drdavidprivacy/reachfirst-model

ReachFirst v1 is the version the Chief AI Officer signed in Week 6 of the AI Governance Practicum (Track 1). The organization, the system, and the people are fictional; the data is public. Educational use only, not for real decisions.

What is in this repository

• reachfirst_v1.pkl: 4,300 bytes, SHA-256 120234d9f9a111028a2d605d2779a9d8e151ff8d6f85df4ec0497f651595b953
• reachfirst_v1_Inside.csv: every fitted column with its weight, and the intercept, as a table a person can read.
• Model-Card_Chen-exemplar.pdf: the card below as a PDF.

Lineage: trained on the remediated working file in drdavidprivacy/reachfirst-data, which links to the file itself in drdavidprivacy/practicum-case-packs. Served by the Space drdavidprivacy/reachfirst. Libraries: scikit-learn 1.9.1, fairlearn 0.15.0; the model does not load on another version of scikit-learn, by design.

Model card below: the AI Governance Lead's exemplar (Lauren Chen), the answer key to the Week 6 Model Card. Students replace it with their own; see "Make this your own" at the end.

---

Model Card: ReachFirst v1 (Crosswind Workforce Partnership)

Analyst: Lauren Chen, AI Governance Lead (exemplar) Organization: Crosswind Workforce Partnership System: ReachFirst Version documented: v1 (recommended for signature), compared with v2 Evidence: Week 6 evaluation notebook, sections 1 to 15, and the documents cited by name

Section 1: Model details

Objective: Identify the model a reader is holding, and every fitted step production has to reproduce.

1.1 Model name and organization

ReachFirst, Crosswind Workforce Partnership's model for ordering the outreach list, so that the outreach team contacts first the households most likely to be at or below the income line the programs serve (organization briefing).

1.2 Version documented

This card documents v1, the champion the Chief AI Officer ratified in her Week 4 memo and the version I recommend for signature (section 8). It was compared with v2, the same recipe refit with the correlation remover (notebook, sections 2 and 5).

1.3 Model family and fixed configuration

Logistic regression, regularization C = 0.1, class weights off, scaled inputs, ten inputs encoded as 58 columns. v2 is fitted on the same 58 columns after the part of each that moves with sex, race, and country of origin is removed (Week 4 rationale, section 2, by reference, and notebook, sections 2 and 5).

1.4 Fitted steps production must reproduce

The one-hot encoding levels fixed on the training rows, the median fill for any numeric gap (the file has none), the scaler, and, for v2, the correlation remover fitted on the training rows (notebook, sections 2 and 5, and Week 4 memo, 'What proceeds with an assumption and a control').

1.5 Mitigation

Fairlearn's CorrelationRemover, fitted on the training rows against sex, race, and country of origin, 24 encoded columns of 82. It removes from every other input the part that moves with those three, then drops the three. Neither version has the three fields as inputs. What the remover changes is that v2's ten inputs no longer carry them (notebook, section 5).

1.6 Artifacts and integrity hashes

reachfirst_v1.pkl, 4,300 bytes, SHA-256 120234d9f9a111028a2d605d2779a9d8e151ff8d6f85df4ec0497f651595b953. reachfirst_v2.pkl, 18,322 bytes, SHA-256 5ecaaa41d3290bc987ad90625fcf7ba66b07935f878326b0c9e62427079e02d5 (notebook, section 15).

1.7 Owner and holder of this card

Larkspur Analytics builds and retrains the model under contract and owns the code. Miguel Santos, Director of Outreach and Navigation, owns the capacity that sets the operating point. The AI Governance Office holds this card and the program file. The Chief AI Officer is the approver of record (Week 4 memo, 'The operating point', and organization briefing).

Section 2: Intended use

Objective: What the model is for, who uses it, and what it may not do.

