Recursive Systems

VSM inquiry with AI · Research prototype

From evidence to inquiry

This research prototype explores how AI can support a VSM inquiry while keeping its interpretations open to challenge.

1 · What the source says
“The committee is responsible for reviewing performance.”
2 · What that supports
A review duty is documented. Whether a review happened remains open.
3 · What to ask next
Did a review happen, and what changed as a result?

A made-up example.

The aim is a useful first inquiry: a reading worth checking, a relationship worth examining, or a question worth taking to someone who knows the work. Whether it earns that place is for practitioners to test. Every claim stays open to inspection, and your corrections stay in your name. I would like to hear about a real problem worth investigating, and about cases likely to show where the method breaks.

What this is, and is not

This is a research prototype. It is not a diagnosis, a measure of how well an organisation works, or a finished product. No result has yet been checked with the people concerned.

  • A quotation that matches its source shows where a reading came from, not that the reading is right.
  • The app’s labels and counts are readings of documents. They do not measure variety, and they do not show that a function works or that a unit is viable.
  • A gap is something to investigate. Missing evidence is not proof that a function is absent.
  • A person can accept a reading for a stated purpose. That does not make it a fact about the organisation.

The case

The case is ScotSTAR, Scotland’s specialist transport and retrieval service. A read-only copy of the application, as it stood on 23 September 2026, is not published yet.

Dates

  • 28 July 2026: I submitted the abstract. It named two public-record cases, Robin Hood Energy and the Charity Commission inquiry into New Wineskins and U-Turn Move on Homes.
  • 6 August 2026: I withdrew Robin Hood Energy after cold probes showed the models already knew the case’s history.
  • 13 to 14 September 2026: I ran an earlier ScotSTAR analysis, on thirteen documents.
  • Friday 18 September 2026: I gave the talk at Metaphorum 2026, Alliance Manchester Business School.
  • 20 September 2026: I ran the prior-knowledge check on ScotSTAR.
  • 21 to 23 September 2026: I ran ScotSTAR through a standard Map (21 September), a Deep Map (22 September) and a Diagnosis (23 September).
  • 23 September 2026: the published copy was frozen, as the application showed the case that day.

The analysis workflow

The route at a glance

Documents and answers Extract entities, relations, quotes Assemble candidates Audit: classify or abstain Map, level by level Diagnosis: findings and questions A person decides rules model diagnosis first function-led inquiry

“Map, level by level” includes a new evidence check each time the inquiry changes its system in focus.

Every stage, in the software’s own terms, is in the technical appendix at the end of this page.

What the app assesses in VSM terms

Start with the transformation and the system in focus. Then look at how operations, coordination, control, independent checking, intelligence and policy relate, and follow them into child systems where the evidence supports it.

Each assessment shows what the evidence supports and what to ask next.

Each function on the Map says whether it rests on a checked quotation, a person’s correction or role evidence. Where the rule’s word list did not cover a typing, it is admitted on the model’s reading and marked provisional. A child boundary nobody has examined is labelled “Not assessed”. A channel whose flow nobody has described is shown as “Unassessed”, and a blank mechanism cell is recorded as unassessed or still open.

Frame the inquiry

Recursion, system in focus, criterion and level

Changing the system in focus to examine an S1 operation as a viable system in its own right, within a wider system.

How the app assesses this
The app looks for operations and asks whether the evidence supports opening a child analysis with its own operations and boundary. It records the criterion (Pérez Ríos calls these recursion criteria: Design and Diagnosis for Sustainable Organizations, 2012, s. 2.2, p. 72), the supporting quotation and the parent–child link.
Where it stops
A criterion the software proposes is recorded with its quotation and treated as working context that the evidence can challenge; a person can record their own. An admitted child is a scope worth analysing.
The source says
Beer: recursion is “a next level that contains all the levels below it” (Diagnosing the System for Organizations, 1985, p. 17), and “the System-in-focus is in the centre of a higher level of recursion, in which it is embedded, and it contains a set of viable systems which exist at the next lower level of recursion” (same book, p. 16).
The wider argument (synthesis)
Beer treats a division’s capacity to be hived off as the mark of a contained viable system: “It could in principle be hived-off as a separate business” (The Heart of Enterprise, 1979, p. 118). A service department is different. Its staff could leave and set up as consultants, but in its corporate role it “simply cannot maintain a separate existence” (p. 119; see also p. 515). Cohesion also concerns the balance between local autonomy and intervention by the wider system. A parent–child link gives a place to investigate that settlement; whether the settlement works is a separate question. See Beer’s discussion of autonomy and minimum intervention (pp. 157–160) and the Law of Cohesion across recursive levels (p. 355).
System One

The primary operations that produce the transformation of the system in focus; each S1 element is itself a viable system in the VSM.

How the app assesses this
The app looks for who performs, or is assigned, the primary transformation. Naming an S1 contribution and opening a child analysis are separate decisions. Map reasoning rates each unit’s autonomy across tasking, professional judgement and control of resources.
Where it stops
An assigned responsibility is not evidence of delivery. Professional discretion is not, on its own, operational autonomy.
The source says
Beer: “The set of these embedments will be known as SYSTEM ONE of the System-in-focus” (Diagnosing the System for Organizations, 1985, p. 19).
The wider argument (synthesis)
Beer treats autonomy together with systemic coherence and interdependence, rather than as unlimited independence (The Heart of Enterprise, 1979, pp. 146–152, 160). Beer’s operational elements are those which “exist to do what the system does” (p. 116).
TASCOI

A declaration of the transformation, actors, suppliers, customers, owners and interveners that helps frame the inquiry.

How the app assesses this
The app can store a TASCOI declaration in the consultant’s name. It frames the inquiry and is kept apart from the evidence for functional claims. It can be revised.
Where it stops
A declaration is a point of view to investigate, not proof of the stated boundary, ownership or transformation.
The source says
Espejo and Reyes: “the mnemonic TASCOI where T stands for the canonical form of the Transformation, A stands for the Actors performing the transformation, S stands for the Suppliers, C stands for the Customers, O stands for the Owners and I stands for the Interveners” (Organizational Systems: Managing Complexity with the Viable System Model, 2011, ch. 7, p. 126).
The wider argument (synthesis)
Espejo and Reyes treat identity declarations as hypotheses open to stakeholder viewpoints and boundary debate (Organizational Systems, 2011, pp. 125–128). They set TASCOI apart from Checkland’s CATWOE: TASCOI names “meaningful chunks of complexity in the world”, while CATWOE names “ideas that might be relevant in the world” (ch. 7, p. 126, note 8).

Examine regulation

System Two

Coordination between S1 operations insofar as it damps their oscillations, without taking over their management.

How the app assesses this
Where several operations are mapped, the app checks whether mapped coordination reaches every one. Partial or missing reach can become a question about how coordination happens.
Where it stops
Reach on the map is not damping in practice. A coordination title, or one link to one operation, is not reach across all of them.
The source says
Beer: “Its function is not to command, but to damp oscillations” (Diagnosing the System for Organizations, 1985, p. 68).
The wider argument (synthesis)
Beer describes an anti-oscillatory service to the operations and treats bodies such as monastic orders as Systems Two, warning against equating it with routines (The Heart of Enterprise, 1979, pp. 184–186).
System Three

Management of the inside and now, seeking cohesion and synergy across operations through resource bargaining, accountability and operational control.

How the app assesses this
The app looks for operational control, resource allocation and performance management, and records resource-bargaining and accountability links between S3 and the operations.
Where it stops
Finding these links does not show that reporting is useful, that autonomy is enough or that the whole benefits. That needs evidence of how the arrangements work.
The source says
Beer: “System Three, then, is responsible for the internal and immediate functions of the enterprise: its ‘here-and-now’, day-to-day management” (Diagnosing the System for Organizations, 1985, p. 86), with the resource bargain as “the ‘deal’ by which some degree of autonomy is agreed between the Senior Management and its junior counterparts” (same, p. 38).
The wider argument (synthesis)
Beer develops S3’s view of the whole and the search for synergy, including contributions from operating managers (The Heart of Enterprise, 1979, pp. 201–203, 207).
System Three-Star

Independent access to operational conditions, supplementing the accounts that reach S3 through routine management reporting.

