Recension Research

Independent AI systems research and engineeringGreater Manchester

AI systems you can inspect, evaluate and control.

Recension Research builds open infrastructure for the parts of an agent system that are usually taken on faith: what the agent did, what the evidence supports, what it was permitted to do, and whether the record still holds. The same methods are available to teams as short, scoped engagements.

Fig. 1 — One run, twice: as behaviour, and as a record.Move the lens

1–3inspect, evaluate and control ] scripsi; trust codd.

recension, n. The critical revision of a text from a comparison of its surviving witnesses; the text so established.

Four open-source projects

Open research

Four instruments, one job each.

Each project takes one responsibility and states its limits. Together they are a division of responsibility, not a dependency graph: none grants another authority, and each can be used alone.

Make capability legible before making it larger.

The question behind all four: how do increasingly capable agent systems remain inspectable, evidential and governable, without pretending that one mechanism solves everything?

  1. TypeScriptApache-2.0Experimental

    Endophasia

    A local-first workbench and research toolkit for coding agents.

    Fig. EComplete frame
    Endophasia: plan, hand over, observe, check, replay Five scenes. First, a task becomes an attempt and a brief, with a plan that states the check to run and the output expected. Second, the brief reaches the agent only after an exact phrase is typed in the terminal; prompt, steer, queue and stop requests wait the same way. Third, Endophasia sits beside the Pi harness, whose loop runs context, model and tool; each step drops an observation onto a trace, and a capability the runtime cannot report is shown as unavailable. Fourth, the operator runs the check: the program prints 20 and exits 0, the receipt records a pass, and the ledger, reading the receipt against the plan, shows the criterion failed because 14 was expected. Fifth, a recording replays exactly, while a run in a changed workspace diverges. DEVELOP OBSERVE RESEARCH BEFORE THE WORK NOTHING SENDS ITSELF THE LIVE SESSION AFTER THE WORK RECORDED RUNS task attempt brief nothing sent PLAN · FROZEN WHEN PREPARED checkprecedence argvnode calc.mjs '2+3*4' expectexit 0 · stdout "14" criterion2+3*4 is 14 what correct looks like, stated before the work brief context model tool pi · its own loop › the exact phrase, typed APPROVED PROMPT STEER QUEUE STOP each, the same way a receipt says accepted · not that it worked approval governs dispatch · it is not containment context model tool HARNESS · PI endophasia · beside, never inside trace as the runtime reported it active-tool control · UNAVAILABLE what a runtime cannot report stays missing $ endo attempt check -- node calc.mjs '2+3*4' stdout 20 · exit 0 receipt · at 6712eae check passed · exit 0 LEDGER · READ AGAINST THE PLAN plannedprecedencefailed at HEAD behaviourexpected "14" · observed "20" criterion2+3*4 is 14a planned check failed the ledger reports · you decide a receipt is not acceptance recording replay new workspace = exact diverged 40 of 40 replays exact recorded model responses held fixed the changed workspace diverged where expected
    Select a step to hold it
    Fig. EThe exit status said the check passed. The plan said what the command should print, so the ledger says the criterion failed. The example is the project's own five-minute tour.

    A coding agent reports that the work is done, and the tests exit 0. Endophasia answers from records, not from the report. You state what a correct result looks like before the work begins, hand the brief to the agent, run the checks yourself, and read one ledger of what is established and what is not. It attaches beside the harness you already run (Pi, today) and never installs into it, patches it or replaces it.

    Use it when
    You need to know whether an agent's work is what it claims to be, why two runs differed, or what a change to a prompt, model or harness does before it ships.
    Works today
    Tasks, attempts and briefs; operator-run checks with receipts, verification plans, an acceptance ledger and working-tree provenance. A live Pi workbench in which every prompt, steer, queue and stop is approved in the terminal. Recording, exact replay, trajectory comparison, and controlled and sandboxed experiments. Build from source: main is well ahead of npm 0.1.0-alpha.2.
    Pencilled in
    An operator interface for evaluation-driven adaptation, model-internal observation, and live control of harnesses other than Pi.
    Does not claim
    That relayed reasoning is a verified record of the model's computation, that a passing check is acceptance, or that terminal approval contains the tools an agent runs.
  2. RustApache-2.0Experimental

    Magpie

    Replayable memory for AI systems. Conserve the log; derive the rest.

    Fig. MComplete frame
    Magpie: a signed log, and standing derived from it by a named policy On the left, a signed, append-only log holds a claim and a piece of supporting evidence. On the right, nothing is stored: the caller names policy v2, which reads direct support, and the claim's standing comes out as supported. A detached receipt for that answer is carried elsewhere to be checked. Refuting evidence is then appended to the log without overwriting anything; asked under policy v4, which reads direct refutation, the same record gives the standing refuted. Finally the derived views are deleted and replayed from the log, with the same result. signed log · append-only derived, on request #1 claim c17 #2 evidencesupports #3 evidencerefutes nothing overwritten replayed from #1 signed · hash-chained claim c17 policy v2 · support policy v4 · refutation standing SUPPORTED REFUTED elsewhere receipt binds history · key · policy no latest · caller chooses views deleted · replayed same history · same result
    Select a step to hold it
    Fig. MThe log holds what was recorded. Standing is the output of a policy the caller names, and a verified history is not thereby a current one.