2.1 Primary use

Order the outreach list so that the outreach team, which can reach about 20 percent of the outreach universe in a cycle, contacts first the households most likely to be at or below the income line, instead of relying on word of mouth and walk-ins (Use Case Assessment, and Week 1 decision memo, by reference).

2.2 Intended users

The Director of Outreach and Navigation and his team of coaches. The score proposes an order of contact. The team keeps judgment over its own week and decides whom to contact (Week 1 decision memo, 'The decision', and Week 4 memo).

2.3 Posture and stage

A priority ordering, not a cutoff. Everyone in the outreach universe stays on the list. The program is before first use: no household has been contacted, or left uncontacted, on a score, and the Week 4 operating-point row is a decision record, not a mailing list. The next stage is the measured first cycle, half of the capacity worked by the list and half by current practice, and the whole budget follows the list only if that comparison says so (Week 1 decision memo, and Program Record, Status tab).

2.4 Out of scope

A score never states or implies that a household is eligible, never decides any benefit, and never removes anyone from the list. No neighborhood batching. The use gate (the Chief AI Officer's permission for a household to be contacted, or left uncontacted, on a score) is closed until the watch-list checks have run and each is closed in writing, the criteria are in plain language with the board's community seats, the bottom tenth of the list has been described by group, the unknown-group check has run on the champion, and 20 percent of a counted outreach universe is in the operating-point record (Week 3 memo, 'My decision on each', and Week 4 memo, 'The operating point').

Section 3: Factors

Objective: The groups and conditions the model was tested across.

3.1 Groups and the fields that define them

The groups the memos name: sex (Female, Male), race as recorded (Amer-Indian-Eskimo, Asian-Pac-Islander, Black, Other, White), age band (under 25, 25 to 34, 35 to 44, 45 to 54, 55 to 64, 65 and over), marital status as recorded (seven values), and country of origin by the measurement copy, in which the 26 values under 100 records are grouped as all other and the placeholder is its own row. The records with no employer recorded are their own row in the unknown-group check (Week 2 memo, and datasheet, section 2).

3.2 Measurement fields as model inputs

In v1, age and marital status are inputs under the proxy test, and sex, race, and country of origin are held out of the inputs and kept in the file for measurement. In v2 the same ten inputs are used, with the part of each that moves with sex, race, and country of origin removed (notebook, sections 1 and 5).

3.3 Groups too small to judge

Groups too small to judge against the tripwires (under 500 records in the test set): the Amer-Indian-Eskimo (83), Asian-Pac-Islander (300), and Other (71) race rows, the 65 and over band (439), the Married-AF-spouse (5), Married-spouse-absent (121), Separated (329), and Widowed (290) statuses, and every country of origin but the United States, Mexico (198) and the all other line (258) included. They are reported with their rates and no conclusion (notebook, section 4).

3.4 Limitations on every group measurement

Every group measurement carries two limitations. The file is from an older records system with no field tying a record to Crosswind's service area, so a rate by group is a rate on yesterday's residents, and every contact rate the program publishes is compared against the current service-area population, never the file. And the label is a person's own income at the 50,000 line, so no group's rate is its share of households that qualify (Week 3 memo, and Program Record, Items tab, Items 12 and 13).

Section 4: Metrics

Objective: Which numbers the card reports, and why those and not others.

4.1 Ranking and count-based metrics

The card reports two ranking metrics: ROC-AUC (how often a randomly chosen at-or-below record outranks a randomly chosen above-the-line record) and PR-AUC (precision averaged over every possible cut, read against the file's 76.1 percent at-or-below rate). The count-based metrics (ranked in, caught, missed, false alarms, precision, recall) are reported at the capacity cut and not at the 0.50 reference threshold, because the team acts on a share of the list and 0.50 ranks in 81.0 percent of the test rows, about four times the capacity (Week 4 rationale, sections 2 and 3, by reference, and notebook, section 3).