How the app assesses this
The app admits System Three-Star evidence when a clause about the unit describes independent audit, verification, inspection or investigation, whether of controls, compliance, performance or risks, or directly of operational objects such as sites, vehicles, samples, wards or crews.
Where it stops
A compliance duty, routine oversight, a job title, or an audit of management reports or accounts does not count on its own. Independence here means access beyond routine management reporting; it need not be a separate department or person. The software stage called Audit is the analyst’s instrument, not the organisation’s System Three-Star.
The source says
Beer: “Such mechanisms work sporadically” and “penetrate straight to the operations themselves”, and “These procedures, which may generically be called ‘audits’, are indicated as the sixth vertical channel” (Diagnosing the System for Organizations, 1985, p. 82). He adds that “routine and regular audits surrender a large part of the variety they generate to no purpose whatsoever” (same, p. 85).
The wider argument (synthesis)
Beer describes direct access to operational activity beyond managerial filtering, distinct from the command axis (The Heart of Enterprise, 1979, pp. 211–212, 215–216). When an audit report is evidence, ask how its observations were obtained, what filtering occurred and whether it gave management access beyond routine reporting. The existence of a report does not settle those questions (pp. 211–212).
System Four

Intelligence and adaptation: understanding the wider environment and the organisation itself, developing possibilities and examining them with S3.

How the app assesses this
The app looks for intelligence about the outside and the future, and whether it changes operational priorities or resource decisions. It can also admit a sentence showing that the organisation uses, tests or revises a model of itself to compare options or adapt. It records whether the evidence is a mandate or reported activity.
Where it stops
The self-model route is deliberately narrow: the model and its adaptive use must appear in the same sentence. A duty to keep a model is not proof that it was used. The app does not measure whether intelligence or adaptation is adequate. Its readings of the S3–S4 balance are interpretations tied to quotations.
The source says
Beer’s phrase is “the outside-and-then”, and he asks more of System Four than that: “SYSTEM FOUR is not only concerned to manage the outside-and-then, but to provide self-awareness for the System-in-focus” (Diagnosing the System for Organizations, 1985, p. 115).
The wider argument (synthesis)
Beer develops S4’s wider environmental perspective and its interrogable model of the whole system, supporting creative adaptation as well as response to change (The Heart of Enterprise, 1979, pp. 227–238, 241–243).
System Five

Identity and policy for the particular system in focus, providing closure and monitoring the balance between present operations and adaptation.

How the app assesses this
The app looks for identity, policy and normative decisions in the system in focus. Where a source describes mediation between S3 and S4, the app can record the mediating body as a comparator. Role evidence belongs to one level; a person who sits at two levels does not make the functions the same.
Where it stops
A recorded comparator is not proof that S5 monitors the balance, and one episode of mediation is not continuing effectiveness. Missing evidence does not mean S5 is absent.
The source says
Beer locates logical closure and monitoring of the Three–Four homeostat in System Five, then connects closure to the assertion of identity (Diagnosing the System for Organizations, 1985, pp. 129–130).
The wider argument (synthesis)
Beer places S5 in relation to the machinery that balances S3 and S4, providing closure without making constant intervention the preferred pattern (The Heart of Enterprise, 1979, pp. 260–263).

Assess relationships and variety

Variety

The number of distinguishable states or possibilities in a specified set, relative to the distinctions an observer can make.

How the app assesses this
The app records the operations, relationships, regulatory functions and communication mechanisms the evidence describes. Its counts show what has been mapped, which helps locate unanswered questions.
Where it stops
Those counts are not a measure of variety.
The source says
Beer: “a measure of complexity: the number of possible states of a system” (Diagnosing the System for Organizations, 1985, p. 35).
The wider argument (synthesis)
Beer relates meaningful distinctions to the purpose and boundary of the system being examined (The Heart of Enterprise, 1979, pp. 32–36).
Requisite variety

The regulatory repertoire needed to counter relevant disturbances and keep essential variables within acceptable limits.

How the app assesses this
The app checks structural coverage. Do the operations have mapped coordination and control? Does control have audit and outside intelligence, is there a policy and identity function above them, and is there an urgent-escalation channel? The coverage check leaves viability and Beer’s First Axiom unassessed. Autonomy is read separately; see System One.
Where it stops
Coverage does not measure disturbance variety against regulatory capacity. Complete coverage does not prove requisite variety, and gaps may only be gaps in the evidence. Counts of vertical and horizontal channels are shown as coverage only; equal counts are not presented as a variety balance.
The source says
Beer: “only variety can absorb variety”, which he attributes in the same line to Ashby’s Law (Diagnosing the System for Organizations, 1985, p. 35). Ashby’s own statement: “only variety in R can force down the variety due to D; variety can destroy variety” (An Introduction to Cybernetics, 1956, S.11/7).
The wider argument (synthesis)
Beer connects managerial, operational and environmental variety with channel capacity and the rate at which relevant information is generated (The Heart of Enterprise, 1979, pp. 97–102). Beer distinguishes six vertical variety absorbers: interactions among environments, operations and their managers, followed by designed metasystem intervention, System Two coordination and monitoring or audit. The first three arise through the units’ interactions; the latter three need deliberate provision. For a document-based inquiry, the question is which of these relationships the sources describe and which remain unexamined (pp. 216–217).
Homeostat

A reciprocal regulatory relationship through which interacting parts maintain relevant variables within acceptable bounds.

How the app assesses this
The app looks only at the S3–S4 relationship. It records the pair, what is being balanced, any comparator a source names, a supporting quotation, the source’s account of the pattern and whether the source says the balance is holding or failing. It then asks whether intelligence affects operational decisions and whether operations respond.
Where it stops
This is a reading of sources, not a measure of dynamic stability. The record holds the two functions, a quotation and the source’s own description, but no timing; a separate worksheet reads the channels between them from the Map. Disagreement between S3 and S4 can be useful deliberation rather than a failure to regulate. This is the only one of Beer’s whole-organisation variety balances the app assesses. From documents alone the app can record a reading: balanced, S3-dominant, S4-dominant or no evidence, each tied to a quotation. “Balanced” means the documents describe the balance working. Beer gathered his evidence for it by interview and by self- and peer-estimates of attention (The Heart of Enterprise, 1979, p. 369; pp. 449–453).
The source says
Beer defines the state rather than the device: “stability of a system’s internal environment despite the system’s having to cope with an unpredictable external environment” (Diagnosing the System for Organizations, 1985, p. 17).
The wider argument (synthesis)
Beer examines reciprocal understanding and influence, the allocation of attention and resources, and continuity in the S3–S4 relationship (The Heart of Enterprise, 1979, pp. 253–259). Beer names the S3–S4 relationship the Three-Four Balancer (Figure 44, p. 257). S5 monitors that regulatory relationship (p. 262). A recorded comparator helps investigate it.
Amplifier, attenuator, transducer

Attenuation reduces distinguishable possibilities; amplification increases available variety; a transducer converts what is communicated across a boundary into a form the receiving system can use. For adequate regulation, the relevant distinctions must survive that conversion.

How the app assesses this
For each channel, the app can record amplification, attenuation or transduction a source describes at either end. It keeps a named mechanism apart from an open question or a practitioner’s “unknown” or “not relevant”. A blank cell is recorded as unassessed or still open, not as absent.
Where it stops
It records mechanisms, not their capacity. It does not say whether a transducer keeps every necessary distinction, whether a mechanism was designed or grew up, or how it pairs with the return channel.
The source says
Beer: an attenuator is “a device that reduces variety”, an amplifier “a device that increases variety” (Diagnosing the System for Organizations, 1985, p. 35). Transduction is his Third Principle of Organization: “Wherever the information carried on a channel capable of distinguishing a given variety crosses a boundary, it undergoes transduction; the variety of the transducer must be at least equivalent to the variety of the channel” (same, p. 47).
The wider argument (synthesis)
Beer distinguishes transmission capacity from meaningful discrimination across a boundary, and considers what happens when information is collected but not acted on (The Heart of Enterprise, 1979, pp. 97–102).

Algedonic: a signal of pain or pleasure that bypasses routine reporting to reach the level that can act (synthesis of The Heart of Enterprise, 1979, pp. 406–407).

