Item profile
TechnoSeal IM-1652 Ionomer Resin
Sole-sourced sealing-layer material, $4.02M annual spend, HHI 1.0. Demonstrates supplier_concentration, low_supplier_diversity flag, and propylene-tracked price discipline.
Progressive disclosure levels
{
"artifact_type": "tabular_manifold",
"artifact_version": "1.2",
"manifold_kind": "entity_profile",
"subject": {
"entity_type": "ca_item",
"ca_item_number": "CA-c1d05a30b8e4.7f1652",
"cleansed_part_number": "TS-IM1652",
"cleansed_part_name": "TechnoSeal IM-1652 Ionomer Resin",
"cleansed_part_description": "Heat-sealable ionomer resin pellet (DuPont Surlyn 1652 equivalent) used as the lidding-seal layer in pharmaceutical blister packs and tamper-evident food-cup seals. Critical sealing-layer material; no qualified alternate grade in current BOMs."
},
"time_window": {
"start": "2024-05-01",
"end": "2026-04-30",
"granularity": "month"
},
"level_0_summary": L0Summary{
"observation_count": 51,
"time_coverage": {3 keys },
"distribution": {6 keys },
"reliability": {6 keys },
"quality_flags": {9 keys },
"financial_summary": {12 keys },
"price_stability": {3 keys },
"supplier_concentration": {4 keys },
"commodity_classification": {4 keys },
"relationship": {4 keys },
"interpretation_hints": [4 items ]
},
"level_1_geometry": L1Geometry{
"monthly_timeseries": {5 keys },
"supplier_rollup": {4 keys }
},
"level_2_telemetry": L2Telemetry{
"row_count_total": 51,
"strategy": "preview_outliers",
"inline_rows": {2 keys },
"retrieval": {5 keys }
},
"token_budget": {
"level_0_tokens_approx": 480,
"level_1_tokens_approx": 1820,
"level_2_inline_tokens_approx": 1140,
"row_count_total": 51,
"inline_row_limit": 6,
"compression_ratio_l1_vs_l2": 7.4,
"compression_ratio_l0_vs_l2": 18.2,
"recommended_strategy": "Read level_0_summary first. The dominant signal (supplier_concentration HHI 1.0 + low_supplier_diversity flag) makes the next step a sourcing-policy question, not a data-investigation question — Level 1 / Level 2 only needed if validating the price trajectory or confirming PO cadence."
},
"lineage": {
"manifold_id": "mfld_ca_item_CA-c1d05a30b8e4.7f1652_20260501",
"computed_at": "2026-05-01T03:14:08.642Z",
"computed_by": "technoflex_item_profile_v2.4",
"qps_entry_id": "qps_technoflex_item_CA-c1d05a30b8e4.7f1652_20260501",
"inputs": [3 items ],
"filters_applied": [3 items ],
"transformations": [3 items ],
"checksum": {2 keys }
}
}
Glossary
Field reference
Domain and spec terms that appear above. The full schema is in the TMS specification; this section explains the parts that are easy to misread without procurement or statistics context.
Envelope
manifold_kind- Identifies the cognitive shape of the manifold. TechnoFlex examples use the v1.2 draft "entity_profile" kind, an extension to the five canonical kinds (timeseries_metric, funnel_conversion, cohort_behavior, inventory_snapshot, anomaly_detection) defined in TMS v1.1.
subject- The entity the manifold describes. Always carries an "entity_type" plus the natural keys for that kind (ca_item_number for items, parent_supplier_id for suppliers, commodity_group_label, sub_commodity_label, etc.).
time_window- Observation window for the underlying data. "granularity" is the aggregation grain used in Level 1 ("day", "month", "quarter"). The window can extend past the last observation (a sparse tail is expected when source data is late-arriving).
Reliability & sample size
sample_size_class- Banded confidence in the L0 summary: sparse (n < 30, treat stats with caution), adequate (30 ≤ n < 100, reasonable but not rock-solid), robust (n ≥ 100, high confidence). Set against observation_count in this manifold's window.
confidence_in_mean- Classical 95% confidence interval around the mean of the leaf metric (typically unit_price). "margin_of_error" is half the interval width. Single-PO items collapse this to zero MoE and are flagged by other means.
data_quality_score- 0–1 score reflecting completeness and attribution of the underlying rows. Anything below ~0.9 means a meaningful fraction had to be imputed or excluded. Pairs with data_quality_notes for the narrative.
staleness.is_stale- True when the most recent observation is older than stale_threshold_days (default 90 for purchasing data). Stale manifolds should not drive sourcing decisions without a refresh.