4.2 Operating point

The operating point is the top 20 percent of the ranked list each cycle, what the outreach team can reach. Miguel Santos owns the capacity, and if his reach moves, the share moves with it. A probability threshold is not the operating point: the cut on the test set lands at a score of 0.986, a figure that moves at every retrain while a share does not (Week 4 memo, 'The operating point', and notebook, section 3).

4.3 Fairness measures

Four fairness measures decide. The false above-the-line check: at the cut, each group's share of its at-or-below records that were not ranked in, against the file's, with the tripwire that no group of 500 or more sits more than 20 percent above the file's rate. The marital-status check: inside each status of 500 or more, the same rate for records marked Female beside Male, with the tripwire that the higher is at most 20 percent above the lower. The unknown-group check: the same rate for the records with no employer recorded against everyone else's, tripwire 20 percent above. The sex-and-race recovery check: a simple classifier guessing sex, and separately race, from the champion's score and the ten inputs, with the tripwire that it does no more than 10 points better than always guessing the most common value. Equalized odds difference (the largest gap between groups in the false above-the-line rate or the false alarm rate) and demographic parity difference (the largest gap in the share ranked in) are reported beside them as context. The tripwires decide (notebook, sections 4, 7, 8, 9, and 12, and Week 3 memo, 'The watch list from my last memo').

4.4 Procedural checks

Two checks are procedural and have no pass line. The fields-in-versus-out check (the champion trained twice, as remediated and with sex, race, and country of origin added back as inputs, and the two lists compared) records what adding the three fields back would change, and for whom. The bottom-tenth check describes who sits in the last tenth of the ranked list, by group, against the file (notebook, sections 10 and 11, and Program Record, Checks tab).

4.5 What the card does not report

Accuracy, printed once in the notebook at 0.853 against the 0.761 baseline of saying at or below to everyone, because that baseline finds no one. Not reported at all: any rate read as the share of residents who qualify, because the label is a person's income and the household mapping is not written. Nothing from the retired prototype (Week 3 memo, the decision gate, the no-eligibility-statement rule, and the no-prototype rule).

Section 5: Evaluation data

Objective: What the model was tested on, and what that data is not.

5.1 The test set

The test set is 9,758 records, 7,422 at or below the line and 2,336 above, 76.1 percent. It was split off in Week 3 by row, stratified on the outcome, 20 percent of the working file, seed 42, after the 52 duplicate rows were removed. It was read once in Week 4 (section 9) to confirm the cross-validation order, and it is read here to evaluate v1 and v2. Nothing in this card was chosen on it (notebook, sections 2 and 3, and Week 3 memo, the decision gate).

5.2 The test-set-read-once rule

The test-set-read-once rule completes here: the cross-validation table was printed before the test-set table in Week 4, v1 was chosen on that table, and this evaluation is the test set's second and last reading for this version. The split and the comparison close on the Program Record with this card (notebook, section 2, and Program Record, Checks tab, Check 09).

5.3 What the test set is not

The test set is not Crosswind's residents and not a record of who qualifies. It is one fifth of a file from an older records system with no date and no service-area field, and the label is a person's own income at the 50,000 line. Every rate here is a rate on that file (Week 3 memo).

Section 6: Training data

Objective: Where the model learned.

6.1 Working file

The working file is crosswind_pilot_extract_remediated.csv: 48,790 records, 10 model inputs, 76.1 percent at or below the line (37,109 records), MD5 ca9ed6d4b9dfd59f3f586b09da05b31f (datasheet, by reference, and notebook, section 1).

6.2 Training rows

39,032 records, at 76.1 percent (notebook, section 2).

6.3 Remediation behind the file

The file is the output of the eleven-row remediation log the Chief AI Officer ratified in Week 3: the label fixed as the person's income with at or below the line as the reporting class, Husband and Wife recoded to one level in a new household field, the ceilings kept and flagged, the sampling weight held out, the placeholder kept as a level, 52 duplicate rows removed under one written rule, country of origin held out with a measurement copy added, sex and race held out and kept in the file, one representation of education kept, and the split rule recorded (datasheet, section 4).

