Journal of Intelligence, Technology & Human Progress
Founding Edition · Issue 001
How AI is moving from answering questions to acting in the world — and why evidence, permissions, recovery and verification become part of the product.
Issue thesis
Models once mostly returned text. Agentic systems can now participate in workflows, operate tools and make changes that persist. That shifts the center of gravity from output quality alone toward state, causality, permission, recovery, timing and proof.
actor → action → state transition → invariant → evidence → verificationReading map
Published · 6 September 2026 · EN / RU / 中文. Four repository roles, six reproduced seed classifications and a proposed model experiment that separates factual support, action admission and observed effects.
Read in English → · Русский → · 中文 →Published · 5 September 2026 · EN / RU / 中文. A fresh synthetic snapshot retains the price but reduces capacity to 600 USDT for a 1,000 USDT route. Seven cases and twelve focused tests show where a boundary recheck stops admission and where later state changes remain unprotected.
Read in English → · Русский → · 中文 →Published. The evidence that AI is moving from answers to actions — and why verification must move from outputs to trajectories.
Read flagship analysis →What happens when software can initiate work, transactions and coordination.
Published. A proposed verification model connecting state, causality, phase, transition, time, recovery and evidence across agent actions.
Read the framework →Published · experimental protocol. OBSERVE → VERIFY → AUTHORIZE → BIND → COMPARE → COMMIT → RECONCILE → PROVE. A framework-agnostic protocol for keeping trust guarantees connected across consequential transitions.
Read the protocol →Published + benchmark. Eight structural failure classes, real failure patterns, and a 16-case reproducible test seed.
Read the research →Published · 95/100 structural baseline. A pinned SDK revision was tested through 597 upstream tests across seven suites.
Open verified report →Published · 8/10 containment. Host filesystem, workspace path and Docker control-plane boundaries held; an ephemeral local peer remained reachable over the Docker bridge.
Open containment report →Published · 10/10 recovery protocol. Blind retry after a synthetic timeout-after-commit produced two side effects; reconcile-before-retry preserved exactly one effect.
Open recovery report →Published · 10/10 uncertainty-preservation protocol. When reconciliation itself timed out or returned stale UNKNOWN, treating uncertainty as retry permission duplicated the effect.
Open ambiguous reconciliation report →Published · 10/10 evidence-conflict protocol. A fresh primary COMMITTED record and conflicting replica ABSENT record caused a duplicate under unsafe arbitration.
Open conflicting evidence report →Published · 10/10 authority-verification protocol. A forged ABSENT record claimed primary authority and caused a duplicate when trusted without verification.
Open evidence authority report →Published · 10/10 authority-lifecycle protocol. A mathematically valid, still-fresh old signature caused a duplicate when revocation was ignored; decision-time key lifecycle checks accepted only the current rotated authority.
Open revoked authority report →Published · 10/10 trust-state freshness protocol. Fresh signed evidence evaluated against a 600-second-old registry snapshot reproduced a duplicate; refresh exposed revocation, while refresh failure preserved TRUST_UNKNOWN.
Open stale trust registry report →Published · 10/10 execution-bound trust protocol. A key was ACTIVE at verification time, then revoked before commit. Reusing the old decision duplicated the effect; trust-epoch binding or commit-time re-verification blocked the race.
Open revocation race report →Published · 10/10 atomic state-version protocol. Two nodes read the same ABSENT/version=100 snapshot. Split check/write produced two effects; compare-and-transition with an expected version produced one winner and one safe conflict.
Open distributed commit race report →Published · 10/10 compounded-hazard protocol. Timeout-after-commit, stale evidence, revoked authority, stale trust and a competing writer were composed in one trajectory. Unsafe handling produced three effects; TTP v1.0 preserved one and exercised all eight stages.
Open TTP end-to-end report →Published · 10/10 real database adapter. Two independent PostgreSQL connections read the same ABSENT/version=100 snapshot. Unconditional writes produced two effects; a version-bound conditional UPDATE allowed one winner, rejected the stale writer and preserved one effect.
Open PostgreSQL adapter report →Published · 10/10 conflict-signal matrix. The same TTP race produced a zero-row precondition failure under READ COMMITTED and SQLSTATE 40001 serialization failures under REPEATABLE READ and SERIALIZABLE. Fresh reconciliation preserved one effect in every safe path.
