymized, illustrative.
CustomerSupply chain disruption on a typed graph
A shock explained by the GNL layers, not a flat feed. Anonymized, illustrative.
CustomerAn energy utility forecasts load as a cone
Peak-load distribution for a grid that hedges. Anonymized, illustrative.
CustomerA fintech hands the regulator the manifest
A credit-decision audit that points instead of narrating. Anonymized, illustrative.
EngineeringWhy we built audit manifests in Rust, not Python.
The non-obvious choice that made bit-for-bit reproducibility actually possible — a deep dive into determinism, type safety, and what regulated AI demands of its substrate.
CustomerOnboarding into a regulator's sandbox
The sandbox as a calibration partner — the regulator sees the loop. Anonymized, illustrative.
CustomerTelco churn, as a defensible distribution
Retention modeled honestly — the shape survives a regulator question. Anonymized, illustrative.
CustomerA hedge fund trades the cone, not the point
An IC adopts distributions — the point estimate retires. Anonymized, illustrative.
EngineeringThe 4-layer Geo-Neural Layer explained
What each layer does, why we split them, and how versioning works across instruments, governments, companies, and actors.
CustomerA central bank's shadow-run calibration
A monetary-policy path cone, scored against the realized path. Anonymized, illustrative.
AuditEU AI Act Annex III: a practical compliance checklist
What Annex III transparency obligations actually require, mapped to specific manifest fields — not a law summary, an implementation list.
CustomerDrift alerts for an asset manager
A portfolio model that flags its own drift before the IC notices. Anonymized, illustrative.
CustomerCat bond backtest forensics, anonymized
A (re)insurer design partner reruns a cat-bond model against the realized season. Illustrative.
EngineeringSister-LLMs: ensemble reasoning over a shared ontology
How five independent LLMs produce a single calibrated probability distribution over a shared, versioned world model.
OracleWe ship cones, not forecasts
The product surface is a probability cone — the shape is the deliverable, not the number.
OracleThe reality-oracle thesis
The founding point of view, expanded — convert events into affirmable facts and real probabilities.
OracleThe auditable oracle category — why now
A convergence of decision intelligence and governed AI is opening a category that did not exist before. Why the timing is structural, not opportunistic.
OracleDEINO is a universal oracle, not a template
Product-neutral — the substrate is reusable, the product is not a fork.
OracleWhat an acquirer's diligence actually looks for
From the v2 audit — the questions a buyer asks that a manifest answers and a dashboard cannot.
OracleA signed paid pilot before any side scales
The gate that protects the claim — no accuracy posture before a paid pilot closes.
Engineering10× token reduction with deino-loom
Context curation via attention budgets that reduces tokens without losing fidelity — a PROJECTED number, and the mechanism behind it.
OracleThe 'AI truth engine' fails procurement
The question no honest system answers yes to — and why the auditable oracle answers the real one.
OracleA category is a property, not a feature list
Writable into a procurement requirement — that is the test of a category.
AuditHow to defend an AI decision to your regulator
What examiners actually look for during a model-risk review, and the documentation to hand them — field-mapped, not vague.
OracleOne beachhead, not a scatter of partners
Go-to-market discipline — one EU-AI-Act-exposed market where demand and regulation match.
OracleDecision intelligence, not descriptive dashboards
From watching to acting — the move from a number on a screen to a cone a CRO can hedge.
Oracle'What if you are wrong?' — the question a distribution answers
The board question that kills a point estimate, answered by a shape.
CustomerSix weeks of audit prep, in six hours
An anonymized walkthrough of a LATAM Tier-2 bank that rewired its BCRA audit cycle around audit manifests.
OracleEntropy is the honesty field
Spread as truth — the number that says how much the model disagrees with itself.
OracleReading a distribution: bimodal means two regimes
How a CRO reads shape — the modes are the contingency.
OracleCalibration earns the word 'probability'
A number without a loop is decoration — the loop is what makes it a probability.
OraclePoint predictions are a liability
Why a single number is the one thing a CRO cannot defend — and why a distribution is the thing they can.
OracleWhy we say 'forecast,' not 'predict'
Voice discipline as substance — the word is the claim.
Oracle'Oracle' means provenance, not omniscience
The metaphor earns its keep — consultable, shows its work, never infallible.
AuditVendor lock-in is a compliance risk
Provider-agnostic as an audit posture — the manifest survives the vendor.
AuditModel risk tiering: not every model gets the same scrutiny
Risk-based tiering done honestly — the tier decides the manifest, not the other way around.
