tenacious-market-space-map

name

tenacious-market-space-map

description

Build the population-level market-space map over the full Crunchbase ODM sample — sector × size × AI-readiness cells, with hand-labeled precision/recall validation and top-cell ranking for outbound allocation. Use this skill whenever the user mentions the market-space map, the Day 6 stretch, the distinguished-tier deliverable, market_space.csv, top_cells.md, methodology.md, hand-labeling validation, or any phrasing like "where should Tenacious point the system at the sector level." The underlying AI-maturity scoring primitive already exists in tenacious-signal-brief; this skill is the population-level extension with the validation discipline that keeps it honest — a superficial map is worse than none because it misdirects strategy with false confidence.

Tenacious Market-Space Map

The Day 6 distinguished-tier stretch. This skill is not required to pass the week — attempt it only if the five-act core deliverables are in good shape. A superficial market map damages the memo because it adds confident-sounding strategy claims on shaky statistical ground.

What this skill is and is not

• Is: a population-level segmentation of the Crunchbase ODM sample (1,001 companies) into (sector × size-band × AI-readiness) cells, with a rank of the top cells for outbound allocation, validated on a hand-labeled sub-sample with reported precision and recall.
• Is not: a generic sector analysis. The output must change how Tenacious allocates outbound effort — that is the distinguished-tier bar. "SaaS companies at 50–200 are promising" fails. "Dev-tools-for-engineers companies in the 50–200 band at AI-readiness 2, with ≥20 open engineering roles and no public Head of AI, are Tenacious's highest-return Segment 1 cell — 23 candidates, median bench-match score 0.81" passes.

The core primitive already exists

tenacious-signal-brief scores AI maturity (0–3) for one company. This skill runs that scoring over the full sample and aggregates. Do not re-invent the scoring rubric; import it. If you find a scoring bug during the population run, fix it in the signal-brief skill and propagate, do not fork.

The hand-labeling is the new work. The scoring rubric is the old work.

The work, in order

Step 1 — Enrich the population

Run the tenacious-playwright-enrichment pipeline over the Crunchbase ODM sample in batches. Use the frozen snapshot, not live crawls — 1,001 companies exceeds the 200 live-crawl cap and you do not need fresh signal at population scale.

Expected outputs:

• 1,001 enrichment bundles (enrichment_bundles.jsonl).
• A coverage_report.md naming how many companies had (a) a Crunchbase match, (b) a job-post page reachable in the snapshot, (c) a resolvable GitHub org, (d) complete enough data for AI-maturity scoring.

If coverage is < 70% on any dimension, flag it in the methodology — the map is lossy by that much and the precision/recall numbers should be reported conditional on complete records only.

Step 2 — Score AI maturity across the population

Run the signal-brief's AI-maturity scoring over every enriched company. Emit a per-company ai_maturity_score.json containing score, confidence, and per-input evidence.

Step 3 — Define sectors and size bands

Sectors: use the Crunchbase industry field as the starting point, but consolidate. Crunchbase has hundreds of labels; you want 10–15 meaningful sectors (e.g., "fintech infra", "vertical SaaS — healthcare", "dev tools", "data / ML platforms", "horizontal SaaS — marketing"). Document the consolidation in methodology.md.

Size bands: align with the ICP classifier's bands for cross-skill consistency:

• 15–80 (Segment 1 range)
• 80–200
• 200–2000 (Segment 2 range)
• 2000+

Readiness bands: 0, 1, 2, 3 (the AI-maturity score itself, not re-binned).

Cells: sector × size × readiness. Expected ~150 populated cells from 1,001 companies; most will have < 5 members (too sparse to act on).

Step 4 — Hand-label a validation sample

This is the step that separates a defensible map from a superficial one.

Sampling:

• Random sample 50 companies stratified across the score bands (roughly 12–13 per score 0/1/2/3).
• For each, spend ~5 minutes reading the team page, careers page, and any linked blog/press. Assign a human-judged AI-maturity score 0–3 with a one-sentence justification.
• This takes 4–5 hours. Budget it on Day 6 morning before touching the cell ranking.

Outputs:

• validation_set.jsonl with {company, automated_score, human_score, human_justification}.
• Precision/recall per score band:
  • For each band B, precision = (companies where automated and human agree on B) / (companies automated scored B).
  • Recall = (companies where automated and human agree on B) / (companies human scored B).
• Confusion matrix, 4×4.

If precision or recall on any band is < 0.6, the map is not trustworthy enough to ship. Options: (a) narrow the claim to bands where numbers are defensible, (b) fix the scoring primitive and re-run, (c) cut the stretch and use the Day 6 time on core deliverables. Explicitly pick one in the methodology.