6.4 Open items on the data

Open items on the data that still stand: Item 07, sex found through the household fields, and marital status and age as inputs. Item 08, race found through country of origin, and the groups too small to test. Item 09, the placeholder records. Item 13, the vintage and the outreach universe. Item 14, the 52 duplicate rows (Program Record, Items tab).

Section 7: Quantitative analyses

Objective: The results, read against the fold range, the tripwires, and the pass lines, for the version you will recommend.

7.1 Table A

Table A. v1 against v2 on the test set.

	v1 (as trained)	v2 (correlation removed)	Notebook section
Cross-validation ROC-AUC, mean (spread)	0.904 (0.014)	0.878 (0.014)	6
Test-set ROC-AUC	0.905	0.881	6
Test-set PR-AUC	0.967	0.956	6
Top 20 percent: ranked in	1,952	1,952	6
Top 20 percent: at-or-below records caught	1,945	1,936	6
Top 20 percent: at-or-below records missed	5,477	5,486	6
Top 20 percent: false alarms	7	16	6
Top 20 percent: precision	0.996	0.992	6
Top 20 percent: recall	0.262	0.261	6
Groups of 500 or more that tripped the false above-the-line check	35 to 44 (22 percent above the file's), 45 to 54 (23 percent), 55 to 64 (24 percent), Married-civ-spouse (35 percent)	45 to 54 (21 percent above), 55 to 64 (25 percent), Married-civ-spouse (34 percent)	4 and 6
Marital statuses of 500 or more that tripped the marital-status check	none (largest gap 8 percent, Never-married)	Never-married (Female 0.743 against Male 0.263), Divorced (0.960 against 0.539)	7
Unknown-group check: the placeholder group's rate against everyone else's	0.491 against 0.756 (35 percent below), pass	0.550 against 0.753 (27 percent below), pass	8
Sex-and-race recovery check: points above the most-common-value rate (sex, race)	13.3 (TRIPPED), 0.1 (pass)	as v2 sees it: minus 0.5 (pass), 0.0 (pass)	9

7.2 v1 against the Week 4 fold range

v1 against the Week 4 fold range: every one of the five metrics sits inside its range. ROC-AUC 0.905 against 0.898 to 0.912, PR-AUC 0.967 against 0.965 to 0.970, and precision, recall, and F1 at 0.50 inside theirs. The Week 4 choice stands on the test set. The challenger's 0.929 is what the cross-validation table said it would be, and at the cut it catches 4 more at-or-below records than v1 out of 1,952, the price the Week 4 memo paid for criteria that fit on one page (notebook, sections 2 and 3).

7.3 Fairness

The false above-the-line check trips under every version, and not where the Week 2 memo expected. Under v1 the groups above the line are three age bands, 35 to 44 at 0.897, 45 to 54 at 0.904, and 55 to 64 at 0.916, 22 to 24 percent above the file's 0.738, and the married status: of 2,490 at-or-below records marked Married-civ-spouse, 7 are ranked in, a rate of 0.997, 35 percent above. The reason is in the same table. The cut is where the young and the never-married sit: under 25 is ranked in at 66.3 percent of its records and Never-married at 50.9 percent, and the top 20 percent of a list ordered by how sure the model is fills up with them before any prime-age married household is reached. v2 moved the bands a little (35 to 44 passes at 0.861, the other two still trip) and the married status not at all (0.990), and it moved sex the wrong way: the Female rate rose from 0.676 to 0.858 and the Black rate from 0.649 to 0.815, both inside the line and both worse. The challenger trips the same bands and the married status, and Divorced as well. So neither version clears the check, and v2 clears nothing v1 trips while worsening two groups v1 served better. The version I recommend is chosen on that (section 8) (notebook, sections 4 and 6).