AuditEU AI Act vs the UK approach: binding law vs pro-innovation
Two jurisdictions, two philosophies — and why the binding tier sets the floor either way.
AuditThe audit reads the substrate, not the slide
What is defensible is what is reconstructable — the slide is the story, the substrate is the evidence.
AuditRisk assessment and transparency are different obligations
Don’t conflate them — one is about the system, the other is about the output.
AuditWhat you do not claim: the refusal that makes the rest credible
Certified ≠ True — the refusal to certify truth is what certifies provenance.
AuditAccuracy is a loop, not a property
Calibration in progress against a disclosed baseline — the honest version of "is it accurate".
AuditThe override is logged, not inferred
Who changed what, when — sealed in the chain, not reconstructed from memory.
AuditState the scope: no examiner trusts a model that claims everything
Boundaries are credibility — the stated limit is the thing the examiner believes.
AuditExaminer windows are long; logs must mean the same thing months later
Durability of meaning — a field name is a contract with the future.
AuditWriting reproducibility into a procurement requirement
The category property is writable — "reproducible, tamper-evident, defensible" goes in an RFP.
AuditModel card vs per-output manifest
Per-output beats per-model for a decision under review — the unit of audit is the decision.
AuditHuman oversight is a design requirement, not a paragraph
Built into the loop, not bolted onto the docs — the override is a typed, recorded thing.
AuditThe Fundamental Rights Impact Assessment
The FRIA under the EU AI Act — the obligation that is most often skipped and most often asked for.
AuditSeparate the law from the guidance
The defensible posture names the force behind each line — binding, supervisory, voluntary.
AuditHand them the manifest, not a slide deck
Point, do not narrate — narration invites argument, a record answers.
AuditWhat an examiner actually looks for
Provenance, reproducibility, boundaries, oversight — the four-part question behind every review.
AuditThe replay anchor: handing a third party the hash
Reproducibility by a stranger — the one field that turns a claim into a check.
AuditTraceability across a model swap
Pin the model by hash, not by name — swap the live model, the chain still verifies.
AuditA log that outlives the system that wrote it
Examiner windows are long — meaning has to survive the upgrade that replaced the logger.
AuditBinding, guidance, voluntary: name the force behind each requirement
The three-tier discipline — the defensible posture never lumps them.
AuditISO/IEC 42001: the AI management system, honestly scoped
What a voluntary cert proves and what it does not — and why the line matters.
AuditNIST AI RMF: the voluntary tier done right
Documentation on top of binding law — useful precisely because it is not oversold.
AuditBCRA model-risk expectations: a LATAM reading
Argentina-specific model-risk pressure, and what a LATAM Tier-2 bank actually faces.
AuditSR 11-7 is guidance, not law — and why that distinction wins an audit
The supervisory guidance tier, and why calling it law reads as someone who does not know the difference.
AuditGDPR Article 22: the right to a human, not a number
Automated decisions and the explanation that follows — what Art. 22 actually reaches for.
EngineeringWhy the web is a static export on distroless
No shell, no Node runtime — the marketing surface is a static bundle on distroless/static.
EngineeringCompile-time-checked SQL, or: why .sqlx is committed
The offline cache lets CI build without a live DB — and why editing SQL means refreshing it.
Engineeringdeino-cli: bootstrapping an Eye Key locally
The new-key flow, and why local minting is the only honest bootstrap.
Engineeringdeino-sdk: the Eye Key is the developer surface
The SDK wraps the developer API — and the contract is one key, shown once.
EngineeringTimescale + pgvector in one image
Why the pg17 image bundles both — and what we avoid by not running two databases.
Engineeringdeino-courier: Brevo, or a logging no-op
Transactional email that fails safe — the no-op adapter is a first-class citizen.
Engineeringdeino-ingest: the signal pipeline
How a raw external feed becomes a typed NodeVersion.
EngineeringGzip frames, when the client asks
Bandwidth discipline on the WS — opt-in Binary frames for clients that say so.
EngineeringRealtime deltas, per tile
One broadcast channel per geohash — a client only receives its viewport.
EngineeringGeoSignal ids are deterministic
uuidv5(source, native_id) — re-ingest updates, never duplicates.
EngineeringAIS: the one streaming source
aisstream.io over WebSocket — the streaming exception, and why it stays an exception.
EngineeringA missing key self-disables
An unset env var returns Disabled, not an error — the platform never breaks on a missing feed.
EngineeringSources are data, not code
Add a feed by adding a SourceDescriptor — never touch the engine.