Four more distinctions used in this guide
Three-Four Balancer
The relationship between S3’s concern with present operations and S4’s concern with adaptation. S5 monitors its regulation (The Heart of Enterprise, 1979, pp. 257–262).
Translation and transduction
Correctly transferring words is not enough: the receiving system must be able to distinguish what matters in the signal (The Heart of Enterprise, 1979, pp. 101, 124–125).
Designed and happenstance regulation
Some variety is absorbed through interactions already taking place; other regulatory relationships must be deliberately provided. Look for both (The Heart of Enterprise, 1979, pp. 216–217).
Declared and observed
A stated purpose or duty is evidence of what is declared. Whether the activity happens, and what effects it has, require further evidence. It is a way of reading sources.

Some further structural checks run separately after a Map, on request, and abstain when the inputs they need are not available.

More detail

› Why readings stay provisional

A language model proposes connections and questions from the evidence it is given. The inquiry still has to examine operational autonomy, coordination, control, intelligence and policy together, at each level of recursion. One label, such as “oversight”, can make three different situations look the same:

Situation What the diagnosis must preserve
A committee is required to review performance. Assigned responsibility.
Minutes show it performed a review. Evidence that the review took place.
Evidence shows the review changed a decision. Evidence that the review made a difference.

Collapsing the three into one label is attenuation: it removes distinctions this inquiry needs, and a matching quotation cannot restore them.

Recorded in the case. Included in a model request. Admissible for a particular claim. Accepted into the working map. Selected for the final report. These are five different states, and the system keeps them apart.

I am one person with a prototype. Read the route as part of an inquiry: how evidence is selected and reshaped, which distinctions survive, how an interpretation is challenged and what the next question changes.

Beer names the error that matters most here: “the most common mistake is to seize on the existing organization chart of the institution, and blithely to assume that every division or department shown as depending from the boss is a viable system in its own right” (The Heart of Enterprise, 1979, p. 204). That sentence is why the app abstains. A language model makes the same mistake fluently and at scale, so proposed S1–S5 labels are checked before admission, and a later error can be corrected on the record.

Three commitments follow. Quotations are checked against their cited sources, stage by stage; a match shows the words are there, not that they produced the reading. Each kind of claim keeps its provenance: a person’s declaration, an analyst’s judgement and a synthesis are never presented as quotations from the organisation. Not knowing is recorded: the analysis says what it could not establish and why, an important gap can become a question, and a failed request is recorded as a failure, not as evidence that a function is absent.

Beer, of models such as a VSM map: “a model is neither true nor false: it is more or less useful” (Diagnosing the System for Organizations, 1985, pp. 1–2).

Recursion: an S1 operation becomes the system in focus

Start with the primary transformation. The S1 elements are the operations that produce it. In the VSM, each is itself a viable system: changing focus means examining its own operations, metasystem and environment. The app needs evidence for that unfolding.

Illustrative organisation A transport service

Follow the primary work: providing care and transferring patients. Ask how operational autonomy and local responses to demand are sustained within the wider service.

Follow an S1: a retrieval operationNow the operation is the system in focus

Examine its own S1 operations and S2 to S5 functions. Which disturbances can it handle locally? What needs coordination, a resource bargain or escalation? Its relationship with the wider service remains part of the inquiry.

And its own S1 elements?

Follow the same criterion into their primary work, where the evidence supports it. A team name or a box on an organisation chart does not make a further viable system.

Open the nested boxes to change the system in focus.

What the app must establish

  1. Frame the transformation. A person frames the system in focus. The analysis can propose a criterion for a child scope, which must be supported by evidence.
  2. Separate two judgements. Identifying a primary operation and finding evidence sufficient to open a child analysis are different steps. The latter needs evidence of internal operations and a boundary.
  3. Assess functions in their own frame. Keep the child linked to its parent. Evidence of the parent’s S5 does not automatically carry over to the child.
  4. Keep the uncertainty visible. If the evidence does not justify another level, decline the proposal or ask a question.
Shared identity, local S5?

Each viable system has an S5 function. Its identity and policy must cohere with the wider whole while allowing local autonomy. Shared purpose, or the same office-holder at several levels, does not make those functional claims interchangeable.

Beer, Diagnosing the System for Organizations, pp. 19, 102–104 and 129–130. Terms, sources and implementation limits.

Where the inquiry stops: a proposed child is not accepted merely because the AI drew one. The evidence and the boundary have to support it, and even then a parent–child decision can be wrong. Read the VSM terms and sources.

Reading the map as an inquiry

Follow a mapped relationship back to its evidence, then ask what would show whether it works. These are questions for the reader, not measurements the app makes.

What a box or arrow establishes

A box is a candidate in the working account, and its proposed function needs evidence in the stated system in focus. An arrow means the analysis thinks these two are connected. Follow it back to its evidence. Check the quotation, the interpretation and any recorded human decision separately.

Follow the return path

For a claimed regulatory loop, follow the signal out and the feedback back. At each boundary, ask what distinctions survive the transducer. Can the channels carry the relevant information at the rate it is generated, and is the loop maintained over time? A single arrow answers only part of that inquiry.

These questions draw on Beer’s Second, Third and Fourth Principles of Organization (The Heart of Enterprise, 1979, pp. 99–102, 258).

Keep the dates and response times in view

First ask whether the documents describe the same period. Then distinguish reporting intervals, the time available to respond and the time needed to recover after a disturbance. A relationship described in one period may not describe another.

Beer’s Fourth Principle concerns the continuing cycle without disabling delays; his discussion of relaxation time asks what happens when a further disturbance arrives before recovery (The Heart of Enterprise, 1979, pp. 258, 390).

How the analysis routes differ

Where the analysis commits itself

The four routes side by side

› Rules assemble the candidates Fixed naming rules group the mentions into candidates; a person can correct them.
What it enables
A repeatable set of candidates, each with a fixed set of evidence a person can check.
What it can attenuate or lose
Two spellings of one team stay apart, unknown kinds are not merged, and a relevant passage can sit unlinked.
Recursive consequence
A naming error can drop, duplicate or split an operation before recursion is considered. Once a person corrects it, the next Map reads the corrected records.
› A model assembles the candidates The model proposes which mentions are one unit, before any functional question.
What it enables
Mentions that no rule would join can be proposed as one body.
What it can attenuate or lose
The model sees the case’s name, level and parent, but not its boundary statement.
Recursive consequence
A child analysis of this unit inherits the model’s reading of what the unit is.
› Deep Map: diagnosis first (after an ordinary Map) The model writes an account of the case first; the Map is built from the parts of it that pass the checks.
What it enables
Alternative mechanisms, risks and disconfirming checks are invited, and the account is kept alongside the structure.
What it can attenuate or lose
The model sees only part of the existing Map, and an account can quote its sources correctly and still be wrong.
Recursive consequence
Only candidates that pass the checks enter the Map, so a hypothesis can be kept without adding anything at the level below.
› Function-led inquiry A fixed question per VSM function, on evidence frozen for the run.
What it enables
The model chooses what to look at next, and a contested interpretation stays reviewable.
What it can attenuate or lose
Budgets limit the search, the model starts fresh for each function, and findings must attach to units already in the snapshot.
Recursive consequence
Nothing enters the Map until a person promotes it, which queues an ordinary Map.
Route Its distinctive commitment What it enables What it can attenuate or lose Recursive consequence
Rules assemble the candidates Fixed naming rules group the mentions into candidates; a person can correct them. A repeatable set of candidates, each with a fixed set of evidence a person can check. Two spellings of one team stay apart, unknown kinds are not merged, and a relevant passage can sit unlinked. A naming error can drop, duplicate or split an operation before recursion is considered. Once a person corrects it, the next Map reads the corrected records.
A model assembles the candidates The model proposes which mentions are one unit, before any functional question. Mentions that no rule would join can be proposed as one body. The model sees the case’s name, level and parent, but not its boundary statement. A child analysis of this unit inherits the model’s reading of what the unit is.
Deep Map: diagnosis first (after an ordinary Map) The model writes an account of the case first; the Map is built from the parts of it that pass the checks. Alternative mechanisms, risks and disconfirming checks are invited, and the account is kept alongside the structure. The model sees only part of the existing Map, and an account can quote its sources correctly and still be wrong. Only candidates that pass the checks enter the Map, so a hypothesis can be kept without adding anything at the level below.
Function-led inquiry A fixed question per VSM function, on evidence frozen for the run. The model chooses what to look at next, and a contested interpretation stays reviewable. Budgets limit the search, the model starts fresh for each function, and findings must attach to units already in the snapshot. Nothing enters the Map until a person promotes it, which queues an ordinary Map.