7.4 The other three checks

The marital-status check passes for v1 in every status of 500 or more (the Female and Male rates within 8 percent of each other) and trips for v2 in two: Never-married at 0.743 against 0.263, and Divorced at 0.960 against 0.539. The unknown-group check passes for every version: the placeholder records are ranked in at 45.3 percent against 18.5 percent for everyone else under v1, and their rate is 35 percent below everyone else's. The model does not sink them. The sex-and-race recovery check trips for sex under v1 at 13.3 points above the most-common-value rate and passes for race at 0.1. The ten inputs alone recover sex at 13.1 points, so the score adds almost nothing: the inputs carry it, through occupation, marital status, and household role, which is what the Week 2 memo said they would. Seen through the remover, v2's inputs and score recover sex at 0.5 points below the majority rate. The remover takes sex out of what the model sees, and the tables above say what that did to the list (notebook, sections 7, 8, and 9).

Section 8: Fairness and explainability decision record

Objective: The recommendation, and the evidence it rests on.

8.1 Recommended version

I recommend the Chief AI Officer sign v1, the champion she ratified in Week 4.

8.2 The two numbers

V2 costs 0.024 of test-set ROC-AUC (0.881 against 0.905) and 9 at-or-below records at the cut (1,936 caught against 1,945), with 16 false alarms against 7. Its fold range is its own, 0.878 mean against 0.904, spreads 0.014 both. For that price it clears one of v1's four trips (the 35 to 44 band) and leaves three, it trips the marital-status check in two statuses v1 passes, and it raises the Female rate from 0.676 to 0.858 and the Black rate from 0.649 to 0.815. 1,394 of the 1,952 records are on both lists. A mitigation that moves the harm onto women and onto Black records is not a mitigation this program can sign (notebook, section 6).

8.3 The false above-the-line check

Neither version clears the false above-the-line check, and the use gate stays closed on both. The Week 3 memo says what follows a trip: contact stops on that item, and the inputs doing the work for the tripped groups are named before any retrain. They are named. Married-civ-spouse is the first driver of the score by mean absolute SHAP value (0.662) and carries the fourth largest fixed weight (minus 1.332), the household role Spouse is the third driver (0.356), age the sixth (0.252), and the three records in section 13 are carried by Never-married and the household role. The trip is not a flaw in the fit. It is what a single cut on a list ordered by certainty does when one side of the line is 76 percent of the file: the surest at-or-below records are the young and the single, and the married households at or below the line, 2,483 of them in the test set, sit below the cut and above the bottom tenth, in the middle of the list where no check so far has looked. I recommend the item escalate with two questions for the Chief AI Officer, neither of which this office can answer. Whether the operating point should stay a single share of one ranked list, or become a share within each group the memos name, which is a change to the posture and belongs to her and the board. And whether the Week 3 tripwire, written for a cut that reaches 15 to 20 percent of a file in which 76 percent qualify on the label, is measuring what she meant it to measure. I do not propose a fix. I propose that no list is mailed, in the first cycle or any other, until she has ruled (notebook, sections 4, 6, and 13, and Week 3 memo, 'The watch list from my last memo').

8.4 Explanation evidence

The criteria the board's community seats receive are v1's fixed weights, the same for every resident: capital gain (minus 2.242), the three advanced-degree levels (minus 1.637 to minus 1.227), Married-civ-spouse (minus 1.332), the three lowest education levels (0.952 to 1.195), and farming and fishing (0.877). A coach at the counter can say that a resident was ranked where they were because the file's record of their schooling, their work, their household, and their marital status put them there, and can say which. For the caught record (file row 42,937, rank 975) and the missed one (file row 42,108, rank 1,953, the first below the cut), the reasons are the same three words, never married and an own child or other relative in the household, and the two records are 978 places apart on hours and occupation alone. For the false alarm (file row 25,135, rank 417), 7th to 8th grade schooling and farm work set the score, and the person's income was above the line. What a coach cannot say is anything about eligibility: the score is an order on an old file, and the no-eligibility-statement rule stands (notebook, section 13, and Week 4 rationale, section 2).

