Source identity
| Field | Recorded value |
|---|---|
| Source type | GitHub source |
| Original record | Janady13/missionmatch / docs/AI_ANI_ROADMAP.md |
| Date | Not specified in source |
| Repository | Janady13/missionmatch |
| Revision | 720d47fae476b0d38477ae4401697d372b1ea2bc |
| Branch | feature/opportunity-to-outcome |
| Path or archive member | docs/AI_ANI_ROADMAP.md |
| Original SHA-256 | bdd7246bdb71fc317cd958b00955a2b0866bb0159cf73ceafbc21301e5c6b1d1 |
| Git blob SHA | 7607b2b8e93288ac690d479c3c821d75d172f364 |
| Reading-copy SHA-256 | bdd7246bdb71fc317cd958b00955a2b0866bb0159cf73ceafbc21301e5c6b1d1 |
Every recorded occurrence
These 12 records contain identical original source bytes. Their paths and revisions remain part of the archive.
| Original record | Branch or archive | Revision / identifier | Original location |
|---|---|---|---|
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | feature/opportunity-to-outcome | 720d47fae476b0d38477ae4401697d372b1ea2bc | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | collision-deconfliction | 4ee5a09972608ed89f8377c6665542d5f450f0a6 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | m5-price-to-win | 41000ad30fff4129fdaf5ec84e15ba14c27d4890 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | master | 1e9249840a4302ec82e123b9c9cb6f75badea3e4 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | mag-substrate | b82c7100a492dec67dfd1eda6b3f53def859be35 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | time-to-fill-survival | b66e1e505863d4f17741a4b9ede73509e45bd6c6 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | graphrag-foundation | b0ddeb9f633fd138d09309d1a17b7d13bc3a67f7 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | presentation-readiness | eb574acaa29d71ce4d0f852572e87c275c79fdb7 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | vendor-intelligence | fe9796a19c261951d3f2565206c0c1b5ac294465 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | phase2-connectors | e038d4c687aec16f8bd9bc4d907eea14d91f8704 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | feature/aioe-upgrade | 2178e7298d0e5d9af83afa12402fd587f0ba6ed8 | Open original |
| Janady13/missionmatch / docs/AI_ANI_ROADMAP.md | capture-readiness | bae6af7cb2997e819174a9e9cfff5e16067a1975 | Open original |
Recovered source transcript
Download complete source text ↓68 lines · 10.9 KB
# missionMATCH AI Layer: the road to best in world ANI
Source: a nine agent grounded research sweep (2026-06-30) across the whole AI layer (RAG, TAG, Graph, parametric), each dimension grounded in the actual codebase first, then ranked by leverage. Every item respects the hard rails: the frozen deterministic engine stays the ONLY per candidate scorer, Postgres plus pgvector only and both cloud, no Python, AEDT compliant, FIPS.
## The headline
The AI layer answers single questions well but cannot yet reason ACROSS questions, measure its own answers, learn from outcomes, forecast, or watch the account. The missing core is a bounded PLANNER over the closed Registry plus the measurement, outcome label, and proactive scaffolding that make multi step reasoning honest, calibrated, and self improving.
## The single biggest unlock: a bounded plan and execute planner over the closed Registry
G6 the loop already proves the whole pattern by hand: one hardcoded fan out over four fixed intents, with small cell suppression, the faithfulness gate, and a deterministic narration fallback. Generalizing that one composition into a data driven TYPED DAG over the existing ~25 closed intents, gated by a plan validity grammar and a per plan verifier, converts the brain from a single shot keyword router (which drops every compound cross plane capture question to RAG narration, the top production failure mode) into a true multi step reasoner. The model still authors ZERO SQL and ZERO params: it picks node intents from a closed enum and binds args only through the extractors graphClassify already uses, so the frozen engine, EEO unreachability, suppression, RLS, both cloud, no Python, and FIPS all hold by construction exactly as for G6 today. It multiplies every intent already built and every one on this roadmap. It is the dividing line between a demo and a best in world capture ANI.
