Mandate Fit Scoring

Turns a list of companies into a list of decisions. One bucket per company, one reason string per bucket, the same rubric applied to every row.

Read first:

• references/scoring-rules.md — the buckets, the hard rules, the audit
• references/ownership-taxonomy.md — which types are actionable
• references/signals-and-evidence.md — because fit and signal are scored separately

Step 0 — Load and scope

1. Read mandate-profile.md. If it does not exist: stop and create it from templates/mandate-profile.md. Scoring without it invents a mandate
2. Read the market map, for the segment language and the sponsor list
3. Read targets.csv
4. Scope: rows with status = classified, or any row whose fit_score is blank

If ownership_type is blank across many rows: stop and run ownership-classification first. Ownership gates scoring per hard rule 7, and scoring without it produces a list you have to re-score.

Scoring an imported list

If the user has their own CSV, map their columns to the schema in templates/targets.csv first and show them the mapping. Anything you cannot map becomes Not found, never an inference from an adjacent column.

Step 1 — Build the rubric in writing

Before scoring anything, write the rubric out from the mandate profile. Every criterion gets a name, a source in the profile, a test, and a bucket effect.

RUBRIC — [sector] — built [date] from mandate-profile.md [date]

HARD EXCLUSIONS (any one → Not a Fit, record excluded_by)
  H1  ownership_type in {PE-backed (platform), PE-backed (add-on), public}
  H2  ownership_type not in mandate section 7 in-scope list
  H3  company appears in mandate section 8 exclusions or portfolio conflicts
  H4  company owned by a sponsor named in the market map
  H5  segment explicitly out of mandate (section 4)
  H6  geography outside mandate section 5, where a constraint is stated

SIZE BOX (from section 3)
  S1  verified revenue inside range               → in box
  S2  verified revenue outside range              → Not a Fit
  S3  labeled estimate outside range              → Near Miss
  S4  revenue Not found                           → not a failure; see hard rule 5

MUST-HAVES (section 4)
  M1  [must-have]  → absent and verified: Near Miss
  M2  [must-have]  → absent and verified: Near Miss

STRONG FIT REQUIRES (all three)
  F1  in box and clear on every hard exclusion
  F2  ownership_type actionable and NOT Not found
  F3  a sourced, quotable signal — signal_strength high or medium

🚧 HARD GATE — rubric approval

Show the rubric in full. Ask specifically about:

• The exclusion list and the sponsor screen — the two most commonly wrong
• Whether Not found revenue should really be non-fatal (it should; explain why if they push back)
• The Strong Fit bar

Do not score a single row before the user approves the rubric. If they change it mid-run, hard rule 8 applies: every previously scored row is stale and the whole list re-runs.

Step 2 — Score every row, identically

Order of operations, per row:

1. Hard exclusions first. Any hit → Not a Fit, write excluded_by with the rule ID. Stop
2. Ownership gate. Unactionable → Not a Fit. Not found → cannot be Strong Fit, continue
3. Size box per S1 to S4
4. Must-haves per M rules
5. Signal check — read signal_strength. If signals have not been read yet, the row cannot be Strong Fit yet. Mark it Fit and flag it for re-scoring after business-research
6. Assign the bucket. Write fit_reason naming the specific rule or criterion

When genuinely torn between Fit and Near Miss, choose Fit. Per references/scoring-rules.md — the expensive error is excluding a real target on thin information.

Never move a row to make the distribution look better.

Step 3 — Write back

Fill fit_score, fit_reason, excluded_by, and set status to scored.

Never overwrite a sourced field with an inference while scoring. Scoring reads; it writes only these three columns and status.

🚧 HARD GATE — before writing

Show:

• Row count about to be updated
• Distribution across the four buckets
• excluded_by breakdown — which rules fired, how often
• Ten sample rows in full: three Strong Fit, three Fit, two Near Miss, two Not a Fit
• Any row where you had to make a genuine judgment call, listed explicitly

Get approval. Then write.

Step 4 — The four-part audit (mandatory)

Run it every time, including when the numbers look fine. Twenty minutes.

1. Sample the Strong Fits. Pull the 10 highest-confidence rows. Open the signal_source URL and verify the quoted signal actually appears there. Report any that do not — that is a fabrication caught, and it is the most important thing this audit does
2. Audit the Not a Fits for false negatives. Pull 20 at random. If any look obviously good, the filter is too tight. Report them by name
3. Re-run the sponsor and competitor screen against the keep pile. Anything that hits was missed upstream. Fix and note it
4. Sanity check the distribution against the healthy ranges in references/scoring-rules.md. 80 percent Strong Fit means the rubric is flattering the user. 5 percent surviving means the sourcing or the rubric is wrong. Either way, stop and reread the mandate

Report the audit results whether or not they are clean. An audit that is never reported is an audit that never ran.

Step 5 — Report

Scoring complete — 145 rows scored under rubric [date]

Strong Fit      14   (10%)
Fit             52   (36%)
Near Miss       41   (28%)
Not a Fit       38   (26%)

Exclusions fired:
  H1 sponsor-owned          19
  H2 ownership out of scope   9   (7 franchisee, 2 corporate)
  H3 portfolio conflict       3
  H4 named sponsor            4
  H5 segment out of mandate   3
  H6 geography                0   (no constraint set)

Size box:
  Verified in box            22
  Estimate in box            31
  Revenue Not found          79   ← not scored as failures
  Verified out of box        13

Strong Fit blocked by:
  Ownership Not found        11   → manual review would unlock these
  No signal read yet         23   → re-score after business-research

AUDIT
  1. Strong Fit signals verified at source:   10 of 10  ✓
  2. Not a Fit false negatives:               1 of 20 sampled
       - [Company] — marked Near Miss on segment, is actually in the
         contract-recurring segment. Moved to Fit. Rule M1 wording tightened
  3. Sponsor re-screen on keep pile:          1 new hit → moved to Not a Fit
  4. Distribution:                            within healthy range

Next: business-research on the 14 Strong Fit and 52 Fit rows, then re-score.

Notes

• The rubric is the deliverable as much as the scores are. Save it in the workspace as rubric-[sector]-[date].md so a score can be argued with later
• Fit and signal are separate questions. Never let a strong signal drag an out-of-mandate company into Fit, and never let a missing signal push an in-mandate company down. Hard rule 9
• Near Miss is a keep, not a reject. Mandates widen and companies grow. Re-scoring a kept row costs nothing
• Re-score after every mandate change. Offer to do it whenever the user edits the mandate profile
• Never score during discovery. The rubric exists so that scoring is inspectable, and an inline judgment call is not