Open isolation-level matrix →Published · 10/10 external-effect boundary. One DB transition became two external effects when ACK loss was followed by a new request identity. Transactional outbox, stable idempotency and external reconciliation preserved one effect.
Open transactional outbox report →Published · 10/10 real network boundary. A separate HTTP container committed an effect then dropped the response. A new retry identity duplicated it; same-key redelivery was deduplicated and GET reconciliation prevented a second POST.
Open HTTP idempotency report →Published · 10/10 restart-memory boundary. A remote effect survived real container replacement while process-local dedupe memory disappeared. Same-key replay then duplicated the effect; durable idempotency state or authoritative reconciliation preserved one effect.
Open restart durability report →Published · 10/10 time-memory boundary. With a 120-second delayed recovery, a 60-second dedupe TTL expired while the remote effect remained durable; the same key was applied again. A 300-second retention window deduplicated the same replay, while authoritative reconciliation after expiry avoided replay entirely.
Open idempotency TTL report →Published · 10/10 clock-authority boundary. The same retention record was ACTIVE for Node A and EXPIRED for Node B while the original effect remained committed. Node-local replay duplicated the effect; a declared authority clock deduplicated it, while a skew window crossing expiry preserved TIME_UNKNOWN and reconciled without replay.
Open clock-skew report →Published · 10/10 monotonic-history boundary. The same durable record visibly moved ACTIVE → EXPIRED → ACTIVE after wall-clock rollback. After expiry cleanup advanced a temporal epoch, an un-fenced same-key replay produced two effects; stale-epoch fencing returned HTTP 409 and a monotonic time watermark preserved one effect.
Open Temporal ABA report →How model capability increasingly depends on the infrastructure beneath it.
Why agent safety expands into tools, credentials, environments and persistent state.
When hypotheses get cheaper, verification, reproducibility and attribution matter more.
Designing collaboration so automation increases human agency rather than obscuring it.
Reports #003–#010 contributed separate executable invariants; TTP v1.0 composes them into one path, #011 exercises it end-to-end, #012–#013 move it into real PostgreSQL, #014 crosses the database/external-effect boundary, #015 makes transport ambiguity real over HTTP, #016 tests whether execution memory survives service restart, #017 tests whether that surviving memory remains valid long enough for delayed recovery, #018 tests who has authority to decide temporal validity when clocks disagree, and #019 tests whether rollback can resurrect an already-retired temporal state.
Beyond success rate: invariant violations, evidence completeness, authority lifecycle, state/trust versions, database conflict signals, outbox lag, redelivery count, idempotency dedupe rate, transport failures, service restart identity, dedupe-memory durability, retention/expiry, node clock disagreement, declared clock authority, skew bounds, temporal-uncertainty decisions, monotonic temporal epochs, fencing outcomes, time watermarks, remote reconciliation latency and proof completeness.
A practical map of verification, identity, atomic state transitions, transactional outbox, stable effect identity, transport evidence, durable execution memory, time-bounded retention, clock authority / skew handling, monotonic temporal epochs, fencing, remote reconciliation and proof infrastructure.
Editorial standard
Issue 001 moves through analysis → framework → benchmark → recovery → evidence authority → authority lifecycle → trust freshness → execution binding → atomic transition → Transactional Trust Protocol v1.0 → PostgreSQL adapter → isolation-level conflict matrix → transactional outbox → real HTTP boundary → durable execution memory → idempotency retention → clock authority and temporal uncertainty → Temporal ABA and monotonic execution history. Report #019 adds the anti-resurrection rule: a previously crossed temporal boundary must not become valid again solely because a wall-clock value moves backward. Recovery must bind consequential time decisions to monotonic history or reconcile before execution.
Evidence model
Define the signal and time boundary.
Link claims to evidence whose source, authority, freshness, provenance, state/trust versions, execution preconditions, database conflicts, delivery identity, transport outcome, service/restart identity, safety-memory durability, retention/expiry window, node-local times, clock authority/skew model, temporal epoch or watermark, fencing result, remote reconciliation path and final proof can be inspected.
Distinguish mechanism from correlation and narrative.
State the limits, conflicts and competing explanations in the evidence.
Provide a path another reader can inspect or reproduce.
Translate evidence into implications without pretending prediction is fact.
Search / Archive
Canonical crawlable index of RESONANCE web articles. Each entry resolves to the same URL used in sitemap.xml and structured data.