8.5 Ruling requested on marital status and age

I ask the Chief AI Officer to keep both, and to record the sex-and-race recovery check's trip on sex with the fields named rather than retrain without them. The check trips at 13.3 points, and the inputs alone trip at 13.1: occupation, marital status, and household role carry sex, as the Week 2 memo found. The one version in front of her that takes sex out of what the model sees is v2, and the tables say what it did. Removing marital status and age outright would be a third version, untested here, and the Week 1 posture says the model must be explainable at the counter, which a model that cannot see a household's structure would not be (notebook, section 9, and Week 3 memo, 'What we are closing').

8.6 The bottom tenth and the measured first cycle

The bottom-tenth check did not find what the Week 2 memo expected. The last tenth of the v1 list is 976 records, 125 of them at or below the line, and it is over-represented by men (1.36 times their share), the four bands 35 and over (1.25 to 1.84 times), and Married-civ-spouse (2.04 times). Unmarried women, residents under 25, and the placeholder records are not at the bottom. They are at the top. The bottom of this list is the married, prime-age men who are mostly above the line, which is where a list ordered by this label should put them. The measured first cycle stands as designed, and it is now the only safe way to run a first cycle at all: half the capacity by current practice is the control that shows whether the married at-or-below households the list leaves in the middle are the ones word of mouth and walk-ins reach today (notebook, section 11, and Week 4 memo, 'What proceeds with an assumption and a control').

8.7 Rows that go to the Chief AI Officer

Rows that go to the Chief AI Officer: Check 04 (the bands and the married status above the line under every version, with the inputs named), Check 05 (sex recovered through the inputs, with the fields named and the ruling asked for on marital status and age), and Check 03 (the bottom tenth, for her memo and the board) (Table B).

Section 9: Caveats and recommendations

Objective: What the card's reader must not assume about the version you recommend, and what happens next.

9.1 Limitation this version adds

The limitation v1 adds: at the 20 percent cut it reaches 7 of the 2,490 married at-or-below households in the test set and about one in ten of those aged 35 to 64, and no document may describe the list as reaching eligible families evenly until the Chief AI Officer has ruled on the operating point. The score can be read for sex by anyone holding the ten inputs, at 13.3 points above guessing, through occupation, marital status, and household role. And the fixed weights are the criteria: a change to any input changes the page the board holds. The items it inherits are in 3.4 and 6.4 (notebook, sections 4, 9, and 13).

9.2 For the Week 8 monitoring plan

The false above-the-line check monthly, by every group the memos name, at whatever operating point the Chief AI Officer rules. The bottom-tenth check as a monitoring table each cycle. The contact-rate check (contact and enrollment rates by group, against the current service-area population, never the file) on the two halves of the first cycle. The drift check (the share scored at or below the line against the file's 76.1 percent) each cycle. And the unknown-group check whenever the model is retrained (Program Record, Checks tab, and Table B).

Section 10: The closed-out watch list

Objective: Every check on the Program Record, closed out in writing.

10.1 Table B

Table B. The checks, closed out.

