Authoritative source, reproduced unchanged. Exact Markdown · Agent copy with resolved links. These instructions govern research runs, not checkpoint interviews.

METHOD, v43. What the method says, and the shelves.

Read whole at stage A1 and again at C2; the method shelf arrives at A1, B1 and C2, the explanation shelf at C3, the adaptation shelf at D1, the hierarchy shelf at D4, the critique shelf at E1. mem.py shelf lists the read tier, mem.py shelf --list the on-request tier. Nothing here overrides the contract.

What the method says, held at every stage

Gioia, Corley and Hamilton (2013) is the spine. Their first-order analysis adheres faithfully to informant terms and makes little attempt to distill: the categories explode on the front end, and you must get lost before you can get found; then similarities and differences among the many categories reduce the germane ones. Their manageable number, twenty-five or thirty, is one head's capacity for one site; here the hands remove that limit and the count is set by kinds (stage B1). Their second-order analysis is theoretical: themes that describe and explain, with particular attention to nascent concepts that have no adequate referent in the literature, distilled further into aggregate dimensions. The data structure is the display of the move from data to theory, and their mantra is no data structure, know nothing; the grounded model shows the dynamic relationships the static structure cannot, and it is the arrows that set everything in motion. The findings are suffused with informant quotes that align with the exemplars in the data structure. The literature is consulted late, in deliberate semi-ignorance. Intercoder agreement is not necessary at the tiers one head builds; where the fleets place items under a written instrument it is reported (contract §6). Their devil's advocate is here one hostile reader at the end of C2, the readers at C5, the labelers at D2, and no system of adversaries.

Grodal, Anteby and Holm (2021) name what you do. Qualitative analysis is categorization, and categorization is eight moves: asking questions, focusing on puzzles, dropping, merging, splitting, relating or contrasting, sequencing, developing or dropping working hypotheses. Rigor is in the moves, and a move is shown rigorous by its fit to the claim it serves, not by being listed (Günther, Joshi and Constantinides, below). Every move you make is logged with its name (tier_history.jsonl, model/history.jsonl, the diary) and snapped when it is made, and a move that changes a tier or the model says what prompted it and why the state after it explains the material better (stage B1 §6); the leg closes at the end of a pass, or earlier when you choose (STANDING-INSTRUCTIONS §2). Their three phases map onto the run once: generating is A1 to B1, refining C1 to C5, stabilizing D1 to D5, and stabilized never means closed. Their paper is method, not theory, and ships at B1.

Günther, Joshi and Constantinides (2026) say what a record of moves is for. Trustworthiness, not transparency as an end: a reader trusts a study when its claims, its analytical moves and its evidence fit one another, and disclosure serves that fit (pp. 854, 868). Two choices set what a claim owes (p. 859). Whose vocabulary comes first: here the participants' (Gioia's first tier, the third tier gerund-led, the literature at C3), so a move is defended by what it preserves of the rooms' meanings, never by a standard. What carries the claim: the words of texts or the shape of the corpus. This study makes both kinds, and each owes its own validation (p. 872): a claim about what rooms did or meant is validated by exemplars that pass its gate and by the return; a claim about a pattern across rooms, a count, a branch, by counts with their denominators (rules 5 and 6), by the documented events of the 73 hours (prior.md §4), and by the alternatives examined at D4. A dated sequence within a room rests on the documented events and the room's forest; how often a sequence occurs across rooms rests on counts with their denominators. Never validate one kind with the other's evidence. For a consequential move, record what prompted it, the move, what changed, and why the new state serves the claim better (pp. 865, 868; stage B1 §6); where the literature changes what a claim rests on, write a switch record (p. 872; stage C3 §1). Readers of one model agreeing with each other show that the instrument decides, not that it is right (FLEET.md §2); the paper's own ground truth discussion (pp. 857 to 858) says why. On request from A1 (CC BY 4.0, 28 pages); its examples are topic models, and its remarks on language models (pp. 872 to 873) are an extension, not a warrant for this design.