Each route has trade-offs worth testing.

Two requisite-variety questions sit behind the comparison. Does the organisation have the responses its environment demands? Does the inquiry have the responses the organisation demands? The software supports the inquiry’s variety so it can investigate the organisation’s. I have not shown that either balance is achieved by design rather than by default.

› Two provenance views Follow a review forward into later runs, or an output back to its sources.
  • Forward: follow one person’s review into every later run, and see whether each request carried the review’s words or only its reference.
  • Backward: follow one output back through its stages and requests. Where a record is missing, the view says “unavailable” rather than guessing.
  • Neither view shows influence: a passage being present does not show that it shaped the output.
The documents

The document library

The library prepared for the ScotSTAR runs

The library prepared for the ScotSTAR runs had sixteen files: the thirteen documents used in the earlier ScotSTAR case, plus the Scottish Perinatal Network’s 2024–25 annual report, the EMRS workload and outcomes study, and a Scottish Government report on neonatal intensive-care demand and capacity. The Perinatal Network and Scottish Government files are carried as whole-page excerpts; the EMRS paper is complete. Links go to the publishers and may change. The files are not republished here.

The case cites 20 public documents. This list is the library as prepared, not a record of what each model request received.

  1. ScotSTAR annual update report, SAS Board, 28 January 2026
  2. ScotSTAR annual update report, SAS Board, 27 November 2024
  3. ScotSTAR clinical lead: job description (NHS Scotland jobs, 2025)
  4. Head of Service, neonatal and paediatric retrieval: job description (2021, attached 2024)
  5. ScotSTAR referral and activation guideline OG028 v2 (2024)
  6. ScotSTAR fatigue guideline SG011 v2 (6 June 2025)
  7. Winter contingency plan, SAS Board, 26 November 2025
  8. Public board minutes of 26 November 2025 (published January 2026)
  9. Board matters arising, 28 January 2026
  10. Public board minutes of 28 January 2026 (published March 2026)
  11. Clinical Governance Committee report, SAS Board, 26 November 2025
  12. Infection prevention and control activity report, SAS Board, 26 November 2025
  13. Health and Care (Staffing) (Scotland) Act Q2 assurance report, SAS Board, 28 January 2026
  14. Scottish Perinatal Network annual report 2024–25 (whole-page excerpt prepared for the run)
  15. EMRS workload and outcomes in the first decade of national coverage (McHenry et al., 2023)
  16. Neonatal intensive-care demand and capacity modelling, Scottish Government final report (whole-page excerpt prepared for the run)
What comes next

Coming next

Next, I want to make the evidence behind each claim easier to inspect, then test whether that helps a person conduct a better inquiry.

  1. 1. The next case

    I will publish another case only after its Map and Diagnosis have been reviewed together. Each visible claim should show which route admitted it and whether its support is a quotation, a person’s correction or role evidence.

  2. 2. The software, not the model, supplies the quotation

    The model should point to evidence rather than reproduce the quotation. The software can then return the exact words, where they sit and enough context to check the reading. Passages given to the model should keep the document’s intended reading order.

  3. 3. One account of evidence and time

    An assigned duty, a reported event, a future intention and an unresolved conflict should remain different all the way into the Map and Diagnosis. The same applies to a document’s publication date, the time it entered the inquiry and the date of the event it describes.

  4. 4. Test the inquiry, not the quantity of output

    I will start with a few labelled cases and recorded practitioner decisions. The useful measures are unsupported interpretations, honest uncertainty and how long it takes someone to verify a claim. Producing more VSM labels is not success.

What I would test later

Recorded decisions would let me re-run the fixed rules against past judgements, then replay whole analyses once several cases exist. A separate classifier could be a calibrated sensor, never the diagnostician.

Statistical work would need real operational episodes and categories declared in advance. I will not turn counts of mapped VSM functions into a viability score. A channel measure would be comparative and would state which distinctions it counts.

Planned research, not current capability.

Practitioner pilots

I am looking for a few practitioners to try bounded inquiries. You might bring a current question from an organisation you work with, or documents and interviews from earlier consulting work, if you have permission to use them. The organisation need not still be trading. I would also like a return visit after the first analysis.

Richard Gunther, richard@recursive.systems

› Research direction: Agents

An AI agent could help someone use the application: retrieve a case, inspect an argument or propose a question.

A team of agents could itself become the subject of inquiry. What work must continue? Who sets its purpose? What can each worker decide, and what feedback reaches someone able to change the arrangement?

That connects the work to AI governance: how a proposed interpretation becomes permitted action, and how people can challenge or revise it before an error propagates. Applied to agents, the method would be an inspectable inquiry into purpose, authority and feedback.

Research direction. Nothing in this section is built.

Conference details and submitted abstract

Metaphorum 2026, Alliance Manchester Business School

From Evidence to Inquiry: Keeping LLM Inference Provisional in Recursive VSM Diagnosis

The abstract as submitted on 28 July 2026

Purpose. Clemson’s VSM ToolBox and Pérez Ríos’s VSMod showed software can support Stafford Beer’s Viable System Model; however, LLMs introduce a cybernetic problem: plausible inference can harden into organisational fact and propagate across recursion levels. This project traces a path from evidence to hypotheses, questions and interventions.

Methods. The prototype ran on two public-record cases: Robin Hood Energy, and the Charity Commission inquiry into New Wineskins and U-Turn Move on Homes. It proposes operations, relations and recursive structure, promoting each only when an evidence-gated test passes, abstaining otherwise. An operation becomes a child system one level down only when at least two cited source units evidence its internal operations and boundary. Structural gaps become focused questions; answers and corrections persist through reanalysis. Hypotheses and prescriptions carry checks that could confirm or refute them. Any LLM-attributed quotation must match its cited source verbatim.

Findings. Both produced evidence-linked maps and explicit gaps. In one, an entered answer that no role monitored the external environment or future was retained verbatim as evidence of a System 4 gap; no function was invented. VSM operates as a reasoning grammar, not only a diagram.

Implications/Contributions. The contribution is an evidence-to-inquiry loop that turns abstention into diagnostic work while model suggestions remain provisional, complementing Metaphorum 2025’s work on AI-augmented VSM decision environments. Results demonstrate software traceability, not diagnostic validity or organisational impact; practitioner testing comes next. The session is a live demonstration: participants answer gap questions and see reanalysis. When should AI-proposed structure become a diagnostic premise?

Update, 6 August 2026: Robin Hood Energy was withdrawn after cold probes showed the models already knew the case’s history.

Technical appendix

The software’s stages, in its own terms, for readers who want the machinery.

TASCOI is Espejo and Reyes’s mnemonic (Organizational Systems, 2011, ch. 7, p. 126). “Recursion criteria” is Pérez Ríos’s term (Design and Diagnosis for Sustainable Organizations, 2012, s. 2.2, p. 72).

› Open the full route map: every stageClose the full route map

How the analysis paths connect

Read downward. A stage that spans several columns is shared by those paths; the others belong to one path. Open any stage to see what it receives, does and produces.

The two Map columns show the two ways a Map can start. Factual preparation is separate: it stops after the first rule-based skeleton and is not a completed Map. Function-first is a separate assessment; promoting its reviewed roles is a human action that queues an ordinary Map. Diagnosis is always requested separately, after an eligible Map has completed.

Some stages depend on a setting; each says how it is set in this release. What actually ran on a case is in that run’s own record.

Grey: shared source preparation Blue: ordinary Map seed Orange: Deep Map seed Purple: Function-first Blue + orange: both Map modes Teal: Diagnosis, a separately queued run Green: human action

A model stage may send several requests, including batches and retries.

