Coaching Conversation Assessment with a Spoken Round for Managers and Internal CoachesNot how you would coach. What you would say next, turn by turn, and where in the conversation it changes.

Methodology: Forty-three original items across six formats: eighteen scenario turns in three conversations of six, eight forced-choice most-and-least items, six select-every-line transcripts, four matching items, four ordering items and three spoken turns. DECLARED RESPONSE INSTRUCTION: one instruction is declared for the whole instrument and it is BEHAVIOURAL TENDENCY: every turn asks what you are most likely to SAY next, never what a coach should say; the forced-choice items ask which of four utterances you are most and least likely to say; nothing tests knowledge of coaching. Behavioural-tendency framing is less fakeable and correlates better with real behaviour than knowledge framing (McDaniel, Hartman, Whetzel and Grubb, 2007). CONSTRUCT STATEMENT: this measures what a manager actually says next in a coaching moment, turn by turn, inside three conversations that are already running, across four parts: whether the next utterance stays on the other person's agenda, widens rather than narrows what they are considering, names what was seen or heard rather than a trait, and leaves the next step with them with a when they chose; and where in a conversation that changes. It does not measure whether the reader is an empathetic person, their coaching qualifications, how much coaching theory they know, the approach they would take across a whole relationship, how anything sounded, or how they would do in a conversation this instrument did not describe. HOW IT DIFFERS FROM THE COACHING INSTRUMENTS IN THIS CATALOGUE: the Coaching and Developing Others battery and the Sales Coaching Judgement instrument measure the coaching APPROACH a manager would take, across a relationship. This measures the single next utterance, turn by turn, inside one running conversation, and reports WHERE IN THE CONVERSATION the reader changes, which no approach-level instrument can see. SCORING DESIGN: a construct-weighted graded situational judgement test with a SPARSE credit vector. Every turn's four options carry an effectiveness value v from 0 to 3 and a construct-credit vector over the four parts, content.construct_credit, keyed to the published evidence: goal-focused coaching keeps the goal the coachee's (Grant, 2003; Theeboom, Beersma and van Vianen, 2014; Jones, Woods and Guillaume, 2016, for the meta-analytic effect sizes of workplace coaching and the finding that the coachee's own goal and reasoning carry them); autonomy support against control (Deci and Ryan, 2000); feedback at the task and process level against the self level, where praise and the Barnum effect live (Hattie and Timperley, 2007; Kluger and DeNisi, 1996); trait labels against behaviour (Dweck, 2006); reflective listening and open questions as the pairing that produces the other person's own change talk, and distorting or verdict-laden reflections as the pairing that produces defence (Miller and Rollnick, 2013; Magill, Apodaca, Borsari, Gaume, Hoadley, Gordon, Tonigan and Moyers, 2018, meta-analysis of the technical hypothesis); advice given before the person has reasoned as suppressing that reasoning and being discounted (Bonaccio and Dalal, 2006; Deelstra, Peeters, Schaufeli, Stroebe, Zijlstra and van Doornen, 2003, on unwelcome help at work); implementation intentions, a when and a where the person sets themselves, for the closing move (Gollwitzer and Sheeran, 2006); and the tell-versus-ask account of coaching practice (Whitmore, Coaching for Performance). Every turn puts AT MOST TWO parts in play, every option loads only on those, and the best option holds the maximum credit on each, so no option loads on more than two parts and picking the best utterance never depresses a part the turn does not credit. S_k is the credit earned on part k over the answered items crediting it, M_k the maximum attainable on those items, C_k the expected credit under each item's DECLARED marginal, and figure_k = 100 (S_k/M_k - C_k/M_k) / (1 - C_k/M_k), in -100 to 100. The forced-choice, transcript, matching and ordering items feed the parts on the same arithmetic, each worth up to three credit: forced choice (credit of MOST + 3 minus credit of LEAST) / 2 against two declared pole marginals; select-every-line the share of lines whose ticked state matches the key, so over- and under-selection both cost; matching and ordering the share of correct placements. The headline is the eighteen turns alone on v against option_base_rate. THE NULL AND ZERO: zero on every figure is the ORDINARY respondent, somebody choosing at each item's declared marginal (option_base_rate on turns and forced-choice MOST, option_base_rate_least on forced-choice LEAST, select_prior per transcript line, match_prior per matching left, position_prior per ordering position). The marginals are authored assumptions stated in the open, non-uniform within every item, and authored so that the ordinary respondent earns about the mean credit of