Eisenhardt (1989) governs the cases. Within-case analysis before cross-case; the cross-case step as a search for patterns across pairs and groups of cases; replication logic, each case confirming or disconfirming what the others suggested; the literature enfolded after the constructs exist; closure when the marginal case adds little. Her ceiling, that beyond about ten cases the volume of data stops fitting one head, is the ceiling this instrument exists to pass. The within-case tier is the coders'. The return's (D1 to D3) is hers only if something is predicted before it is seen: at D1 the handoff writes, per room, which elements the model expects to instantiate given the room's placements and which to be absent, and the ledger is scored against that expectation at E1. The return is in-sample throughout. Concepts absent from element citations are uncited evidence, not held-out observations: the analyst may already have read them. Record that coverage separately.

Van de Ven and Poole (1990) give the grammar of the record. Incidents are bracketed into events by a rule written before the bracketing; the forest is that grammar, and a coded turn is an incident. Langley (1999) supplies strategies for the dynamic part of the model; the run uses narrative, grounded theory, visual mapping and temporal bracketing (with brackets from prior.md) for the dynamics. The taxonomy at D4 and D5 compares rooms on global measures, which Langley calls the synthetic strategy and classes as variance theory, not process theory: say so where the branches are described, and keep the dynamic claims on the forest and the clock. Suddaby (2006) says what grounded theory is not: not raw data presented, not theory testing, content analysis or word counts, not a mechanical routine. Read that list beside every count you are tempted to report as a finding.

The adaptation shelf, at D1 to D3. Fisher and Aguinis (2017): theory elaboration as a mode of research between generation and testing, with seven tactics (horizontal and vertical contrasting, new construct specification, construct splitting, structuring specific, recursive and sequence relations); the operations a resumed case lead may perform are this instrument's operationalization, not the authors' own coding schema, and SCHEMA-adaptations.md §2 gives the one mapping. Busse, Kach and Wagner (2017): boundary conditions as the conditions under which a theory holds, explored by taking the theory to cases where it may fail, including construct refinement, moderators and mediators. Our restrict and contradict operations record different discrepancies; evidence decides whether to bound, refine, extend or replace the claim. Busse et al. do not prescribe that every discrepancy becomes a boundary. On request: Vaughan (1992), Ketokivi and Choi (2014), Sætre and Van de Ven (2021), Makadok, Burton and Barney (2018), Welch et al. (2022), Eisenhardt (2021), Ragin (2023), Robinson (1951), Turner (1953), Lakatos (1970). The hierarchy shelf, at D4 and D5. Doty and Glick (1994): a typology is a theory only when its types are ideal types with stated constructs and stated relations among them, and a case is placed by its distance to a type; a branch description that lacks either is a label. Their programme predicts outcomes from fit to type; that stays secondary here, and you take the definition, not the programme. Kozlowski and Klein (2000), the section on emergence: a higher-level property is composed when it is the same kind of thing as the lower-level properties that make it, and compiled when it is a configuration of different contributions; these are the words for saying what a form is relative to its rooms. McKelvey (1975): guidelines for classifying organizations empirically, the discipline a taxonomy owes before it names a branch. Pavitt (1984): a taxonomy of sectors built from observed patterns and turned into a theory, the nearest model for what stages D4 and D5 do (named on the shelf; its text is not held yet). On request: Hannan and Freeman (1986), Fiss (2011), Rich (1992), Grimm et al. (2005), Gray, Bryant and Greenhill (2010), Hsu and Hannan (2005).

The explanation shelf, at C3. Read: Puranam, Alexy and Reitzig (2014), the standing claim; Felin and Zenger (2017), the theory-based view: actors' theories choose what is pursued, and their 2016 Organization Science essay, a theory for the firm as the chosen bundle of problems (held in its SSRN version); O'Mahony and Ferraro (2007), governance emerging in a volunteer community; Bechky (2006), role-based coordination in a temporary organization; March (1991), read after these and weighted below them: a room holds many individual codes, and the subreddit carries an organizational code that newcomers learn and that changes as new ideas come in (the mutual learning model). On request: Machlup (1967), Felin and Zenger (2009), Felin et al. (2024), He et al. (2018), Lewin and Volberda (1999), Lewin, Long and Carroll (1999), Nickerson and Zenger (2004), Mintzberg (1979, the coordinating mechanisms, by OCR). Anything else you bring is recalled (stage C3). A work named on a shelf and not held says so in its file.

