PROJECT FACTS
CLIENT
Deino (two-founder, pre-seed)
SECTOR
Auditable AI infrastructure for regulated finance
FACTA SERVICES
Fractional CAIO · CTO advisory · Rust engineering · Premortem facilitation
TIMELINE
4 months (thesis → POC + Gate 1 green)
TEAM
2 founders + 3 FACTA engineers
STATUS
POC shipped · Gate 1 (replay green 30 days) passed · Founding Pilot open
Executive summary
Provenance-grade infrastructure, gated by premortem discipline.
Deino arrived at FACTA as a two-founder thesis with an uncomfortable honesty problem: the vision was a five-sided reality oracle (citizens, verified actors, intel subscribers, market prediction, investment funds), but the most likely death was shipping all five sides at once. The premortem — run in three passes — named dispersion of focus, conflation of provenance with truth, and a calibration loop that might never beat baseline as the three co-fatal failures. The fix wasn't more features. It was sequence, gates, and firewalls.
FACTA took the role of end-to-end technical partner: fractional CAIO direction to collapse the two-identity problem into one sequenced narrative, CTO advisory to architect the Rust provenance core, and engineering delivery of 17 core + 6 GNL crates — including the BR-842 hash-chained audit manifest, the 4-layer GNL world model, a 5-sister-LLM + 23-agent A2A CRDT swarm, and the deino-eval calibration loop. The POC passed Gate 1 (bit-for-bit replay green for 30 consecutive days) on schedule. The five sides stay frozen behind Gate 1b until calibration beats a disclosed baseline.
OUTCOMES
Rust crates delivered
23
17 core + 6 GNL, inert-until-calibration
BR-842 replay
30 days green
Bit-for-bit, the existential gate
Swarm convergence
23 agents
A2A CRDT, no central orchestrator
Premortem fixes applied
11
Website + one-pager + sequencing
The client
Deino — the provenance thesis for regulated AI.
Deino is a pre-seed, two-founder company built on a premise that's uncomfortable for the AI ecosystem: regulated firms are being told to ship AI they can't defend. Models make decisions with legal effect, and when a supervisor asks how a specific output was reached, the answer takes weeks of forensics — if it can be answered at all. Deino's thesis is that provenance-grade infrastructure, not another model, is what closes that gap. The same auditable core is designed to expand into a five-sided reality oracle, but only after the wedge earns the right to live.
The challenge
Build a provenance core that survives its own premortem.
The scope was not 'build an AI product'. It was 'build the auditable layer under regulated AI, in a way that survives every plausible death we can name in advance.' FACTA ran three premortem sessions using Gary Klein's prospective hindsight method. The first named vanity-traction and an unprecise forecast as the most-likely and most-dangerous failures of the original plan. The second — across the full portfolio — surfaced a structural finding: Deino was entering every pitch with two incompatible identities (sober B2B infrastructure vs. viral consumer social network), and no single investor archetype could lead the round. The third tracked residual risk after the eleven fixes were applied.
CONSTRAINTS
Provenance ≠ truth
The manifest attests provenance and bit-for-bit reproducibility — never ground-truth correctness. A certified statement is not a verified-true statement. This had to survive every page and every output.
Determinism firewall
BR-842 replay had to stay bit-for-bit green even as non-deterministic consumer surfaces merged into the same audited ledger. Any replay-hash flip within 48h of a consumer merge would block the merge.
Liability firewall
Side 2 (verified actors) is the most exposed to defamation, GDPR Art. 22 and DSA risk. Confidential had to be the default for contested affirmations, with a written external-counsel opinion required before any consumer surface shipped.
Calibration that beats baseline
Gate 1b required a gnl_shadow_run cohort to show signal_calibrated = true against a forecastable series, with a named owner, a disclosed baseline and a dated stop-work trigger. If it didn't beat baseline, the forecast / oracle / intel / fund sides would be cut, not shipped.
Two-founder bus-factor
The sequencing had to be survivable with two founders. Sides 3–5 were frozen; the second architect engineer was prioritized before the first close.
One beachhead market
EU AI Act-exposed compliance and risk teams in regulated finance. One buyer archetype. One deep-tech / applied-AI infrastructure investor narrative. No 'trusted by' logos, no invented social proof.
FACTA's approach
Triple role: direction, architecture, execution — plus premortem facilitation.