Ordinary Map: build from extracted candidatesRules or model
Deep Map: interpretation-led seedRuns when a Deep analysis is requested
Function-firstSeparate assessment run
Document ingestion and parsingShared by all three paths
Receives
Uploaded documents
Does
Parse content into stored source spans; extraction runs only when requested.
Produces
Stored document spans
Document extraction: reasoningWindowed extraction; repeats over longer documents
Receives
Document spans in windows
Does
Produce a prose evidence ledger; carry prior structure to the next window.
Produces
Prose ledger
Document extraction: formattingSeparate from extraction reasoning
Receives
Prose evidence ledger
Does
Format and resolve extracted nodes against the document spans.
Produces
Resolved nodes and source references
Relation recoveryOnly when the formatting pass resolved two or more entities and no relation at all
Receives
Resolved nodes and spans
Does
Recover document-local relationships for emission when that condition holds. When it does not, the stage records that it was not needed and no model is called.
Produces
Pending relation emissions, or a recorded “not needed”
Document channel signalsOptional fourth stage of the same attempt, after relation recovery; off unless the operator turns it on
Receives
The window’s document spans
Does
When the flag is on, extract grounded channel signals in a separate model call with its own budget. With the flag off the attempt is byte-identical to the producer before the feature existed.
Produces
Signals with verbatim source phrases
Evidence emissionJoins extraction, relations and optional signals
Receives
Resolved extraction, relations and optional signals
Does
Emit document evidence and preserve source links; stored evidence can be selected for later analysis.
Produces
Document evidence events and their source spans
Framing declarationShared case frame; independent of ingestion order
Receives
System in focus, recursion criterion, TASCOI and presenting concern
Does
A person declares the frame used for analysis. The selected frame and sources feed each later run; source bytes remain evidence.
Produces
Recorded framing
Map run scope and the analytical basisIn the software both columns use the Map strategy named ‘standard’; they differ in how the Map is seeded
Receives
Frame, scoped evidence and chat pointers
Does
A queued run carries its scope, and building the skeleton is its first step. The inputs the analysis actually used are captured just before the reasoning stage’s model call, and are attached to the run only if the run completes. A failed run has no recorded basis.
Produces
A run-scoped basis on success, nothing on failure
Function-first preparation and own snapshotOff unless the operator turns it on: no new Function-first run can start otherwise
Receives
Canonical working set, source units and frame
Does
Capture a bounded repeatable-read basis and admit a prepared assessment run.
Produces
Sealed Function-first snapshot and admitted run
Skeletonisation choiceDeterministic unless the model skeleton is chosen
Receives
Scoped evidence and chat pointers
Does
Choose deterministic skeletonisation or the single structured model pass below. With the Deep route on, a deep run uses the other seed lane instead, unless the zero-node fallback is switched on and the deep seed returned nothing.
Produces
Skeleton nodes and channels
Free-form diagnosis seedA deep run, with the Deep route on in the shipped configuration and able to be switched off; switched off, a deep run falls back to the seed lane on the left
Receives
Whole corpus, case frame and prior structure for orientation
Does
Generate a diagnostic seed through the diagnosis-first engine. If its content cannot be decoded, the seed is refused.
Produces
Claims, actor material and diagnostic seed
Execution authorisationSeparate from preparation
Receives
Prepared sealed run
Does
An authenticated person authorises all six functions or one selected pilot function. Provider configuration alone is insufficient.
Produces
Bound execution selection and authorisation
LLM skeleton Pass 1One structured pass, only when the model skeleton is chosen; the default skips it
Receives
Scoped source evidence
Does
Produce structured node and channel proposals in one structured pass; a failure or an empty result retries, which costs more calls. Then validate deterministically, resolve source handles and write the admitted skeleton; there is no separate prose-to-formatting model pass here.
Produces
Validated, source-resolved skeleton candidates
Backward groundingWithin the Deep Map seed
Receives
Free-form claims and source corpus
Does
Resolve the claims back to source passages; retain the distinction between source-anchored observations, provisional interpretations and unresolved claims.
Produces
Grounded findings and grounding resolutions
Dispatch and reservationProvider identity and budget required
Receives
Authorised selection and sealed basis
Does
Check the selection, the provider and the budget, and reserve a bounded amount of model work for each turn and tool call. Whether the evidence is still current is checked once, at authorisation, not again during the run.
Produces
Admitted request or visible refusal
Factual preparation runA Map requested as factual preparation only: deterministic skeletonisation only, then it stops
Receives
Frame and scoped evidence
Does
Wire the deterministic skeletonise stage and nothing else. Trace, audit, structuring, the cross-level channel check and Map reasoning are left unwired, so the run terminates after preparation. Requesting it together with a deep run is refused at the run row.
Produces
A skeleton and its captured basis, recorded as a factual-preparation run. Not a Map that Diagnosis can read.
↓
Gap probeReads backward-grounded findings and the corpus
Receives
Grounded findings, raw corpus and prior structure
Does
Probe missing evidence; its open questions join the Deep seed result.
Produces
Gap observations and open questions
Track B: function assessment loopS1 · S2 · S3 · S3* · S4 · S5
Receives
Sealed basis, compact index and selected function
Does
For each authorised function the model takes a turn, makes a bounded number of searches and fetches, reads the passages returned, then either takes another turn or gives its verdict. Within a function, earlier exchanges stay in the request. Only selected functions run, and budgets or connection failures can stop the work.
Produces
Assessment verdict and retained exchanges
↓
Seed projection and persistenceCompletes the skeleton slot in a Deep run
Receives
Grounded Deep seed result, gaps and scoped frame
Does
Persist untyped carrier nodes and missing-element markers. Failures abandon the run. A successful run that produced no node falls back to deterministic skeletonisation only when that fallback is switched on, and it is off unless the operator turns it on.
Produces
Persisted seed nodes and missing-element markers
Quote grounding verificationChecks sealed bytes and windows actually returned
Receives
Verdict anchors and tool receipts
Does
Each quotation must appear exactly in a passage this function was shown. That proves its source, not that it supports the claim.
Produces
Verified anchors or refused terminal/binding
Map traceBoth seed modes; conditional on observed chat signals
Receives
Skeleton and channel material
Does
Trace channel content when a chat signal is observed. Both seed modes enter the same conditional trace gate.
Produces
Channel content states, or an unrun stage
Persist assessments and findingsCan stop here; no Map Audit required
Receives
Accepted terminals, bindings and anchors
Does
Persist assessments, findings and supporting receipts. Evidence status and execution status remain separate. Unassessed does not mean absent; findings remain provisional model interpretation. Subsequent proposal review and promotion are separate, optional work.
Produces
Durable function findings and execution outcome
Audit evidence capsulesCandidate-local assembly; joins trace states at audit
Receives
Candidate nodes and grounding events
Does
Gather the evidence the Audit may see for each candidate. A candidate with no grounding is refused; channels are checked separately. This builds the request.
Produces
Audit inputs and grounding closure
Derive eligible proposalsOptional feature-gated completed-run derivation
Receives
Completed assessment findings
Does
Derive immutable proposals for review when the proposal feature is enabled; derivation failure is separate from settled assessment execution.
Produces
Proposals linked to findings
Map Audit: assess functions and relationshipsProposed typings, relationship readings and unresolved outcomes
Receives
Candidates, evidence capsules and trace states
Does
Assess candidate actors and relationships against selected source passages. A model proposes functional roles and supported relationship readings; code checks references, quotations and the applicable admission rules. The deterministic evidence gate covers node typings; channel typings have their own checks. The current homeostat assessment is limited to S3–S4 interactions. This checks the proposed representation, not the organisation’s actual practice.
Produces
Admitted typings, channel assessments and explicit unresolved outcomes
Human review decisionsAccept, caveat, defer, reject or withdraw
Receives
Immutable proposals and evidence
Does
Record append-only human decisions. Acceptance alone does not add canonical Map roles.
Produces
Review decision chain
Pre-structuring completeness checkAfter audit; not universally a stop gate
Receives
Typed/abstained node set
Does