The critique shelf, at E1 and nowhere earlier. The published case for and against generative AI in qualitative research, with the call for papers the study is written for. It arrives after the findings are fixed, so it can shape how the audit answers the critics and not what the model says. Read: Schildt (2026), the design case for predictability, transparency and control; Jowsey et al. (2025), the case against AI in reflexive qualitative work; O'Sullivan and Ravasi (2026), prompting as a sensemaking aid to the researcher; the call for papers, AI-enabled frontiers in organizational science. On request: Ravasi et al. (2025), Bechky and Davis (2025), Goldenstein et al. (the Organization Studies special issue on interpretative computational methods), Shrestha et al. (2021), Choudhury, Allen and Endres (2021), Miric, Jia and Huang (2023), von Krogh, Roberson and Gruber (2023), Tidhar and Eisenhardt (2020), Gartenberg et al. (2026).

The shape of the run

Stages A1 to D3 are Gioia with the return: five tiers, the loop at C2 as the constant comparison, the return (D1 to D3) as its test, with the labelers writing each room's model in the global model's words and the adaptations as the diff. Stages D4 and D5 read the taxonomy off the adaptations and describe each branch at every level. Stage E1 audits the run from its logs, rebuilds sampled rooms from the global model plus their adaptations, and reports where the files and the analyst's account of itself disagree. There is no crossing in v38.

The four licences a judgement stage names in its final message (contract rule 17): Gioia's explosion at the front end; an in-sample return in the spirit of Eisenhardt's replication logic; Langley's strategy, declared; the count set by kinds rather than by a head's capacity.

The fleets are a prosthesis, not a panel. They are how one reader reaches seven and a half thousand second-order themes when one context cannot hold them, working under that reader's written instrument. The researcher is still the instrument. A fleet, a labeler or a hostile reviewer is a prosthesis or a critic and is never a judge (rule 13).

Three things Gioia has that this design does not, and what stands in for each. Theoretical sampling: the rooms were handed over as a census. What stands in is sampling inside the corpus, which stage C2 requires of every arrow, and the return (D1 to D3), where local discrepancies are resolved and shared adaptations examined (rules 12 and 16). Saturation: at stage B1 the stopping rule reads rooms in a seeded random order until its markers and a random draw say little is left, and the PI decides at marker 3 (tools/saturation.py; the report extrapolates to the rooms not read); kinds are recorded against second-order themes read, with each minting stamped by the harness (the leg, the rooms read so far, the second-order themes read so far, the tokens read so far, the room and theme that prompted it), and at C2 as passes of the loop with the same stamp on every element (contract §6). The informant relationship: the first tier is archived text. Claim informant voice nowhere.

What to write in a methods section. Claim Gioia for A1 to D3 with the five tiers and the loop; cite Grodal, Anteby and Holm for the move record, not for replicability; say the coder tier and the fleets are Shrestha, He, Puranam and von Krogh's division between pattern detection and abduction; say the return is an in-sample return modelled on Eisenhardt's replication logic, never a hold-out test (Shrestha et al. make the hold-out the test; this run holds none out): in-sample at the tier the model was built from (the coders' second-order themes, Gioia's first-order concepts consolidated within a case) with uncited concepts and quotes used as additional in-sample checks, never called out-of-sample; say the taxonomy classifies by shared changes to an explanatory model (shared innovations, not resemblance), that branches are a similarity hierarchy with no claim of descent, and that the chunk-share tree is kept as the theory-free comparison; state the four departures (the tier split between coders and analyst; no overlap between the coders' collection and your analysis; the census in place of sampling; the in-sample return); say what the canvas is and is not (contract §5); say in one sentence that you held every room's outcome while building and where the record declares its use (rule 7). On replicability: every prompt, input, output, transcript and model identifier is archived, so any claim can be traced; the study does not promise that a rerun reproduces it.

Operational limits preserved by v40

The unit and level of each explanation are explicit. A within-room loop needs a complete within-room dated chain; an inter-room loop needs jointly linked dated evidence and an account of the interactions. Neither similarity nor a graph cycle establishes a mechanism. Empirical construct bindings and their mapping to prior organizing theories remain separate, so zero or one dimension, a warranted remainder or no shared branch may be reported.

Local reconstruction retains supported particulars. Silent, absent, unnecessary and untestable are different evidence states. Fixed-global diagnostics, current adapted-local status and shared explanations retain their model versions. Pending claims remain unresolved. The D3 model stays frozen through D6 and E1; new D6 insights remain unvalidated proposals until a later affected-case return. These checks establish traceability and declared evidential state, not truth from file validity.