ID	Check	Status	Result against the pass line	Where it is printed
Check 01	The plain-language criteria	Not due this week	Week 7 deployment materials. v1's fixed weights, the criteria, are printed in the notebook and listed in section 8 of this card.	Notebook, section 13. Program Record, Checks tab
Check 02	The contact-rate check	Not due this week	Week 8 monitoring plan, on the two halves of the first cycle, against the current service-area population. No contact has occurred.	Program Record, Checks tab
Check 03	The bottom-tenth check	Checked by inspection	Procedural, run and printed: 976 records, 125 at or below. Over-represented: Male (1.36 times its share), the bands 35 and over (1.25 to 1.84), Married-civ-spouse (2.04). Under 25, Never-married, Female, and the placeholder records sit at the top, not the bottom.	Notebook, section 11
Check 04	The false above-the-line check	Failed and escalated	v1: 35 to 44 at 0.897, 45 to 54 at 0.904, 55 to 64 at 0.916, Married-civ-spouse at 0.997, against the file's 0.738 and the line 0.886. v2: 45 to 54 at 0.894, 55 to 64 at 0.922, Married-civ-spouse at 0.990. Challenger trips the same and Divorced. Sex and race pass under both. Inputs doing the work named in section 8.	Notebook, sections 4 and 6
Check 05	The sex-and-race recovery check	Failed and escalated	Sex: 79.7 percent right against 66.4 for always guessing Male, 13.3 points, line 10. The inputs alone: 13.1 points. Fields named: occupation, marital status, household role. Race: 0.1 points, pass. v2 as v2 sees it: minus 0.5 and 0.0, pass.	Notebook, section 9
Check 06	The marital-status check	Passed	v1: every status of 500 or more within 8 percent (Divorced 1, Married-civ-spouse 2, Never-married 8), line 20. v2 trips Never-married (183 percent) and Divorced (78 percent). The four statuses under 500 are count only.	Notebook, section 7
Check 07	The unknown-group check	Passed	v1: 0.491 against 0.756 for everyone else, 35 percent below, line 20 percent above. Challenger 24 percent below, v2 27 percent below. Owed since Week 4, run here on the champion.	Notebook, section 8
Check 08	The fields-in-versus-out check	Checked by inspection	Procedural, run and printed: with sex, race, and country of origin added back, ROC-AUC 0.905 against 0.905, 1,945 caught against 1,945, 85 of 1,952 records (4.4 percent of the cut) differ. The inputs already carry the three fields.	Notebook, section 10
Check 09	The test-set-read-once rule	Passed	Run in Week 4 and completed here: the test set read once in Week 4 for the order and once here for the evaluation. v1 refit reproduces 0.905; every metric inside its fold range. Accuracy printed once, 0.853 against 0.761, and judges nothing.	Notebook, sections 2 and 3. Week 4 notebook, sections 5, 9, and 10
Check 10	The no-eligibility-statement rule	Checked by inspection	No sentence in this card states or implies a household's eligibility from a score, and no group's rate is called its eligibility rate.	This card
Check 11	The no-prototype rule	Checked by inspection	Nothing in this card or the notebook cites the retired prototype.	This card. Notebook, section 14
Check 12	The duplicates question	Waiting on an answer	Neither Larkspur nor the source has reported that any of the 52 removed rows was a second person. The split stands.	Program Record, Items tab, Item 14
Check 13	The drift check	Not due this week	No current residents have been scored. Runs each cycle from the Week 8 plan against the file's 76.1 percent.	Program Record, Checks tab

---

Make this your own

This page carries the AI Governance Lead's exemplar so that the repository is complete on day one. Your portfolio should carry yours. Three steps, about ten minutes:

1. On Hugging Face, the + menu, then New Model. Owner: your account. Name: reachfirst-model. License: cc-by-4.0. Public. Create it.
2. Export your Model Card (your Week 6 deliverable) as a PDF and upload it to the new repository under the name Model-Card.pdf (Files, then Contribute, then Upload files).
3. Upload a README.md: copy README_TEMPLATE.md from this repository, put your name in it, and replace YOUR-ACCOUNT with your Hugging Face account name everywhere it appears (including inside the metadata block at the top, which otherwise stays as it is). The template links to your PDF. If you would rather have your card render as the page itself, paste the text of your Model Card below the metadata block instead.

Then add the new repository to your collection beside your Space and your dataset repository. Do not upload the Lead's exemplar, the memos, the audit briefings, the notebooks, or the data.