Calculate audit coverage. With structuring off, this is the census location; with it on, final product census emission is deferred. Strict evaluation and product routes have different failure rules.
Produces
Coverage calculation and incomplete-audit signal
Sign and freeze the reviewed assessmentExact proposals, decisions and currentness
Receives
Human decisions and sealed assessment basis
Does
Sign an immutable review assembly over exact identities and hashes. Signing is separate from promotion.
Produces
Signed Function-first review report
Assess possible child scopesOn in the shipped configuration and can be switched off; the two recursion mechanisms never both run
Receives
Audited structure and focal system
Does
With structuring on, structuring decides which S1 units may open a child analysis. With it off, the older recursion step may run on its own eligibility check.
Produces
Eligible recursive children
Promote accepted roles into a new MapExplicit human command
Receives
Current signed report and eligible accepted proposals
Does
Recheck the person’s authority, the quotations and the exact selection, record the accepted roles and a receipt, then queue one ordinary Map. With multi-role on, units whose roles were accepted can skip the next Audit, but other units, channels, child systems and the later stages can still need work.
Produces
New Map run; no direct shortcut to Diagnosis
Final post-structuring censusWhen structuring ran; before the cross-level channel check
Receives
Revised typed/abstained node set
Does
Recheck and emit the final census after structuring. The census records how much of the map has been checked. Product runs carry on even when it is incomplete.
Produces
Run-scoped census and incomplete signal
Assess channels between recursive levelsOn in the shipped configuration and can be switched off, and the prerequisite stages succeeded
Receives
Parent System 1 and child-system pairs
Does
Assess the vertical channels between each parent S1 and its child, once children exist. It is skipped if a required earlier stage failed or is incomplete.
Produces
Typed cross-level channel observations
Map reasoningLast analytical producer; relevant gates must pass
Receives
Revised typed structure, channels and audit abstentions
Does
Produce provisional hypotheses, gaps and actor overload observations over the revised Map. A malformed answer is retried once.
Produces
Reasoning outputs and exact reasoning inputs
Map settlement and input basisSuccessful run captures its final reasoning basis
Receives
Map structural writes and reasoning outputs
Does
Persist the Map and bind a successful eligible run to captured analytical inputs. Earlier stage outputs are written within the Map transaction.
Produces
Run-scoped Map, basis and outcome
Diagnosis input captureSeparate requested run; requires an eligible current Map
Receives
The latest Map run, which must itself have completed and still be current, and its basis; frame, evidence and observations
Does
Capture coherent repeatable-read inputs. Stale, incomplete or unavailable Map authority blocks Diagnosis; factual prep alone is insufficient. Function-first must first enter a new ordinary Map through human promotion.
Produces
Diagnosis basis bound to the Map run
Diagnosis reasoningSeparate from the Deep Map seed
Receives
Captured frame, sources, structure and observations
Does
Generate the diagnostic certificate against retained inputs.
Produces
Certificate JSON
Check the structured diagnostic outputRetries once on a malformed answer; on success the run continues straight into the gate
Receives
Certificate JSON and rendered evidence
Does
Validate the primary target catalogue and mechanism comparisons. Each later write gate checks its own quote fields.
Produces
Validated findings, prescriptions and gaps
Structural hypothesis reasoningOptional parallel branch when live observation keys exist
Receives
Structural seeds and grounded prompt
Does
Generate provisional prose structural hypotheses alongside the primary findings path. With no eligible observations, the path continues without this branch.
Produces
Prose hypotheses
Structural hypothesis harvestOptional branch’s formatting model call
Receives
Prose hypotheses
Does
Format the hypotheses and key them to observations. Any hypothesis left with no observation key is dropped.
Produces
Structured hypotheses, appended at settlement as their own records
Diagnosis confidence gateRuns on the primary findings path, before the optional structural branch
Receives
The certificate’s “confirmed” findings, and a pre-call record of which Map hypotheses the prompt included.
Does
Apply deterministic confidence rules, including backing-hypothesis source grounding.
Produces
Gated provisional findings
Diagnosis emissionRun-scoped records
Receives
Gated findings and commitments
Does
Emit prescription reasoning events while preserving finding, hypothesis and human commitment distinctions.
Produces
Retained diagnostic records
Diagnosis renderSaved records become visible
Receives
Emitted records
Does
Render retained findings with evidence and human-review surfaces.
Produces
Consultant-visible Diagnosis
Compose diagnostic reportAutomatic best-effort, or explicitly requested
Receives
Run-scoped Diagnosis findings and commitments
Does
Compose for a terminated completed/abstained prescription run only; Function-first is excluded. Composition failure does not change the analytical run outcome.
Produces
Deliverable or separate composition failure
Map inquiry agendaBranch from Map settlement, not the report
Receives
Map gaps, open needs and answered gaps
Does
Build authoritative inquiry needs for the consultant.
Produces
Questions to investigate
Function-first inquiry needsOwn post-assessment read model; not automatic chat
Receives
Valid function assessments and open needs
Does
Where a function was considered but not evidenced, the run review can show a request for evidence. A function left unassessed by a run problem stays separate and never becomes a question about absence.
Produces
Eligible Function-first inquiry needs, separate from execution failures
Map question generationConditional generation from the Map agenda
Receives
Allowed Map inquiry needs
Does
The Map’s question generator can generate on a cache miss when enabled; retained generations are distinct from a fresh question call.
Produces
Rendered or retained questions
Consultant chatUser-triggered Map inquiry branch
Receives
Agenda, available Map hypotheses and conversation
Does
Reply using the selected inquiry context. The reply itself emits no observation events: it returns the turn to the caller and saves it.
Produces
One saved chat turn
Selected question agentsAlternative branches for selected Map question kinds
Receives
Selected question kind and permitted context
Does
Dispatch to the wired question agent appropriate to the selected question; the four wired agents are alternatives. The agent that records amplifiers, attenuators and transducers on a channel makes one grounded observation per turn, dropping any it cannot resolve to a source passage.
Produces
Question-specific inquiry response, and the variety-mechanism observations
Chat extractionDurable optional job → a later Map input
Receives
Saved turn text
Does
Extract scoped chat evidence into source-backed reasoning events, including the inter-party observations a reply never emits.
Produces
Recorded chat events, including inter-party observations, for later Map input
Explicit audited Map resumeOptional non-deep dispatch
Receives
A selected source Map run and its retained durable Audit
Does
Verify source-run identity and audit coverage with a read-only proof: no model calls or writes in that check. Only an explicit eligible non-deep continuation can skip seed, trace and fresh Audit, then enter the downstream shared Map stages. This needs its own proof.
Produces
Proven continuation inputs, or refusal
Standalone diagnosis-first resultEngine stopping point; not another app route
Receives
Source corpus and case frame
Does
The standalone engine call runs diagnosis → backward grounding → gap probe, separates source-anchored observations from hypotheses and unresolved claims, and returns. It does not project Map nodes or run Map Audit. The integrated Deep Map seed above calls this engine and then projects its result into Map.
Produces
Diagnostic result without a Map
Answers and interview materialOptional case-level evidence contribution
Receives
Questions, human answers and interview material
Does
Record source material and queue the routing of that evidence, retaining what was admitted and the receipt for it. An answer can queue a rerun of the analysis, or record that a rerun is owed. Neither is a required step between analysis stages.
Produces
New source units, admitted evidence, routing receipt and rerun/debt record
Corrections and structural feedbackNew correction events; never rewrite old evidence
Receives
Map structure, findings and their sources
Does
Record new correction/feedback events and, when requested, select a later Map capture.
Produces
Correction records and later Map inputs
Analytical reviewSeparate from Function-first report signing
Receives
Findings, hypotheses and their source evidence
Does
A person records an analytical review. Human declaration, canonical promotion and report review remain distinct authorities; later Map or Diagnosis captures can consume eligible review records.
Produces
Human review records
Commitments and observed outcomesOptional follow-up; feeds later Diagnosis
Receives
Chosen actions and subsequently observed outcomes
Does
Record human commitments and later outcomes separately from the model’s prescriptions; use eligible records in a later Diagnosis capture.
Produces
Commitment/outcome records
How the analysis works