    Magpie keeps a signed, append-only history of what was recorded. Claims, evidence, provenance and conclusions are derived from that history, and a conclusion (a claim's standing) is always the output of a policy the caller names. A remembered claim therefore keeps the distinction between what was said, what was checked and what may follow. Contradicting evidence is added to the record, never written over it.

    Use it when
    A system's memory or conclusions must be auditable later, whether by you, a customer or a regulator.
    Works today
    Source release v0.3.0. An Ed25519-signed hash chain with local SQLite persistence, typed evidence and provenance, explicit standing policies, deterministic replay and detached standing receipts. Independent verifiers in Rust, Python and Go.
    Does not claim
    Truth, key ownership, freshness or permission to act. Verification establishes only the predicate checked against the inputs supplied.
  3. RustApache-2.0Research prototype

    Deadbolt

    A permission engine for AI-operated systems. The model may suggest; authority stays outside the model.

    Fig. DComplete frame
    Deadbolt: proposals meet an authority boundary First, a comparison: a sandbox walls the agent off with a copy of the system, while Deadbolt lets it act on the real system through narrow typed routes. Then three proposals approach a checkpoint that applies route, policy, lease and confirmation. "Set rounding to 0" holds an exact lease, is confirmed locally, changes the real desktop and leaves a receipt. "Change everything" holds the same lease but is refused as too broad. "Delete files" has no route and is refused. Finally the receipt is used to restore the previous value. sandbox agent copy of system walled off: nothing real changes deadbolt agent narrow route real system real work, but only through typed routes route · policy · lease · confirm default: locked set rounding → 0 proposal LEASE real desktop prior value rounding: 0 restored change everything proposal LEASE delete files proposal confirmed granted receipt + rollback data refused too broad refused no route
    Select a step to hold it
    Fig. DEach lease opens one route, and every effect leaves a receipt. The live routes really are this small: an unremarkable desktop setting gives permission and recovery somewhere concrete to be worked out.

    Deadbolt does not hand a model a shell and then try to make the shell safe. Each operation is a typed route that fixes which request it accepts, which policy applies, what lease is required, whether a person must confirm, what is recorded and how the effect is undone. An agent can do real work on a real system without inheriting permission to do everything else.

    Use it when
    An agent must make real changes to configuration, infrastructure or files, and a general credential is not an acceptable answer.
    Works today
    Two live routes that change a real Hyprland desktop, each with a typed request, local policy, an exact signed lease, explicit local confirmation, a receipt and rollback. Read-only style planning, and a read-only MCP surface for review.
    Pencilled in
    Execution containment, and Home Manager routes.
    Does not claim
    General desktop automation, an agent shell, or that arbitrary computer use has been made safe. A valid lease or receipt grants nothing merely by existing.
  4. RustApache-2.0Experimental

    Cogitator

    A tamper-evident recorder and verifier for agent runs.

    Fig. CComplete frame
    Cogitator: run records folded into a witness root Five committed records from a run appear in order: two trace entries, two tool calls that ran, and one call that policy blocked before dispatch and that is recorded all the same. Each record is then encoded and folded in order into a hash chain that starts from an initial block over the run's metadata. The final link is the witness root, compared against a copy kept outside the bundle. If one record is edited afterwards, every link from that point changes and the root no longer matches. With timestamps and hostnames kept outside the seal, the same run on another machine produces the same root. committed records trace call call′ blocked trace call ran not run ran policy is applied before dispatch an allowed call runs, and is recorded a blocked call does not run, and is recorded a refusal is evidence too init root BLAKE3 chain · domain-separated outside copy match mismatch record edited after the run time · host outside the seal root · B another machine =
    Select a step to hold it
    Fig. CA matching root shows the record is intact. It does not show that the run took place: for that, keep a copy of the root somewhere outside the bundle.

    Cogitator records what an agent did, which tools it asked for, which calls ran and which calls policy stopped, then seals the committed record under a single fingerprint, the witness root. Another machine can recompute the root from the bundle and check that the record still matches. Machine details such as timestamps and hostnames are kept as provenance, outside the seal.

    Use it when
    You must show what an agent did, what it attempted and what it was prevented from doing, to someone who was not present.
    Works today
    Release v2.2.2. Reference recorder and verifier for the draft Cogitator Witness Protocol: canonical events, policy applied before dispatch, byte-stable replay with drift reports, and witness roots. Agents and tools are fixtures for now.
    Does not claim
    That a run took place, that the machine which produced it was trustworthy, or that an external model is deterministic.
Findings

Results, with their scope.

A result is only as useful as the conditions it was measured under, so each is printed with them. One square is one trial.

  1. Recorded sessions replayed exactly, all forty times, with the model's responses held fixed. A changed workspace diverged where expected.