## What best in world means here
Not a better single retrieval call. A brain that (1) reasons across the bench, market, clearance, lifecycle, and teaming planes in ONE compound question and renders an auditable capture brief; (2) proves on every release that its answers are faithful, its retrieval recalls the right people, and it never drifts into disparate impact; (3) gets measurably better at THIS customer's win patterns by recording real bid outcomes and calibrating confidence against observed recruiter and award behavior; (4) tells the analyst unprompted that a winnable recompete just dropped or a key clearance is expiring; (5) hands back the smallest proposal team that covers the PWS plus the named residual gap and flags any scarce cleared person double committed across competing bids. Every one rides the existing governance rails: the frozen engine is the only per candidate scorer, every learned or predictive layer scores an OPPORTUNITY or emits a DISPLAY ORDER (never a second candidate score), EEO is unreachable by package walk, counts are k=10 suppressed, the model authors no SQL, and every threshold or coefficient move is an auditable model registry version bump. Most capable AND most provably compliant: the defensibility (continuous four fifths time series, conformal controlled abstention, plan trace audit, vendor entity provenance) IS the moat for a FedRAMP and LL144 buyer, not a tax on capability.
## Tier 1: highest leverage, buildable now, zero or trivial new data (alongside or right after G6)
- HIGH/L Bounded plan and execute planner over the closed Registry (typed DAG of existing intents), deterministic Go DAG runner, hard node and depth caps. The biggest unlock.
- HIGH/M Plan validity gate plus per plan faithfulness and self consistency verifier plus plan trace audit (reject any out of set or unbound or kind mismatched plan; verify the narration against the UNION of node evidence; sample the plan K times and run only the stable shape, else the deterministic G6 fallback).
- HIGH/S Snippet on the Hit (ts_headline span on retrieval results) so the gate and narration reason over text, not just names. A precondition.
- HIGH/M Recall at k, nDCG, MRR gold set retrieval harness over the live RRF HybridSearch (versioned qrels in testdata, Go harness), so any retrieval regression is caught before it ships.
- HIGH/S Continuous adverse impact and four fifths time series with alerting (a scheduled snapshot of eeo.AdverseImpact into an append only fairness_snapshots table, trend alert at 0.80).
- HIGH/M Minimal cover optimal slate construction (greedy weighted set cover) returning the smallest best fit team that covers every required skill, clearance, cert, and key personnel slot plus the named uncovered remainder; engine Total is the tiebreak.
- HIGH/M Cross opportunity win collision deconfliction flagging every scarce cleared person assigned or proposed on two or more competing open bids, with a deterministic suggested primary a human resolves.
## Tier 2: the labels and the learned and predictive layer (scaffold now, models time gated on label volume)
- HIGH/M M7 bid_outcomes ledger plus win pattern base rate read (a tenant scoped RLS forced table captured at the existing human decision points). Ship it dormant NOW so labels accrue. The single most reused unlock.
- HIGH/M Shared versioned model coefficient and calibration store (forecast_models, coefficients, calibration_bins; migrator owned, SELECT to app) that every predictive and learned intent reads, registered like aedt_model_versions.
- HIGH/L Interpretable calibrated win probability intent over opportunities, stored coefficient logistic or EBM scored in SQL, features from G1 incumbent share, G9 recompete, set aside fit, segment past performance; per feature additive contribution as the trace.
- HIGH/L LambdaMART style learned slate re ranker that emits DISPLAY ORDER ONLY, never a second score, trained on recruiter approve and reject labels, consuming only the engine's already computed sub scores and structural features.
- HIGH/M Outcome calibrated confidence (isotonic or Platt map on realized labels) replacing the deterministic recommendation confidence.
- HIGH/M Recompete timing hazard intent plus time to fill survival forecast (interpretable discrete time logistic hazard, per month hazard list as evidence, cumulative survival as a bounded recursive CTE).
- MEDIUM/M Price to win wage band estimator that fills the G6 honest floor leg (BLS OEWS percentile band plus a loaded rate range cross checked against award amounts).
- HIGH/M Bid or no bid explainable weighted scorecard (GO, REVIEW, NO GO) over signals already grounded, every factor and weight in the trace.