The method

The groups below describe the declared reference workflow. The route map above shows how the paths connect; retained run records establish what actually ran.

› Every stage, in readable parts 11 parts. Each stage says what it does and, where it applies, where a person can question or change the interpretation.

The analysis is a chain of stages. Some are paid calls to a language model with a fixed instruction and a strict output format; most are ordinary code that checks, records or displays. Model outputs that assign a function, a channel type, a mechanism or a balance must quote the passage they rest on; how the quotation is checked depends on the stage. A few intermediate outputs, such as a run’s reading of its own purpose, are not quote-checked and are not shown as findings. Treating a silent channel as healthy is a default, and an attenuation: “healthy” and “unknown” collapse into one state, so it needs review. A person can question or overrule an interpretation at several points.

› Prepare source material Parsing, document extraction and factual preparation of what the analysis reads. 7 stages
  1. Parse and span documents

    Prepares data

    Turn an uploaded PDF into stored source passages with a local parser. No model is called and nothing is paid.

  2. Read a document into a ledger

    Model call

    The first paid reading of each part of a document: the model writes a prose ledger of the entities, roles and relations the passages support, quoting them.

  3. Transcribe the ledger to structured entities and relations

    Model call

    Turn the ledger into a structured proposal of entities, relations and quotation hints, which the software then matches to the exact passages.

  4. Recover missed relations

    Model call

    A bounded third call over the resolved entity catalogue: propose relations between known entities that the transcription missed, grounded in the same spans.

  5. Flow-trace channel signals

    Model call

    A separate model call that reads the spans for information flows with a byte-exact verbatim phrase and within-document locator, emitted as channel-signal events.

  6. Channel-signal backfill (operator script)

    Model call

    Record channel-signal observations for documents already held, so the Map and Diagnosis prompts carry provisional signals, without turning the ingest stage on.

  7. Record document evidence events

    Saves

    Write the resolved entities, relations and signals as document evidence, safe to replay after a failure through the stored extraction attempt.

› Frame the case and select evidence The consultant’s declared frame and the exact inputs each analysis captures. 4 stages
  1. Declare the case frame

    Saves

    The consultant’s explicit frame: system in focus, recursion level and criterion, TASCOI and the presenting problem. A person declares it.

    Where a person steps in. The consultant declares the system in focus, the recursion level and the scope; the analysis reads that frame and may question it, never override it.

  2. Decide whether new evidence reruns the analysis

    Prepares data

    When an answer or interview arrives, give any rerun an identity tied to that evidence and track whether a rerun is owed, so a new run can be traced to the evidence that caused it.

  3. Select the current Map inputs

    Prepares data

    One coherent read-only snapshot of what a Map run would see now: typed nodes and channels, source units, the diagnosis packet and frame values.

  4. Capture the Diagnosis inputs

    Prepares data

    Bind the latest Map run and its input basis, check it is current, and assemble the render context the Diagnosis prompt is built from.

› Initial analysis and gap exploration The free-form stages of a Deep run. 2 stages
  1. Free-form initial analysis (deep run)

    Model call

    One pass over the whole corpus by a frontier model: bold diagnosis with per-fact verbatim quotes, which the backward grounder then checks before anything is written.

  2. Gap exploration probe (deep run)

    Model call

    A normative VSM completeness sweep over the corpus and the diagnosis so far: what is absent or under-evidenced, each with a probing question. Gaps are inferences, never citations.

› Build and examine the system map The structural and reasoning stages of a Map run, from the skeleton to the reasoning. 11 stages
  1. Skeleton from evidence (deterministic)

    Prepares data

    Build the first nodes and channels from extraction events without a model call. This is the default skeletonize mode.

    Where a person steps in. After the skeleton is built the consultant corrects the working set: merges duplicate actors, renames, retypes; each correction is a recorded event the next run reads.

  2. Skeleton by model

    Model call

    Identity-only proposal of nodes and channels from source units and chat pointers, with verbatim quotes and no S-labels.

  3. Trace signal flow

    Model call

    Follow each observed signal through the skeleton: origin, first visible point, degradation state and any algedonic bypass, quoting the signal text verbatim.

  4. Assemble the audit evidence capsule

    Prepares data

    For each candidate, select the closed set of evidence the Audit may see: its grounding, one step along its extracted relations, and the source passages those cite. The Audit request is built from this set, never from the whole case.

  5. Audit: classify or abstain

    Model call

    The canonical typing authority: assign S-labels to nodes and Beer channel sub-types to channels only where the evidence supports it, otherwise abstain with a reason.

    The gate exists to avoid false holism: a map that looks complete because the model filled the empty boxes.

    Where a person steps in. Audit classifications and abstentions are visible per node; a consultant’s correction supersedes the model’s classification and is kept as its own record.

    › What this stage will not admit

    These are the gate’s own refusal sentences, quoted from its declared vocabulary.

    • Semantic evidence gate declined the typing because a name or role title is not evidence of organisational function.
    • Semantic evidence gate declined the typing because the target-bound quote supports the proposed function but does not distinguish a formal mandate from described activity without overclaiming enactment.
    • Semantic evidence gate declined the typing because the target-bound evidence negated or prohibited the proposed function rather than assigning or describing it.
    • Semantic evidence gate declined the typing because the same target-bound evidence supports more than one organisational function and the scalar label cannot preserve that role overload honestly.
    • Semantic evidence gate declined the synthesis typing because a contribution did not verbatim-anchor to an exact grounding event in the cited source.

    The gate has 28 declared grounds; these are 5 of them.

  6. Completeness census

    Validates

    After audit: every node must be typed or explicitly abstained. On the product route an incomplete census is recorded and structuring, recursion and reasoning still run; only the strict evaluation routes stop on it.

  7. Structuring: System One operations

    Model call

    Identify System One: the operational elements, which “exist to do what the system does” (Beer, The Heart of Enterprise, 1979, p. 116). Record proposed environment/S1 corrections as events; they are not applied automatically. Pass eligible candidates to the recursion step.

  8. Cross-level channels

    Model call

    For each parent S1 / child VSM pair, classify the app’s five channel kinds (resource bargaining, accountability, legal/corporate, audit and algedonic) strictly from verbatim evidence. These are the app’s kinds, not Beer’s six vertical components of corporate cohesion (The Heart of Enterprise, 1979, p. 217; Diagnosing the System for Organizations, 1985, p. 82, where audit is “the sixth vertical channel”). In Beer, the algedonic signal is a separate arrangement (Diagnosing the System for Organizations, 1985, p. 133).

  9. Map reasoning: hypotheses

    Model call

    The heaviest Map pass: over the typed structure and the diagnosis packet, emit structural, systemic and general hypotheses, S3/S4 balance, autonomy and gaps, each verbatim-grounded.

    Where a person steps in. Each hypothesis shows its quotation; a consultant can record an analytical review that challenges, accepts or rewrites it, with the review kept beside the original.

  10. Recurse: spawn child systems

    Model call

    Per typed S1 candidate, decide whether to spawn a child VSM (evidence of its own operations and boundary, with a stated criterion) or abstain.

  11. Canonical structure and reasoning events

    Saves

    Every Map output lands as events through a single write point that checks each quotation appears in its source; canonical nodes and channels are projections of those events.

› Assess VSM functions The separate Function-first assessment route. 3 stages
  1. Function-first assessment loop (Track B)

    Model call

    A tool-using loop over a frozen basis: for each of six VSM functions the model searches evidence, opens dossiers, fetches passages and records a verdict with quotes.

  2. Quote grounding verification

    Validates

    Every quote in a verdict must be byte-exact in a source the loop actually opened.

  3. Function assessment findings

    Saves

    Findings, turns, tool receipts and transport receipts retained per run.

› Develop the diagnosis The prescription certificate, structural hypotheses and the report. 8 stages
  1. Diagnosis certificate

    Model call

    The diagnosis itself: findings, gap probes, prescriptions and structural inferences, in plain language, against a closed list of targets. A finding counts as “confirmed” when every quotation it rests on was found in its cited source. An overlap of duties is never reported as confirmed.

  2. Certificate and citation validation

    Validates

    Validate the certificate against the closed target catalogue, resolve mechanism comparisons, and check the certificate’s quotations against the rendered source spans.

  3. Structural hypotheses: reasoning

    Model call

    Free-prose, forward-looking hypotheses keyed to recorded observations rather than to a quotation, extending the certificate’s cross-cutting inferences.

  4. Structural hypotheses: harvest to JSON

    Formats

    A formatting-class model turns the prose into hypothesis rows keyed to observation keys, predicted risk and how to verify; keys outside the live set are dropped.

  5. Confidence gate

    Validates

    A finding with no source grounding is downgraded to a gap probe; its confidence cannot exceed its backing’s; if the backing cannot be looked up, the check refuses.