the four options on every part a turn credits: a fixed position, which the platform's shuffle turns into a coin, must not beat the ordinary respondent, and the probe checks that every fixed habit (each position, longest, shortest, always ask a question, always a what-or-how question, always reflect, always advise, always the utterance that begins with I will, select all, select none, identity) lands within seven points of zero on the headline and on every part. Observed shares replace the declared marginals once live data exists. SPARSITY, PROVEN: the inter-competency correlation matrix from a reference simulation of three thousand sittings through this instrument's own builder, with four latent tendencies sharing half their variance and each option's attractiveness tilted by its own credits, is printed on the report (maximum off-diagonal .66, under the .70 above which a report tells the reader the same thing four times); a second run with independent tendencies gives the correlation the credit vectors alone manufacture (maximum .34). Both are declared assumptions and observed data replaces them. RELIABILITY: McDonald's omega in its Spearman-Brown form from the answered item count and an assumed average inter-item correlation of .24, printed as an assumption; every figure carries its standard error of measurement from a declared, modelled standard deviation, its 68 and 95 per cent bands, and a three-way placement against the ordinary respondent (McDonald, 1999). REPORT DESIGN: the heat strip item map ordered along the CONVERSATION: one strip per part, one cell per turn in conversation order, four greyscale-distinguishable steps each with a value label and a word, a turn not answered drawn empty and labelled no answer, a turn not crediting the part drawn not in play. Because the x-axis is the conversation the finding is a shape: the mean quality on the opening third (turns 1 and 2) against the closing third (turns 5 and 6), pooled across the three conversations, per part and overall, printed with the difference and the measurement threshold 1.96 times a declared within-person per-turn standard deviation times the square root of the sum of the reciprocals of the two counts; a difference inside the threshold is within measurement error and is drawn flat. The within-strip spread is printed beside each strip with a word, even, mixed or spiky, at declared cuts of .20 and .35. THE SPOKEN ROUND: the pipeline that commissioned this product asked for an AI-scored spoken role play; the platform delivers a spoken turn per conversation, recorded, transcribed and rubric-graded on the TRANSCRIPT against four content criteria, the same four things the written turns measure. No pronunciation, tone-of-voice, warmth, pace or fluency score reaches the builder and none is claimed; a coaching conversation is partly how it sounds and this instrument does not score that. The spoken round is reported entirely apart, enters no figure, and a turn whose transcription or grading did not complete is printed DEFERRED in words rather than as a score. LOCALE: the three conversations are set in an Indian workplace with mixed-seniority pairs, and where deference to seniority changes the better next utterance the item says so in its scenario rather than hiding it in the key; that item is the flagship, tagged for expert review, and its keying follows the finding that autonomy support in a high power-distance setting means scaffolding the step rather than withdrawing from it (Hofstede, 2001; Deci and Ryan, 2000). REFUSAL RULES: an unanswered item leaves numerator, denominator and chance term; an empty sitting scores exactly zero everywhere and is not reportable; fewer than twenty-two of the forty keyed items answered is withheld with every refusal printed where its figure would have gone; a part with fewer than eight answered items or omega under .70 on the unrounded value carries a placement word and no figure; a shape with fewer than two answered turns in either third is not computed; no total across the four parts, no percentile anywhere. RE-MEASUREMENT: the report names a re-sitting in about six weeks and prints, per part, the movement it would have to exceed to count as real change, 1.96 times the declared standard deviation times the square root of 2 (1 - omega), from the Reliable Change Index (Jacobson and Truax, 1991), and distinguishes it from the smaller difference that would matter to the people the reader coaches. FURTHER SOURCES: de Haan, Duckworth, Birch and Jones, 2013, for what carries coaching outcomes; Ely, Boyce, Nelson, Zaccaro, Hernez-Broome and Whyman, 2010, for the evaluation of leadership coaching; Baumgartner and Steenkamp, 2001, for response styles. No proprietary coaching model, questioning framework or credential is named or used; the constructs are public and are cited; no item from any published instrument is reproduced or adapted. All forty-three items are original works written for this instrument. No competitor is named anywhere, and the instrument is not affiliated with or endorsed by any author or organisation named above.