Inspiration
While building an institutional research system, I ran into a failure more subtle than hallucination. A number could be real. A reconstruction could be mathematically correct. Yet an AI could still promote that valid fact into authority it had never earned: strategy causality, repeatability, capital, publication, release, or closure.
That led to one question:
At the declared decision time, under this exact policy version, which authority stage did the evidence actually earn?
I built AsOf because high-stakes AI should be free to reason continuously without being allowed to talk its way across missing institutional receipts.
What it does
AsOf is evidence promotion CI for high-stakes AI. It decomposes a declared agent output into atomic claims, evaluates evidence available at decision time, runs deterministic source and authority gates, advances an ordered promotion graph only through earned predecessor receipts, and issues a content-addressed, replayable, policy-bound receipt.
The signature synthetic finance case begins with a fact that passes: a broker receipt proves a $1.2 million account gain. AsOf preserves it. It separately blocks the claims that the named strategy caused the gain, that its edge is validated forward, and that it deserves five times more capital.
The profit is real. The promotion is not.
AsOf returns four explicit outcomes with a fixed conservative lattice: BLOCK > INSUFFICIENT_EVIDENCE > REVIEW > PASS, plus claim-local reason codes and smallest-repair guidance.
How it works
The React workbench shows the original output, per-claim verdicts, the decision-time evidence clock, the exact authority path, deterministic gate failures, GPT-5.6 challenges, and separate hash-consistency and canonical-replay checks.
The TypeScript kernel is the policy authority. Seven versioned executable adapters define domain-local claim classes, evidence classes, exact authorities, predecessors, and target stages for finance, software, policy, biotech, security, science, and enterprise analytics. A manifest may select an installed adapter; it cannot weaken a requirement, truncate the path, or move the target backward.
A successor stage must link to the exact receipt that earned its predecessor. Each claim binds an adapter-owned required stage, the claim coverage must account for the complete output, and each evidence item declares its freshness and source health.
The 1.2.0 run identity binds the complete execution manifest, trusted adapter policy, and kernel semantics. The authority checksum binds deterministic state. A separate envelope checksum also binds timing, replay, and model provenance.
GPT-5.6 is an adversarial reviewer, not a judge. Through the OpenAI Responses API and strict structured output, it identifies assumptions, contradictions, authority leaps, and falsifiers. Its schema has no verdict field. Unknown, unrelated, or stage-less references are preserved in a signed quarantine ledger and excluded from admitted findings. The server reruns canonical deterministic authority, and the browser rejects any changed verdict, gate, stage, run identity, or authority checksum.
Why it matters
Authority teleportation appears anywhere AI output crosses an institutional boundary:
Finance: booked outcome -> attribution -> forward validation -> capital Software: build identity -> release validation Policy: decision-era rule -> effective-rule resolution -> authorization Biotech: database lock -> reviewed publication Security: triage -> remediation execution -> environment validation -> closure Science: preregistration -> confirmatory result -> replication Enterprise: period definition -> close validation -> quarter publication
The authority architecture transfers. Each domain keeps its own clocks, evidence classes, authorities, and falsifiers.
How I used Codex and GPT-5.6
Codex was my engineering collaborator. It helped turn a three-ledger research insight into typed contracts, an earned-receipt state machine, versioned adapter policies, anti-weakening checks, hostile tests, the bounded GPT-5.6 interface, the workbench, deployment, and the demo film.
Codex also helped falsify my first demo framing. Broker-booked P&L can be valid even when strategy attribution is not. That correction became the matched control at the center of the product.
The key human decisions stayed explicit: use a clean-room synthetic project, preserve valid facts instead of reflexively blocking them, keep model confidence outside authority, make policy versioned and executable, and state every limitation plainly.
GPT-5.6 attacks the claim graph at runtime. On the public finance control it surfaced grounded assumptions, contradictions, authority leaps, and falsifiers. The deterministic Block verdict, run identity, authority checksum, and canonical replay remained unchanged.
What I built and verified
Seven synthetic scenarios backed by seven executable trusted adapters Eight deterministic claim gates plus earned-predecessor promotion checks Full execution, policy, and kernel digests bound into a 1.2.0 run identity Deterministic replay pinned to run identity and authority checksum 74 passing graph, poisoning, tamper, adapter, and model-boundary tests A production TypeScript/Vite build with zero known dependency vulnerabilities A same-origin OpenAI boundary with strict schemas, limits, timeouts, bounded cache, coalescing, concurrency, and quotas An always-on public judge workbench with a no-key deterministic fallback A live GPT-5.6 review whose receipt passed hash consistency and canonical replay A fresh-clone rehearsal that passed install, test, build, and replay
Limits
This contest build evaluates canonical synthetic manifests. It does not yet ingest arbitrary answers, automatically extract claims, authenticate upstream sources, externally sign receipts, or certify production safety in any domain. Its unkeyed checksums establish content consistency, not issuer authenticity. The contest cache and quotas are process-local rather than distributed.
The next production steps are signed source resolvers, durable transparency logs, governed adapter publication and key rotation, untrusted arbitrary-input claim proposals, explicit source-route exhaustion, and domain-specific pilot suites.
Try it
Open the hosted workbench with no account:
https://asof-decision-ci.onrender.com
Start with Finance / Profit-to-capital leap, then switch to Software / Defensible release for the positive control.
Source and setup:
Built With
- codex
- gpt-5.6
- openai-responses-api
- react
- render
- typescript
- vite
- vitest
- zod
Log in or sign up for Devpost to join the conversation.