  6. Translate and record findings

    Saves

    Translate the gated certificate into recorded prescription reasoning and commitments, through a write point that checks each quotation appears in its source.

    Where a person steps in. Findings are proposals until a person commits, revises or withdraws them; the case record shows which state each is in.

  7. Generate the diagnosis report

    Formats

    Compose the deliverable from one Diagnosis run’s retained commitments and findings. A separate action from the Diagnosis itself and from reviewing it; no model call.

  8. Diagnosis page

    Renders

    Present the retained findings, their sources and history; nothing is generated here.

› Prepare investigation questions The routes that produce questions: the deterministic agenda and the model-rendered engine. 3 stages
  1. Question agenda (deterministic)

    Prepares data

    The authoritative agenda: Map gaps, open inquiry needs and answered gaps, ranked by the server-side inquiry policy without a model.

  2. Question rendering by model

    Model call

    The model words the needs the rules selected as plain-language questions, with answer types and grounded options; it cannot choose the question.

    Where a person steps in. Investigation questions are answered by the consultant in the conversation; an answer becomes a cited source unit the next analysis can read.

  3. Differential questions

    Model call

    Episodic differential-diagnosis questions that split competing hypotheses.

› Conversation The consultant conversation: reply, turn extraction, side agents. 3 stages
  1. Conversation reply

    Model call

    The synchronous next move in the consultant conversation: the agenda owns question selection; surfaced quotes are substring-checked before the reply is returned.

  2. Turn extraction

    Model call

    Asynchronously read each consultant turn into typed reasoning events with a verbatim quote and confidence; no VSM typing, no hypotheses.

  3. Per-question side agents

    Model call

    Helper agents for one selected question, each with a shared instruction and one for its question type; they record only confirmed items, each with a verbatim quotation.

› Human feedback Answers, interviews, reviews, promotions, corrections and outcomes recorded by people. 9 stages
  1. Consultant answers a question

    Saves

    A consultant’s answer becomes a cited source passage and closes the gap it answers; a later correction is a new event, never an edit.

    Where a person steps in. An answer to a gap question is recorded with its author and time and is cited like any other source.

  2. Map corrections and identity confirmations

    Saves

    Human corrections to the Map and to observations, and confirmed actor identities, recorded as new events that the next run reads.

  3. Duplicate-actor suggestion

    Model call

    Asks whether two canonical records denote the same real-world entity, given the passages cited for each and any passage naming both. A SUGGESTION beside the pair: it merges nothing, and the consultant’s own ruling remains the only write to structure.

    Where a person steps in. A model may suggest that two records are the same actor; only the consultant’s ruling merges them, and the ruling names who decided and why.

  4. Staff interview answers

    Saves

    Observed staff testimony captured against the interview agenda becomes cited source passages, read by the Map and the Diagnosis like consultant answers.

  5. Route human answers into admitted evidence

    Saves

    A durable job that takes consultant, staff-interview and chat-turn answers and records their analytical evidence admission with a routing receipt, so later analyses read admitted evidence rather than raw text.

  6. Intervention outcomes and failed attempts

    Saves

    After a Diagnosis: decisions on prescriptions, observed interventions, judged outcomes, learning, and recorded failed attempts, which the next Diagnosis reads.

  7. Consultant analytical review of a finding

    Saves

    The consultant records a stance on one named finding, a gap probe, a critical flow, a breakpoint, a structural hypothesis or a Map gap, with cited source passages. It is one recorded review, append-only: a later review or withdrawal retires it.

  8. Review, sign and freeze, then promote assessed structure

    Saves

    Human decisions on function proposals, signing and freezing the assessment report, and promoting accepted roles into the structure the next Map reads. Separate from reviewing a Diagnosis finding and from generating the report.

    Where a person steps in. A signed and frozen function report can be promoted into the structure only by a person; the promotion is an event with its own authority record.

  9. Contextual contribution

    Saves

    Neutral historical context recorded on the framing with exact prompt, response, speaker and capture time. The analysis ignores it.

› Evaluation Opt-in probes and harnesses. 2 stages

Corpus selection · criterion 4

ScotSTAR’s prior-knowledge check, its limits and the paired prompts and answers are on the lessons page. Read the check and the records →

  1. Corpus harness (Wells end-to-end)

    Prepares data

    An end-to-end test: create an organisation and frame, load a corpus through the product’s own entry points, run a Map and a Diagnosis, and score the typing against a reference.

  2. Live LLM evaluation probes

    Model call

    Grounding and S4-discrimination probes that reuse the audit system prompt with a synthetic frame.

› Shared machinery Run lifecycle, transport, structured-output validation, embeddings. 4 stages
  1. Run lifecycle: claim, heartbeat, terminate

    Saves

    Every Map, Diagnosis and function-first run is a run record: claimed for dispatch, kept alive with regular signals while it works, guarded against a newer superseding run, and closed once with its outcome and input basis.

  2. Crash sweep

    Saves

    A scheduled table scan that marks runs whose heartbeat stopped as crashed.

  3. Structured output: validate and retry

    Validates

    The one loop every schema-bound call shares: truncation guard, JSON repair, the caller’s validator, and one retry on a malformed answer.

  4. Embeddings

    Model call

    Vectorise events and canonical nodes for recall; no prompt, an embedding request.

› The complete workflow diagram Every stage and connection the product declares and tests. A route is drawn because it exists, not because it ran on a case. This is the software’s workflow, not an organisation’s VSM, and it uses the software’s own stage names.
PREPARE SOURCE MATERIAL FRAME THE CASE AND SELECT EVIDENCE INITIAL ANALYSIS AND GAP EXPLORATION BUILD AND EXAMINE THE SYSTEM MAP ASSESS VSM FUNCTIONS DEVELOP THE DIAGNOSIS PREPARE INVESTIGATION QUESTIONS CONVERSATION HUMAN FEEDBACK Parse and span documents Read a document into a ledger A Transcribe the ledger to structured entities and relations Recover missed relations Flow-trace channel signals A Channel-signal backfill (operator script) Record document evidence events Declare the case frame Decide whether new evidence reruns the analysis C Select the current Map inputs Capture the Diagnosis inputs Free-form initial analysis (deep run) A Gap exploration probe (deep run) D Skeleton from evidence (deterministic) Skeleton by model A Trace signal flow A Assemble the audit evidence capsule B Audit: classify or abstain B D Completeness census B D Structuring: System One operations B Cross-level channels A B Reason as the VSM: hypotheses A Recurse: spawn child systems A B Canonical structure and reasoning events A Function-first assessment loop (Track B) A Quote grounding verification A Function assessment findings Diagnosis certificate D Certificate and citation validation A Structural hypotheses: reasoning Structural hypotheses: harvest to JSON Confidence gate A D Translate and record findings A Generate the diagnosis report Diagnosis page Question agenda (deterministic) D Question rendering by model D Differential questions Conversation reply A D Turn extraction A Per-question side agents A Consultant answers a question C Staff interview answers C Route human answers into admitted evidence C Map corrections and identity confirmations C Duplicate-actor suggestion C Intervention outcomes and failed attempts C Consultant analytical review of a finding C Review, sign and freeze, then promote assessed structure C Contextual contribution
Model call (a paid request to a language model)
Ordinary code, a check or a saved record
Conditional stage (runs only on some routes)
Colour band: which part of the journey the column belongs to
Data passed to the next stage, on the ordinary consulting route
The same, on a side route (deep run, function-first, chat, feedback)
Branch: taken only when its condition holds
Retry
A person’s later contribution flowing back into the analysis

What the talk claims, and where it runs

A Every claim the model promotes quotes the passage it rests onDeclared exception: structural hypotheses are keyed to recorded observations rather than to a quotation. The check differs by stage; see the stage entries.
B Structure is linked at the next recursion level only when an evidence gate passes
C A person’s corrections persist through reanalysis as their own recorded eventsEight of the nine feedback stages plus reanalysis; contextual contributions are deliberately kept out of analysis.
D Unresolved gaps become diagnostic questions rather than invented answersOnly supported structural gaps become questions, and the question agenda is capped.

A stage is marked only where the catalogue entry itself says it does that work; a stage can carry more than one claim.