Engineeringdeino-atlas: a gateway, not a scraper
The Facade pattern — one poller per source, clients read the cache, upstream call volume is ours to bound.
EngineeringGate 1b: the calibration gate
A shadow run must beat a disclosed baseline on one series before any forecast side scales.
EngineeringGate 1: the reproducibility gate
The full-chain-walk — before any accuracy claim, the bytes have to match.
EngineeringShadow run vs ground truth: how calibration actually works
gnl_shadow_run against gnl_ground_truth — the loop that earns the word probability.
Engineeringdeino-eval: Brier and log-loss, not accuracy
The calibration harness scores against a realized timeline — accuracy is a loop, not a property.
EngineeringThe Eye Key is shown once
HMAC + fingerprint on disk, plaintext in memory only — INV-2, and why it is non-negotiable.
EngineeringThree auth regimes on one router
Eye-Key, user JWT, Atlas WS — one router, three contracts, three threat models.
Engineeringdeino-whisper: HMAC-signed webhooks that survive a retry storm
The delivery worker, the WhisperStore port, and why a webhook is an audit artifact.
Engineeringdeino-tide: SSE with backpressure and replay
Streaming cones without drowning the client — the stream is part of the manifest.
EngineeringL4 actors: public signals only, by design
The strictest governance layer is a choice, not a default — and it is load-bearing.
EngineeringOpenAI-compatible only: why we didn’t pick a direct Anthropic SDK
One interface, many providers — vLLM, OpenRouter, Together — and a deliberate non-dependency.
EngineeringEvery run stamps its prompt version
The ensemble that produced an output is reconstructable — the prompt is part of provenance.
EngineeringPersonalities are TOML, not code
Edit a sister without recompiling — the lens is data, the engine is fixed.
EngineeringRouting the globe with geohash prefixes
World Monitor tile keys computed in deino-core::geo — no PostGIS dependency.
EngineeringOne Postgres, no PostGIS: why the whole platform runs on one database
Ledger, events, embeddings, simulations, queue, geo cache — one pool, one migration story.
EngineeringBounded awakening: why per-query cost stays flat as the graph grows
The tractability property at planetary scale — a budgeted neighborhood, not the universe.
EngineeringCRDT convergence: how the swarm agrees without a coordinator
Conflict-free merged agent traces — disagreement becomes data, not a deadlock.
EngineeringA2A: 23 agents that converge without a conductor
Why the swarm uses Google A2A and CRDT merge, with no central orchestrator.
EngineeringBR-842: the audit manifest, field by field
What each field attests — and the one thing the manifest refuses to assert.
EngineeringAxisOverride: a typed, recorded override
Overrides are first-class values, not comments in a Slack thread.
EngineeringNodeVersion is immutable, by construction
Append-only, hash-linked: the world at any past moment is reconstructable.
EngineeringWhy we report entropy in nats
The unit choice is the honesty field — spread as a number a CRO can read.
EngineeringProbabilities normalize in exactly one place
The deino-oracle invariant (INV-1): sum approximately 1.0, sorted, entropy in nats — a guarantee, not a hope.
EngineeringSlice-owned ports: when a table belongs to one use-case
The deino-scry pattern — a private table gets a trait + adapter in its own crate.
EngineeringWhy domain crates depend on traits, never sqlx
The port rule (INV-5): tests swap mocks, business logic never touches a pool.
EngineeringDependencies point down: the rule that keeps the core pure
Why deino-core has zero I/O, and why every other crate is glad it does.
Behind the Architecture
Rust + GNL + A2A deep-dives. For platform engineers, CTOs, and curious builders.
Audit-First AI
EU AI Act, SEC, BCRA practical guides. For compliance officers and risk managers.
The Oracle Layer
Category creation, market analysis, why now. For investors and analysts.
Customer Stories
Anonymized case studies, audit walkthroughs, deployment patterns.
Subscribe to The Oracle Council.
Monthly: one founder note, regulatory updates, the engineering post of the month, a customer story. Two-minute read. No filler.
- What does the DEINO blog cover?
- Four areas: Rust + GNL + A2A architecture deep-dives, audit-first AI guides for the EU AI Act / SEC / BCRA, the oracle-layer thesis for investors and analysts, and anonymized customer stories.
- How often is the newsletter sent?
- The Oracle Council newsletter ships monthly — one founder note, regulatory updates, the engineering post of the month, and a customer story. A two-minute read, no filler.
- Who writes the posts?
- Posts are written by the DEINO founders and the Oracle Council. Engineering deep-dives are authored by the team building the platform.