Quality flags
quality_flags- Booleans that nominate this manifold for attention. Designed so that an agent can read L0 and decide whether to drill into L1 / L2 based on which flags are true. Agent prompt pattern: 'If any quality_flag is true, read level_1_geometry; if outliers are flagged, read level_2_telemetry.'
high_supplier_concentration- Set when supplier_concentration.hhi exceeds the FTC 'highly concentrated' threshold of 0.25. On the TechnoFlex Colorants commodity this is 0.898, deep in the band.
low_supplier_diversity- Item-level flag set when a single SKU is sourced from exactly one supplier. Often the most actionable flag in procurement: it surfaces single-source risk independent of price discipline.
level_1_item_truncated- True when the Pareto-truncated item_rollup at Level 1 dropped rows. Tells the agent that the tail_summary is the only signal it will get for the truncated items without going to Level 2.
has_unattributed_supplier_spend- True when some PO-line spend in the window could not be attributed to a parent supplier ID (typically rebate credits or one-off contractor lines). Quantified in financial_summary.unattributed_spend_usd.
Distribution & statistics
cvCoefficient of variation- Standard deviation divided by the mean. Unitless measure of relative dispersion. Large CV on a global metric is often misleading when the rows span heterogeneous SKUs; see weighted_avg_item_cv for the corrected per-SKU read.
distribution- Five-number summary plus mean, stddev, and CV computed across the leaf metric (unit_price for entity profiles). Built from all rows in the window, not just the Pareto-truncated head.
Concentration
hhiHerfindahl-Hirschman Index- Market-concentration metric: sum of squared spend shares across distinct entities. Range 0–1 (or 0–10,000 by FTC convention). Bands: <0.15 unconcentrated, 0.15–0.25 moderate, >0.25 highly concentrated. HHI 1.0 is a monopoly.
top_supplier_pct- Share of spend held by the single largest supplier for this entity. Paired with HHI, this distinguishes 'concentrated because of one giant vendor' from 'concentrated across a small but balanced group'.
primary_commodity_group- The dominant commodity group inside a supplier's portfolio, used for the supplier-side commodity_concentration block. For multi-commodity vendors this often reveals where they are strategic vs. opportunistic.
Pricing discipline
weighted_avg_item_cv- Spend-weighted average of per-item CVs. Isolates pricing volatility from product-mix effects (the global CV conflates the two when the supplier sells across heterogeneous price points). The right metric for vendor discipline.
discipline_rating- Banded version of the weighted_avg_item_cv: Excellent (<0.10), Good (<0.25), Fair (<0.50), Poor (≥0.50). Thresholds are spelled out in rating_basis so the consumer can re-band if their tolerance differs.
pricing_data_coverage_pct- Fraction of spend eligible for per-item CV. Items with only one PO line cannot contribute a stddev and are excluded; this metric tells you how much of the portfolio the discipline_rating actually covered.
Pareto truncation
truncation.method- Always "pareto" in v1.2 entity_profile: rows are sorted by rank_metric (spend), and only enough rows are emitted to reach target_coverage of the total, bounded by [min_rows, max_rows]. The tail is summarized, not deleted.
target_coverage- Cumulative share of spend that the included rows must cover. 0.80 for supplier_rollup / item_rollup, 0.95 for industry / commodity rollups where the long tail is more informative.
tie_break- "tie_extended_by_spend_v2" means rows tied at the truncation boundary are included if they cover incremental spend beyond the threshold. Prevents arbitrary cutoffs when two rows are within rounding of each other.
tail_summary- Aggregate stats over the rows that were truncated. Always includes rows_truncated, spend, pct_of_spend. Type-specific extras: distinct_commodity_groups for item rollups, tail_industry_mix for supplier rollups.
Telemetry & retrieval
level_2_telemetry.strategy- What the inline_rows preview contains. Common values: preview_outliers (z-score-flagged rows), preview_recent (last N), full_inline (all rows, only if row_count_total < 50), preview_concentration_evidence (rows that drove an HHI flag). Always paired with a retrieval block for the full set.
z_score- Per-PO-line price deviation from the entity's mean, in standard deviations. |z| > 3 is usually flagged. In the TechnoFlex examples, look for z_score and flag together: 'price_outlier_high', 'uom_likely_misencoded', 'price_peak_period'.
retrieval.method- How the consumer fetches the rest of Level 2: mcp_tool (recommended for agents, opaque), sql_query (the literal template, used inline by tekni-mcp), or none (preview is all there is). sql_fallback is always present in the TechnoFlex examples so a human can copy-paste the query.
Token budget
token_budget- Estimated token cost per level plus a recommended_strategy string. The compression ratios (typically ~7× L1 vs L2, ~18× L0 vs L2) are why progressive disclosure matters: 95% of the time an agent only needs L0.
recommended_strategy- Natural-language instruction for the agent on how to read this specific manifold. Used as the "manifold-shaped" half of the system prompt when the agent receives the JSON.
Lineage & provenance
manifold_id- Stable identifier for this build of this manifold. Encodes the entity ID and the snapshot date. Used as the cache key and as the link target from agent reasoning traces.
qps_entry_id- Pointer into a companion Query Provenance Store (see /spec/qps). Lets an auditor replay the exact query that produced this manifold and detect drift against the original generation.
checksum.method- How the lineage checksum was computed. "sha256_row_hash" hashes the sorted, serialized rows; "row_count_only" is the weak-but-cheap option. Used by QPS replay to flag drift.
inputs[]- Source datasets the manifold was built from, each with dataset_id, version, row_count, and as_of_timestamp. The version field is what lets QPS detect upstream changes.