    Endophasia · scopePi 1.0.1, four recorded sessions and a negative control.Study

  2. With the environment pinned and sampling fixed, repeated live runs agreed on every judged layer, on every task. The unpinned control did not.

    Endophasia · scopeThree small tasks, Pi 1.0.1 and qwen3.8-27b on one machine. Not a universal determinism result.Study

  3. A path that should not matter changed behaviour. The same steering message was followed at 27 of 30 working-directory paths on one task, and at 6 of 30 on another.

    Endophasia · scope360 trials across 30 paths. Acceptance and consumption were measured separately from following the instruction.Study

  4. On two tasks, most failures were a response cut off by a token cap. Doubling the cap raised success from 10 of 24 to 19 of 23.

    Endophasia · scope192 trials, four arms and four paths, on two selected tasks. The interval excludes zero only with all four paths included.Study

  5. One request, four different trajectories, the same answer. Attribution from agent traces is distributional, not causal.

    Endophasia · scopeA feasibility probe: two measurements on a real model.Study

  6. Three independent verifiers, written in Rust, Python and Go, agree exactly on every one of 432 frozen cases.

    Magpie · scopeOne corpus and one signature profile. Conformance evidence, not a proof of verifier correctness.Details

  7. A fixed 50-task regression suite has its witness root pinned in CI, so accidental protocol drift fails the build.

    Cogitator · scopeStubbed tools and a stubbed model. It catches protocol changes; it is not an agent benchmark.Details

Some raw recordings and governed execution receipts remain private, so the public summaries do not make every run reproducible from the repositories alone.

Method

Separate the jobs.

Observation, evidence, policy and permission are related. They are not interchangeable, and no single mechanism should do all four. The work rests on six distinctions that are easily run together.

observation ≠ inference
What a runtime reported is shown as reported. Anything derived names the method that derived it, and missing evidence stays missing.
evidence ≠ standing
Records support or contradict a claim. Whether the claim holds is the output of a named policy, chosen by the caller.
proposal ≠ permission
A model may propose an action. Authorisation is decided outside the model, for that action alone.
acceptance ≠ completion
A runtime accepting a request does not show that the request ran, or that it achieved what was intended.
evaluation ≠ promotion
A better score is a result. Replacing a running system is a separate decision, and needs its own authority.
simulation ≠ execution
A simulated outcome is labelled as simulated, and is never presented as a real run.
Engagements

The same discipline, applied to your system.

Short, scoped engagements for teams putting agents into production, who need to know what those agents can reach, what they did, and how either would be demonstrated.

Systems and assurance review

A structured review of an existing or planned agent system: its workflows and data boundaries, the case for hosted or private deployment, and where authority and failure concentrate.

  • Workflow and data-boundary mapping
  • Deployment architecture
  • Security and control recommendations

You receiveA written implementation plan, with risks in priority order.

Agent evaluation and control

Instrumentation for an agent you already run, reproducible tests for the ways it fails, and an account of what it is able to reach.

  • Adversarial evaluation scenarios
  • Runtime instrumentation
  • Permission and failure analysis

You receiveEvidence your team can inspect and rerun.

Private AI and agent pilots

A working prototype under your real operating constraints: local or private inference, bounded agent workflows, and the harness needed to evaluate them.

  • Open-weight and API model paths
  • Local and private inference
  • Bounded pilot deployment

You receiveA working pilot and a documented handover.

How an engagement runs

  1. A conversation

    The system, the constraint, and what a finished engagement would have to show.

  2. A written scope

    Deliverables and boundaries, agreed in writing before any work begins.

  3. The work, in the open

    Findings are written up as they are made, not held back for a final report.

  4. A clean handover

    Code, tests, evidence and documentation that your team can keep running.

If an agent performs well in demonstration and the open question is what happens after that, it is a good moment to talk.

Email George
About

Research that holds under deployment.

Recension Research is an independent AI systems research and engineering practice, founded by George Gallagher and based in Greater Manchester.

The practice works in two directions. Client engagements begin with deployment problems: privacy, reliability, auditability, permissions and evaluation. The open research builds reusable mechanisms for evidence, control and verification. Each disciplines the other. Real constraints decide what is worth building, and a mechanism is kept only if it survives one.

The projects are experimental research software. Claims are kept narrow, failure states are made visible, hostile cases are tested, and each system documents what it does not establish.

Current work

Now
Using Endophasia's development loop on real tasks and hardening it: the whole cycle with a real model, behavioural checks and working-tree provenance included.
Next
Governed continuation studies, before any general claim to restore or fork a live run.

Working surface

Languages
Rust, TypeScript, Python
Runtimes
Pi, MCP, ACP, OpenAI-compatible APIs, local agent harnesses
Inference
llama.cpp, vLLM, SGLang, TensorRT-LLM, GGUF and quantisation
Systems
Linux, Nix, systemd, containers and virtualisation
Correspondence

Building or deploying an agent system?

If privacy, reliability, permissions, provenance or evaluation is the constraint, write with a short description of the system and what you need to be able to show.

george@recensionresearch.co.uk