## Tier 3: measurement depth, entity truthfulness, proactive reach (mostly buildable now)
- HIGH/M Event driven proactive alert engine (the watch loop): a cron that runs fixed trigger intents against newly changed rows and writes typed alerts (winnable recompete dropped, clearance expiring, gap now fillable).
- HIGH/M Vendor and UEI entity resolution (deterministic Fellegi Sunter record linkage into a canonical vendor_entities table, every merge through the existing two person curation gate, never auto merge).
- HIGH/M RAGAS in Go answer quality benchmark plus conformal calibrated abstention (per route operating points so the brain abstains at a guaranteed maximum hallucination rate).
- HIGH/M Topological link prediction over the G2 vendor graph (Common Neighbors, Jaccard, Adamic Adar, Preferential Attachment) surfacing the likely next teaming partner as ranked structure with a cited path.
- HIGH/M Champion and challenger offline replay harness with a no adverse impact regression gate; promotion is a human action; on drift the orchestrator falls back to the frozen engine order.
- MEDIUM/M Propensity and IPS correction on recruiter feedback labels (weight by 1 over propensity from the display rank) so the learned ranker is unbiased to position.
- MEDIUM/S Revive the reranker label agreement channel and reuse it for active learning (persist the learned rank and label alongside each frozen decision, flag lowest confidence slates to label next).
- MEDIUM/L Multi requirement maximum cardinality matching (Hopcroft Karp) so reported portfolio coverage is provably maximal, with an augmenting path explanation; keep greedy as the fast path and oracle.
## Tier 4: data and time gated, higher architecture risk, or aspirational (after the substrate proves out)
- HIGH/M M1 subaward and FSRS ingestion so the actual staffing subcontractor is visible, not just prime of record. Makes the teaming graph and incumbent intelligence true.
- MEDIUM/L Section M and contract clause ontology plus the one justified generative extraction pass (M6): a temperature 0 grammar constrained self hosted extractor tagging each opportunity with its evaluation factors and weights. Tells the team HOW a bid is scored.
- MEDIUM/L node2vec structural graph embeddings in pgvector (biased walks as recursive CTEs, an in house Go skip gram trainer, a 384 dim embedding per UEI and agency) for analog partner and analog competitor reasoning and as a link prediction feature.
- MEDIUM/M Full NICE and DCWF skill graph plus O*NET SOC crosswalk (M5 depth) giving G3 substitution closure real depth.
- MEDIUM/L Candidate availability and attrition risk index from structural signals only, itemized signed terms, suppressed and audited, surfaced to the human, never ranked into the match.
- MEDIUM/S Clearance verification provenance at onboarding (self reported, fso attested, diss confirmed) on the existing clearance_provenance field, with an expiring alert.
- MEDIUM/L Retrieval depth batch: resume chunking under HNSW rolling up to the candidate, weighted and tunable RRF, multi hop query decomposition, span level citation to chunk id, the bge query prefix fix plus a re embed verify cmd, and the deferred cross encoder reranker via an ONNX sidecar (flag off, RRF fallback).
## Recommended sequencing (relative to the current plan)
Insert measurement before reasoning, and labels before learning.
1. Land the three cheap preconditions alongside or right after G6: snippet on the Hit (S), the recall at k gold set harness (M), the continuous adverse impact time series (S). They make every later change provable.
2. Ship M7 bid_outcomes EARLY (dormant, flag gated) so labels start accruing the day it lands, because every learned layer is time gated on label volume.
3. Build the planner (the biggest unlock) as the natural generalization of G6, with its plan validity gate and plan trace audit, BEFORE the heavier learned ranker (the planner needs no labels and immediately multiplies the intents already built).
4. Keep M5 SOC crosswalk and M8 portfolio roughly where they are, but reframe M8 as the deterministic minimal cover set cover plus collision deconfliction plus the Hopcroft Karp upgrade (all no new data, buildable now).
5. Defer the genuinely data and time gated items (the learned re ranker, isotonic calibration, win probability and hazard models, propensity correction, champion and challenger, node2vec, Section M extraction) to their natural window once M7 labels and the eval harness exist.