Step 5 — Compute cell-level features

For each (sector × size × readiness) cell:

{
  "cell_id": "devtools__50-200__readiness_2",
  "sector": "dev tools",
  "size_band": "80-200",
  "readiness_band": 2,
  "population": 23,
  "avg_funding_last_12mo_usd": 18400000,
  "avg_hiring_velocity_delta_60d": 12.4,
  "bench_match_score": 0.81,
  "combined_score": 0.74,
  "confidence": "medium"
}

• Population: companies in the cell.
• Avg funding: mean last-round amount for companies with a funding event in last 12 months; null if < 3 companies in cell have one.
• Avg hiring velocity: mean delta_60d across cell members with complete data.
• Bench-match score: 0–1 score of how well the cell's inferred stack needs match bench_summary.md. A Python-heavy cell with a Python-heavy bench scores 1.0; a Rust-heavy cell with no Rust bench scores 0.1.
• Combined score: weighted sum of (funding recency, hiring velocity, bench match, readiness fit). Document the weights in methodology.md. Do not tune the weights to make the top cells look prettier — the weighting is a prior, not an outcome.
• Confidence: derived from cell population. < 5 companies → low regardless of other metrics; 5–15 → medium; 15+ → high.

Step 6 — Rank and write outputs

Three files:

market_space.csv — one row per cell. Columns match the JSON schema above. Order by combined_score descending. Include cells with population = 0 at the bottom for completeness; someone reading the CSV should see the whole grid, not just the populated bits.

top_cells.md — the 3 to 5 top-ranked cells with confidence: high. For each:

• One-paragraph profile (what these companies look like, what their hiring pattern is, what they do not have that the top quartile does).
• Specific outbound allocation recommendation (e.g., "Allocate 40% of Segment 1 weekly outbound here; expected yield 3–4 discovery calls per 25 touches based on signal-grounded reply-rate benchmark").
• Explicit risk notes — what could make this cell a bad bet (niche saturation, recent sector-wide layoffs, publicly visible competitor plays).

methodology.md — the methodology and the honesty register:

• Sector consolidation rules (the Crunchbase-label → sector mapping).
• Size-band boundaries and rationale.
• Scoring weights in the combined score.
• Hand-labeling protocol — sample size, stratification, labeler identity (probably you), inter-rater agreement if multiple labelers (usually N/A for a solo week).
• Precision and recall per readiness band, with CIs or at least the denominator.
• Known false positives and false negatives in the scoring — the "quietly sophisticated, publicly silent" and "loud but shallow" patterns from the signal-brief honesty rules apply at population scale and must be named here.
• What a reasonable user should NOT conclude from the map.

Honesty at population scale

Every population-level claim extends the per-company honesty rule. Specifically:

• Publicly silent companies are over-represented in score 0. If your map says "only 8% of the fintech-infra 50–200 band is at readiness 2+," caveat with "of companies with public signal — silent sophisticated companies are invisible to this measurement."
• Loud shallow companies are over-represented in score 2+. The validation set's confusion matrix will show this; quote the number.
• Sector consolidation masks heterogeneity. "Dev tools" covers both infra-for-AI and JS build tooling; readiness distributions differ meaningfully. If your cells aggregate across these, flag the heterogeneity.
• The Crunchbase ODM sample is not a random sample of the population. It is 1,001 companies selected for some reason by whoever assembled the sample. The map's generalizability is bounded by the sample's sampling method. Name this in methodology.md.

Why this is not automatically distinguished-tier

The rubric says distinguished-tier credit goes to a map that "would change how Tenacious allocates outbound effort." A technically-correct map that reproduces the intuition of a Tenacious partner (e.g., "post-Series-A SaaS is where we fish") does not earn it, even if the precision and recall are pristine. The originality credit lives in the specific cells you surface — the non-obvious ones with defensible evidence. If all your top cells are the ones Tenacious would have guessed, the output is sound but not distinguished.

One way to surface non-obvious cells: look for cells where bench_match_score is high AND combined_score is moderate AND population is adequate. These are the "underfished" cells — plausibly high-return but not the obvious pick. Name them in top_cells.md explicitly as "non-obvious but defensible."

When to stop and cut the stretch

Day 6 morning check:

• If precision or recall on any readiness band < 0.6 after hand-labeling, AND you cannot fix the scoring primitive in 2 hours, cut the stretch. Use the remaining time on Act V memo polish.
• If coverage of the enriched population < 60%, cut the stretch. The map is too lossy to claim strategic findings.
• If the top 5 cells all have population < 5 companies, cut the stretch. There is nothing to act on.

Cutting is not failure — a superficial map in the memo is explicitly penalized. The Act V memo with only the core five-act deliverables is a stronger submission than an Act V memo with a shaky market map.

What goes in the memo

If the map ships, it goes in the Skeptic's Appendix — not Page 1. One paragraph referencing the top cells + methodology file, with an explicit precision/recall statement. The memo's main decision argument stays grounded in the per-lead traces from Act IV; the market map is contextual evidence that informs pilot scope, not the core claim.

If the map does not ship, remove every reference to it from the repo. An unfinished market_space.csv in the deliverables folder will be read as a failure artifact.

Never do this

• Never ship without hand-labeled validation. Unvalidated scores are false confidence.
• Never tune combined-score weights post-hoc to make the top cells match intuition. That inverts the analysis.
• Never claim a cell-level finding with population < 5. Low-N cells go in the CSV for completeness but stay out of top_cells.md.
• Never reference a named Tenacious client as an example of a cell archetype. The case studies are redacted for a reason.
• Never let the stretch compromise Acts I–V. If Day 6 is eating into Act V memo time, stop.