FACTA came in as the technical partner that could simultaneously hold the product thesis, the Rust architecture, and the engineering delivery — and run the premortem that kept the team from building six half-built surfaces:
01
Direction — Fractional CAIO
- ·Collapsed two identities into one sequenced thesis: wedge → flywheel → expansion
- ·Defined the beachhead (EU AI Act-exposed compliance / risk teams) and the single investor archetype
- ·Ran the three premortem passes and mapped every named failure to a concrete fix
- ·Authored the honesty invariant: pre-seed, pre-revenue, sides INTENDED, certified ≠ true
02
Architecture — CTO advisory
- ·Hexagonal Rust core with tokio + axum + sqlx
- ·BR-842 hash-chained, append-only audit manifest with bit-for-bit replay
- ·Versioned 4-layer GNL world model (Instruments · Governments · Companies · Actors, public signals only)
- ·Determinism and liability firewalls; provider-agnostic provenance via BYO LLM keys
03
Delivery — Engineering team leadership
- ·Led the build of 17 core + 6 GNL Rust crates (inert-until-calibration)
- ·Implemented the 5-sister-LLM + 23-agent A2A CRDT swarm with no central orchestrator
- ·Shipped deino-eval (Brier / log-loss) and the gnl_shadow_run calibration cohort
- ·Built the six POC products over the shared core: Living Map, Path Queries, Risk Maps, Custom Axes, Backtest Forensics, Drift Alerts
The technical solution
One auditable core. Six products over it. Five sides gated behind it.
Deino's architecture organizes everything around a single auditable artifact: the per-output, hash-chained manifest (BR-842). The same manifest certifies a regulated B2B output, a public reality-affirmation and a confidential actor record. Six commercial products sit on top of one shared core; the five-sided super-app stays frozen behind the gates.
MODULES
GNL WORLD MODEL
Versioned 4-layer ontology (Instruments · Governments · Companies · Actors), public signals only, every node addressable and replayable by version.
AUDIT MANIFEST (BR-842)
Per-output hash-chained, append-only ledger of agents, data versions and model versions. Bit-for-bit replayable on demand, years later, even after models are swapped.
SISTER-LLM SWARM
Five sister-LLMs and a 23-agent A2A CRDT swarm that converge with no central orchestrator. Depth, not a wrapper.
CALIBRATION LOOP
deino-eval scores Brier / log-loss against a disclosed baseline. Gate 1b: beat the baseline on a forecastable series or cut the forecast sides.
SEQUENCING & GATES
Gate 1 (replay green 30 days) and Gate 1b (beat baseline + one signed paid pilot). Firewalls keep determinism and liability from contaminating the core.
FOUNDING PILOT PIPELINE
One CTA — Apply for the Founding Pilot — no 'start free'. North Star is a paid pilot before month 6, not vanity signups.
STACK
Pre-seed · Pre-revenue · POC stage · Certified provenance ≠ certified truth
REPOSITORY
github.com/facta/deinoResults
Four months. POC shipped. Gate 1 green. Five sides still frozen.
Deino went from a two-founder thesis to a Rust POC with the existential gate passed: BR-842 bit-for-bit replay stayed green for 30 consecutive days. The six POC products were built over the shared core and left inert-until-calibration, exactly as the sequencing demanded. The five-sided reality oracle stayed gated behind Gate 1b — which is the result the premortem said would keep the company alive.
METRICS
“FACTA ran the premortem that named every plausible death, then architected the Rust core and the gates that made the plan survivable. The hardest call was to freeze four of the five sides — and it was the right one. We passed Gate 1 because the manifest was built to be replayed, not pitched.”
Founder
Deino
Replicable lessons
What makes a premortem-driven build survive.
The fix isn't more sides — it's sequencing them.
The premortem's structural finding was that two incompatible identities kill the round. One sequenced thesis to one investor archetype (wedge → flywheel → expansion) was the only fix, and it cost zero engineering.
Certified ≠ True must survive every edit.
The manifest attests provenance and reproducibility, never truth. Putting that on every page that touches a forecast is the only defense against the defamation / GDPR / DSA event the premortem named as the most dangerous portfolio death.
Calibration that beats baseline is co-fatal.
Without Gate 1b (named owner, disclosed baseline, dated stop-work trigger), the platform collapses into an audit log. The fix is a gate, not a feature.
Pre-seed + BYO keys is a trust signal, not a weakness.
Regulated procurement fears lock-in and acquired-and-shut-down above all. BYO LLM keys and customer-owned audit chains remove both — and disclosing the stage openly is what an audit vendor is supposed to do.
About FACTA
Fractional AI leadership with real technical execution.
FACTA is a consultancy specialized in fractional Chief AI Officer and CTO roles for companies that need to lead the transition to AI without hiring a full-time executive team. Unlike generalist consultants who deliver PowerPoint, FACTA delivers systems in production.
We combine ten years of experience in AI architecture with operational expertise in multi-agent systems, Graph RAG, A2A, Rust engineering and premortem-driven planning. We work with post-seed startups through Series B scale-ups across Spanish- and English-speaking regions.
SERVICES
01 · Fractional CAIO
C-suite-level AI direction. Product thesis, roadmap, investor relations, beachhead selection, premortem facilitation.
02 · Fractional CTO
Architecture, stack decisions, technical hiring, engineering oversight, governance and platform trade-offs.
03 · Rust full-stack engineering
Implementation with an in-house team: Rust, multi-agent swarms, audit manifests, calibration loops, on-prem infrastructure.
04 · Premortem facilitation
Gary Klein prospective hindsight, run in multiple passes, mapped to concrete fixes that survive every edit.
Building auditable AI?
If you're shipping AI under model-risk regulation and need provenance, replay and a partner who can run the premortem — let's talk.
Book a 30-minute callhola@facta.dev